Neil Lawrence's Talkstalks given by Neil Lawrence
http://inverseprobability.com/talks/
Mon, 25 Oct 2021 09:20:18 +0000Mon, 25 Oct 2021 09:20:18 +0000Jekyll v3.9.0AI, Data Science and the Covid19 Pandemic<p>With the world watching case numbers increase and publics and policymakers scrutinising projections from epidemiological models, the covid-19 pandemic brought with it increased attention on the use of data to inform policy. Alongside this scrutiny came a new wave of interest in the ability of data and artificial intelligence (AI) to help tackle major scientific and social challenges: could our increasing ability to collect, combine and interrogate large datasets lead to new insights that unlock more effective policy responses?</p>Mon, 18 Oct 2021 00:00:00 +0000
http://inverseprobability.com/talks/notes/ai-data-science-and-the-covid19-pandemic.html
http://inverseprobability.com/talks/notes/ai-data-science-and-the-covid19-pandemic.htmlnotesAccess, Assess and Address: A Pipeline for (Automated?) Data Science<p>Data Science is an emerging discipline that is being promoted as a universal panacea for the world’s desire to make better informed decisions based on the wealth of data that is available in our modern interconnected society. In practice data science projects often find it difficult to deliver. In this talk we will review efforts to drive data informed in real world examples, e.g., the UK’s early Covid19 pandemic response. We will introduce a framework for categorising the stages and challenges of the data science pipeline and relate it to the challenges we see when giving data driven answers to real world questions. We will speculate on where automation may be able to help but emphasise that automation in this landscape is challenging when so many issues remain for getting humans to do the job well.</p>Fri, 17 Sep 2021 00:00:00 +0000
http://inverseprobability.com/talks/notes/access-assess-address-a-pipeline-for-automated-data-science.html
http://inverseprobability.com/talks/notes/access-assess-address-a-pipeline-for-automated-data-science.htmlnotesEmulationIn this session we introduce the notion of emulation and systems modeling with Gaussian processes.Wed, 15 Sep 2021 00:00:00 +0000
http://inverseprobability.com/talks/notes/emulation.html
http://inverseprobability.com/talks/notes/emulation.htmlnotesMachine Learning and the Physical World<p>Machine learning technologies have underpinned the recent revolution in artificial intelligence. But at their heart, they are simply data driven decision making algorithms. While the popular press is filled with the achievements of these algorithms in important domains such as object detection in images, machine translation and speech recognition, there are still many open questions about how these technologies might be implemented in domains where we have existing solutions but we are constantly looking for improvements. Roughly speaking, we characterise this domain as “machine learning in the physical world.” How do we design, build and deploy machine learning algorithms that are part of a decision making system that interacts with the physical world around us. In particular, machine learning is a data driven endeavour, but real world systems are physical and mechanistic. In this talk we will introduce some of the challenges for this domain and and propose some ways forward in terms of solutions.</p>Tue, 13 Jul 2021 00:00:00 +0000
http://inverseprobability.com/talks/notes/ml-and-the-physical-world-tuebingen.html
http://inverseprobability.com/talks/notes/ml-and-the-physical-world-tuebingen.htmlnotesMachine Learning and the Physical World<p>You can’t have trust without understanding. We must view the systems we build as our tools, because if we can’t manipulate these systems, then we are at risk of being manipulated by these systems. Inspired by the centrifugal governor, this talk describes how statistical emulation provides a possible root for giving understanding to complex AI systems at a level of abstraction that allows humans to view the system as a tool, rather than being a tool of the system.</p>Wed, 07 Jul 2021 00:00:00 +0000
http://inverseprobability.com/talks/notes/ml-and-the-physical-world-trustworthy-ai.html
http://inverseprobability.com/talks/notes/ml-and-the-physical-world-trustworthy-ai.htmlnotesA Retrospective on the 2014 NeurIPS Experiment<p>In 2014, along with Corinna Cortes, I was Program Chair of the Neural Information Processing Systems conference. At the time, when wondering about innovations for the conference, Corinna and I decided it would be interesting to test the consistency of reviewing. With this in mind, we randomly selected 10% of submissions and had them reviewed by two independent committees. In this talk I will review the construction of the experiment, explain how the NeurIPS review process worked and talk about what I felt the implications for reviewing were, vs what the community reaction was.</p>Wed, 16 Jun 2021 00:00:00 +0000
http://inverseprobability.com/talks/notes/the-neurips-experiment.html
http://inverseprobability.com/talks/notes/the-neurips-experiment.htmlnotesAI Can’t Fix This: Happenstance Data, Modelling, and the Covid19 Pandemic<p>With the world watching case numbers increase and publics and policymakers scrutinising projections from epidemiological models, the covid-19 pandemic brought with it increased attention on the use of data to inform policy. Alongside this scrutiny came a new wave of interest in the ability of data and artificial intelligence (AI) to help tackle major scientific and social challenges: could our increasing ability to collect, combine and interrogate large datasets lead to new insights that unlock more effective policy responses? Experiences from the DELVE Initiative, convened to bring data science to bear on covid-19 policy, suggests achieving this aim requires wider adoption of open data science methods to deploy data science and AI expertise and resources to tackle real-world problems.</p>Thu, 20 May 2021 00:00:00 +0000
http://inverseprobability.com/talks/notes/ai-cant-fix-this-happenstance-data-modelling-and-the-covid19-pandemic.html
http://inverseprobability.com/talks/notes/ai-cant-fix-this-happenstance-data-modelling-and-the-covid19-pandemic.htmlnotesPost-Digital Transformation: Intellectual DebtDigital transformation has offered the promise of moving from a manual decision-making world to a world where decisions can be rational, data-driven and automated. The first step to digital transformation is mapping the world of atoms (material, customers, logistic networks) into the world of bits. But the real challenges may start once this is complete. In this talk we introduce the notion of ‘post digital transformation’: the challenges of doing business in a digital world.Mon, 17 May 2021 00:00:00 +0000
http://inverseprobability.com/talks/notes/post-digital-transformation-intellectual-debt.html
http://inverseprobability.com/talks/notes/post-digital-transformation-intellectual-debt.htmlnotesMachine Learning and the Physical World<p>Machine learning technologies have underpinned the recent revolution in artificial intelligence. But at their heart, they are simply data driven decision making algorithms. While the popular press is filled with the achievements of these algorithms in important domains such as object detection in images, machine translation and speech recognition, there are still many open questions about how these technologies might be implemented in domains where we have existing solutions but we are constantly looking for improvements. Roughly speaking, we characterise this domain as “machine learning in the physical world.” How do we design, build and deploy machine learning algorithms that are part of a decision making system that interacts with the physical world around us. In particular, machine learning is a data driven endeavour, but real world systems are physical and mechanistic. In this talk we will introduce some of the challenges for this domain and and propose some ways forward in terms of solutions.</p>Wed, 05 May 2021 00:00:00 +0000
http://inverseprobability.com/talks/notes/ml-and-the-physical-world-data-centric-engineering.html
http://inverseprobability.com/talks/notes/ml-and-the-physical-world-data-centric-engineering.htmlnotesAI Faith Panel DiscussionTue, 20 Apr 2021 00:00:00 +0000
http://inverseprobability.com/talks/notes/ai-faith-panel-discussion.html
http://inverseprobability.com/talks/notes/ai-faith-panel-discussion.htmlnotesAuto AI: Resolving Intellectual Debt in Complex Systems<p>Machine learning solutions, in particular those based on deep learning methods, form an underpinning of the current revolution in “artificial intelligence” that has dominated popular press headlines and is having a significant influence on the wider tech agenda. Our capability to deploy complex decision-making systems has improved, but our ability to explain them has reduced. This phenomenon is known as intellectual debt. The reality of deployed systems is they are constructed from interacting components of individual models. While a lot of focus has been on the explainability and reliability of an individual model, the real challenge is explainability and reliability of the entire system.</p> <p>In this talk we introduce the concept of Auto AI and give a road map to achieving fair, explainable and transparent AI systems.</p>Tue, 23 Mar 2021 00:00:00 +0000
http://inverseprobability.com/talks/notes/auto-ai-resolving-intellectual-debt-in-complex-systems.html
http://inverseprobability.com/talks/notes/auto-ai-resolving-intellectual-debt-in-complex-systems.htmlnotesInterpretable Models<p>The great AI fallacy is that we are building the first generation of automation that will adapt to humans rather than humans adapting to us. The more sobering reality is that we are building complex algorithmic decision making system that we are unable to explain. A FIT model is fair, interpretable and transparent. The machine learning community has placed effort into understanding how to improve interpretability into individual models, but the real challenge is how to build FIT systems. At the heart of the development of machine learning is the notion of separation of concerns, but this can obscure the real challenge which is responding to the human.</p>Tue, 09 Mar 2021 00:00:00 +0000
http://inverseprobability.com/talks/notes/interpretable-models.html
http://inverseprobability.com/talks/notes/interpretable-models.htmlnotesUncertainty, Procrastination and Artificial IntelligenceIn this talk I will introduce the importance of uncertainty in decision making and describe how it provides a mathematical justification for procrastination through the game of Kappenball.Mon, 01 Mar 2021 00:00:00 +0000
http://inverseprobability.com/talks/notes/uncertainty-procrastination-and-artificial-intelligence.html
http://inverseprobability.com/talks/notes/uncertainty-procrastination-and-artificial-intelligence.htmlnotesLaplace’s Gremlin: Uncertainty and Artificial Intelligence<p>With breakthroughs in understanding images, translating language, transcribing speech artificial intelligence promises to revolutionise the technological landscape. Machine learning algorithms are able to convert unstructured data into actionable knowledge. With the increasing impact of these technologies, society’s interest is also growing. The word <em>intelligence</em> conjures notions of human-like capabilities. But are we really on the cusp of creating machines that match us? We associate intelligence with knowledge, but in this talk I will argue that the true marvel of our intelligence is the way it deals with ignorance. Despite the large strides forward we have made, I will argue that we have a long way to go to deliver on the promise of artificial intelligence. And it is a journey that our societies need to take together, not just as computer scientists, but together by rediscovering the interdisciplinary spirit that Celsius, Linnaeus and their contemporaries did so much to demonstrate.</p>Thu, 11 Feb 2021 00:00:00 +0000
http://inverseprobability.com/talks/notes/laplaces-gremlin-uncertainty-and-artificial-intelligence.html
http://inverseprobability.com/talks/notes/laplaces-gremlin-uncertainty-and-artificial-intelligence.htmlnotesAI and the Future of WorknullWed, 03 Feb 2021 00:00:00 +0000
http://inverseprobability.com/talks/notes/ai-future-of-work-hsm.html
http://inverseprobability.com/talks/notes/ai-future-of-work-hsm.htmlnotesIntroduction to Machine Intelligence<p>With breakthroughs in understanding images, translating language, transcribing speech artificial intelligence promises to revolutionise the technological landscape. Machine learning algorithms are able to convert unstructured data into actionable knowledge. With the increasing impact of these technologies, society’s interest is also growing. The word intelligence conjures notions of human-like capabilities. But are we really on the cusp of creating machines that match us? We associate intelligence with knowledge, but in this talk I will argue that the true marvel of our intelligence is the way it deals with ignorance. Despite the large strides forward we have made, I will argue that we have a long way to go to deliver on the promise of artificial intelligence. And it is a journey that science and artificial inteligence need to take together.</p>Tue, 02 Feb 2021 00:00:00 +0000
http://inverseprobability.com/talks/notes/introduction-to-machine-intelligence.html
http://inverseprobability.com/talks/notes/introduction-to-machine-intelligence.htmlnotesData Trusts SalonThu, 17 Dec 2020 00:00:00 +0000
http://inverseprobability.com/talks/notes/ostrom-workshop.html
http://inverseprobability.com/talks/notes/ostrom-workshop.htmlnotesWhen Scientists Work with Government<ul> <li><p>The Delve initiative is a group that was convened by the Royal Society to help provide data-driven insights about the pandemic, with an initial focus on exiting the first lockdown and particular interest in using the variation of strategies across different international governments to inform policy.</p></li> <li><p>Drawing from a multidisciplinary team of domain experts in policy, public health, economics, education, immunology, epidemiology, and social science, alongside statisticians, mathematicians, computer scientists and machine learning scientists, DELVE set out to provide advice and analysis that could feed into live policy decisions.</p></li> <li><p>The main philosophy of the Delve group was to follow the “Supply Chain of Ideas”, connecting scientific evidence to policy questions.</p></li> </ul>Wed, 16 Dec 2020 00:00:00 +0000
http://inverseprobability.com/talks/notes/when-scientists-work-with-government.html
http://inverseprobability.com/talks/notes/when-scientists-work-with-government.htmlnotesAutoAI: Systems, Machine Learning and Mathematics<p>Deployed artificial intelligence solutions consist of interacting components often trained as the result of <em>supervised machine learning</em>. Automatic training of these sub-components is known as AutoML. But the real world challenges of deployment consist of the monitoring of system performance in the real world, in terms of accuracy but also for fairness and bias. To make such systems easily maintainable there is a need for automation of the process of monitoring and redeploying models as well as checking the quality of the overall system decomposition. In contrast to AutoML, we call this system-wide approach “Auto AI”. This is the subject of my Turing Fellowship</p>Wed, 16 Dec 2020 00:00:00 +0000
http://inverseprobability.com/talks/notes/auto-ai-systems-machine-learning-and-mathematics.html
http://inverseprobability.com/talks/notes/auto-ai-systems-machine-learning-and-mathematics.htmlnotesAccelerate-Spark Information SessionFri, 04 Dec 2020 00:00:00 +0000
http://inverseprobability.com/talks/notes/accelerate-overview.html
http://inverseprobability.com/talks/notes/accelerate-overview.htmlnotesAuto AI and Machine Learning Systems Design<p>It seems that we would like to design artificial intelligences, robust decision-making systems that understand the broader context of the decisions they are making, including the history and nature of human experience. At least, that is what the global hype around artificial intelligence implies we are doing. The reality is very different. In practice, we are designing and deploying data-driven decision-making systems within complex software systems with little to no understanding of the downstream implications. At the heart of the challenge is standard practice around the design and construction of modern, complex, software systems. In particular, we have resolved the challenge of the mythical person-month through separation of concerns. Decomposition of the task into separate entities, each of which has defined input and outputs and each of which is normally developed and/or maintained by a single software team. The challenge with such large-scale software systems is that they have incredible complexity. Separation of concerns enables us to deal with such complexity with a decomposition of components. Unfortunately, this means that no team is ‘concerned’ with the overall operation of this system. Modern artificial intelligence is based on machine learning algorithms. In deployment these become components of the larger system that make decisions through observing historic data around those decisions and emulating those decisions through fitting mathematical functions to the data. The field of machine learning is closely related to statistics, but in contrast to statistics, less emphasis has traditionally been placed on the interpretability of model outputs or the validity of decisions in the sense of some form of ‘statistical truth’. This released the field from the constraints of the simpler models that statisticians have typically focussed on, but the success of these models has triggered a wave of head scratching around the fairness, explainability and transparency of such models (FET models). FET models are an active area of machine learning research with their own conference. The challenge we are interested in is deeper: FET systems. When separation of concerns has been deployed, even if an individual model is FET then there is no guarantee that the entire system of interacting components will be FET. That would require composition of our criteria for fairness, explainability and transparency. Other authors have already pointed out the challenges of <em>technical debt</em> in machine learning systems. Technical debt is the challenge of building systems that are <em>maintainable</em> in production without significant additional labour, but the deeper problem is one of <em>intellectual debt</em>. We are deploying systems that are not <em>explainable</em> in production without deeper significant additional intellectual labour. This presentation is a call for help. We urgently need the expertise of the UK Systems Community around these issues to ensure we can construct safe, maintainable and explainable artificial intelligence solutions through FET systems.</p>Wed, 25 Nov 2020 00:00:00 +0000
http://inverseprobability.com/talks/notes/auto-ai-and-machine-learning-systems-design.html
http://inverseprobability.com/talks/notes/auto-ai-and-machine-learning-systems-design.htmlnotesPolicy, Science and the Convening Power of Data<p>With the world watching case numbers increase and publics and policymakers scrutinising projections from epidemiological models, the covid-19 pandemic brought with it increased attention on the use of data to inform policy. Alongside this scrutiny came a new wave of interest in the ability of data and artificial intelligence (AI) to help tackle major scientific and social challenges: could our increasing ability to collect, combine and interrogate large datasets lead to new insights that unlock more effective policy responses? Experiences from the DELVE Initiative, convened to bring data science to bear on covid-19 policy, suggests achieving this aim requires wider adoption of open data science methods to deploy data science and AI expertise and resources to tackle real-world problems.</p>Tue, 24 Nov 2020 00:00:00 +0000
http://inverseprobability.com/talks/notes/policy-science-and-the-convening-power-of-data.html
http://inverseprobability.com/talks/notes/policy-science-and-the-convening-power-of-data.htmlnotesScience, Evidence and Government; Reflections on the Covid-19 Experience<p>This high-profile event brings together four senior academics from across the University of Cambridge who have all been advising policy-makers during the Covid-19 pandemic. The speakers will draw on their extensive experience of advising and being consulted by policy-makers, and will reflect on some of the lessons, debates and controversies associated with governmental responses to the pandemic. And they will consider what this episode tells us about the relationship between science, evidence and public policy in times of crisis.</p>Tue, 10 Nov 2020 00:00:00 +0000
http://inverseprobability.com/talks/notes/science-evidence-and-government-reflections-on-the-covid-19-experience.html
http://inverseprobability.com/talks/notes/science-evidence-and-government-reflections-on-the-covid-19-experience.htmlnotesData Sharing and Data Trusts<p>Computational biologists know better than perhaps any other domain the importance of data sharing in progress in understanding complex decisions. Underlying the revolution in “artificial intelligence” is really a revolution in data. But when data is persona or has legal protections placed upon there are challenges to data sharing. In this talk we introduce the ideas behind data sharing and the model of data trusts, an approach to data sharing that relies on trust law to form its governance structure.</p>Tue, 20 Oct 2020 00:00:00 +0000
http://inverseprobability.com/talks/notes/data-sharing-and-data-trusts.html
http://inverseprobability.com/talks/notes/data-sharing-and-data-trusts.htmlnotesDeploying Machine Learning: Intellectual Debt and AutoAI<p>From the dawn of cybernetics, and across the last eight decades, we’ve worked to make machine learning methods successful. But now that these methods are being widely adopted we need to deal with the consequences of success. Many of those consequences can only be understood when a holistic approach to the machine learning problem is considered: the deployment of a method within a context for a particular objective. In this circumstance, it’s easy to see that questions of interpretability, fairness and transparency are all contextual. In this talk we summarize this challenge using Jonathan Zittrain’s term of 'intellectual debt', we discuss how it pans out in reality and how this challenge could be addressed using machine learning techniques to give us 'Auto AI'. This work is sponsored by an ATI Senior AI Fellowship.</p>Tue, 06 Oct 2020 00:00:00 +0000
http://inverseprobability.com/talks/notes/deploying-machine-learning-systems-intellectual-debt-and-auto-ai.html
http://inverseprobability.com/talks/notes/deploying-machine-learning-systems-intellectual-debt-and-auto-ai.htmlnotesAI and Data Science<p>Waves of automation have driven human advance, and each wave requires humans to The promise of AI is to launch new systems of automated intellectual endeavour that will be the first systems to adapt to us. In reality, the systems we have will not achieve this, and it is the biological sciences that teach us this lesson most starkly. In this talk I will review some of the successes and challenges of AI and its deployment and propose practical visions for the future based on approaches that have worked in the past.</p>Tue, 22 Sep 2020 00:00:00 +0000
http://inverseprobability.com/talks/health/ai-and-data-science.html
http://inverseprobability.com/talks/health/ai-and-data-science.htmlhealthWill AI Make the Workplace - Wherever it is - More Equal?<p>COVID-19 has brought more flexible working, particularly homeworking, for many. Will those changes be sustained after the pandemic and allow previously excluded workers into the labour market? And how will the artificial intelligence revolution affect the jobs we do and who does them? With Drs Christopher Markou, Helen McCarthy, Neil Lawrence and Stella Pachidi. This event is taking place in partnership with The Hay Festival (<www.hayfestival.com>)</p>Sat, 19 Sep 2020 00:00:00 +0000
http://inverseprobability.com/talks/notes/will-ai-make-the-workplace-wherever-it-is-more-equal.html
http://inverseprobability.com/talks/notes/will-ai-make-the-workplace-wherever-it-is-more-equal.htmlnotesDeep GPs<p>In this talk we introduce deep Gaussian processes, an approach to stochastic process modelling that relies on the composition of individual stochastic proceses.</p>Wed, 16 Sep 2020 00:00:00 +0000
http://inverseprobability.com/talks/notes/deep-gps.html
http://inverseprobability.com/talks/notes/deep-gps.htmlnotesFIT Machine Learning Systems<p>As machine learning becomes more widely deployed, it is important that we understand what we have deployed. There has been a lot of focus in machine learning research on the fairness and interpretability of individual models, but less attention paid to how this fits into a wider machine learning system. In this talk I’ll motivate the importance of fair, interpretable and transparent machine learning systems. I’ll outline the challenges and highlight some of the directions we are considering to address these challenges. This work is sponsored by an Alan Turing Institute Senior AI Fellowship.</p>Tue, 15 Sep 2020 00:00:00 +0000
http://inverseprobability.com/talks/notes/fit-machine-learning-systems.html
http://inverseprobability.com/talks/notes/fit-machine-learning-systems.htmlnotesIntroduction to Machine Learning SystemsThis notebook introduces some of the challenges of building machine learning data systems. It will introduce you to concepts around joining of databases together. The storage and manipulation of data is at the core of machine learning systems and data science. The goal of this notebook is to introduce the reader to these concepts, not to authoritatively answer any questions about the state of Nigerian health facilities or Covid19, but it may give you ideas about how to try and do that in your own country.Fri, 24 Jul 2020 00:00:00 +0000
http://inverseprobability.com/talks/notes/ml-systems.html
http://inverseprobability.com/talks/notes/ml-systems.htmlnotesOpen Challenges for Automated Machine Learning: Solving Intellectual Debt with AutoAIMachine learning models are deployed as part of wider systems where outputs of one model are consumed by other models. This composite structure for machine learning systems is the dominant approach for deploying artificial intelligence. Such deployed systems can be complex to understand, they bring with them intellectual debt. In this talk we’ll argue that the next frontier for automated machine learning is to move to automation of the systems design, going from AutoML to AutoAI.Sat, 18 Jul 2020 00:00:00 +0000
http://inverseprobability.com/talks/notes/open-challenges-for-auto-ml-solving-intellectual-debt-with-auto-ai.html
http://inverseprobability.com/talks/notes/open-challenges-for-auto-ml-solving-intellectual-debt-with-auto-ai.htmlnotesFuture of AI and Machine Learning<p>Machine learning technologies have driven a revolution in artificial intelligence. Our machines are now able to identify objects in images, transcribe spoke language, translate between languages and even generate text of their own. In this talk we consider what this means for the future of AI and our own intelligence with a particular focus on what the opportunities and pitfalls for businesses are.</p>Wed, 10 Jun 2020 00:00:00 +0000
http://inverseprobability.com/talks/notes/future-of-ai-and-machine-learning.html
http://inverseprobability.com/talks/notes/future-of-ai-and-machine-learning.htmlnotesThe Great AI Fallacy<p>Artificial intelligence is a form of intellectual automation. The promise of artificial intelligence is that it will be the first generation of automation that adapts to humans, rather than humans having to adapt to it. I see no evidence that this is true, but this fallacy is having very real effects on the way we think about creating and deploying artificial intelligence solutions. In this talk I introduce the Great AI Fallacy and discuss strategies for deployment that pre-emptively deal with the problems it will trigger.</p>Tue, 21 Apr 2020 00:00:00 +0000
http://inverseprobability.com/talks/notes/the-great-ai-fallacy.html
http://inverseprobability.com/talks/notes/the-great-ai-fallacy.htmlnotesIntellectual Debt and the Death of the ProgrammerTechnical debt is incurred when complex systems are rapidly deployed without due thought as to how they will be <em>maintained</em>. Intellectual debt is incurred when complex systems are rapidly deployed without due thought to how they’ll be <em>explained</em>. Both problems are pervasive in the design and deployment of large scale algorithmic decision making engines. In this talk we’ll review the origin of the problem, and propose a roadmap for obtaining solutions. It’s a journey that will require collaboration between industry, academia, third sector, and government.Mon, 09 Mar 2020 00:00:00 +0000
http://inverseprobability.com/talks/notes/intellectual-debt-and-the-death-of-the-programmer-bbc.html
http://inverseprobability.com/talks/notes/intellectual-debt-and-the-death-of-the-programmer-bbc.htmlnotesIntellectual Debt and the Death of the ProgrammerTechnical debt is incurred when complex systems are rapidly deployed without due thought as to how they will be <em>maintained</em>. Intellectual debt is incurred when complex systems are rapidly deployed without due thought to how they’ll be <em>explained</em>. Both problems are pervasive in the design and deployment of large scale algorithmic decision making engines. In this talk we’ll review the origin of the problem, and propose a roadmap for obtaining solutions. It’s a journey that will require collaboration between industry, academia, third sector, and government.Fri, 14 Feb 2020 00:00:00 +0000
http://inverseprobability.com/talks/notes/intellectual-debt-and-the-death-of-the-programmer.html
http://inverseprobability.com/talks/notes/intellectual-debt-and-the-death-of-the-programmer.htmlnotesR250: GP IntroIn this talk we give an introduction to Gaussian processes for students who are interested in working with GPs for the the R250 module.Fri, 24 Jan 2020 00:00:00 +0000
http://inverseprobability.com/talks/notes/r250-gp-intro.html
http://inverseprobability.com/talks/notes/r250-gp-intro.htmlnotesCommunication and Remote WorkingThu, 23 Jan 2020 00:00:00 +0000
http://inverseprobability.com/talks/notes/communication-and-remote-working.html
http://inverseprobability.com/talks/notes/communication-and-remote-working.htmlnotesCoconut Science and the Supply Chain of Ideas<p>Ideas help companies innovate. Different businesses have different approaches to innovation. Some companies centralise their innovation, other companies deploy scientists close to the business. There are two types of business, those where the demand for ideas is driven by customer needs (customer led), and those where ideas are being imposed by a business on the population (technology led). The focus in companies is on the generation of ideas, but this is an error. The focus should be on the supply chain of ideas. The process by which ideas are translated from their point of origin to solving a business task.</p>Wed, 22 Jan 2020 00:00:00 +0000
http://inverseprobability.com/talks/notes/coconut-science-and-the-supply-chain-of-ideas.html
http://inverseprobability.com/talks/notes/coconut-science-and-the-supply-chain-of-ideas.htmlnotesMachine Learning and Emergency MedicineThu, 09 Jan 2020 00:00:00 +0000
http://inverseprobability.com/talks/notes/machine-learning-and-emergency-medicine.html
http://inverseprobability.com/talks/notes/machine-learning-and-emergency-medicine.htmlnotesFrom Innovation to DeploymentIn this talk we introduce a five year project funded by the UK’s Turing Institute to shift the focus from developing AI systems to deploying AI systems that are safe and reliable. The AI systems we are developing and deploying are based on interconnected machine learning components. There is a need for AI-assisted design and monitoring of these systems to ensure they perform robustly, safely and accurately in their deployed environment. We address the entire pipeline of AI system development, from data acquisition to decision making. Data Oriented Architectures are an ecosystem that includes system monitoring for performance, interpretability and fairness. The will enable us to move from individual component optimisation to full system monitoring and optimisation.Wed, 04 Dec 2019 00:00:00 +0000
http://inverseprobability.com/talks/notes/from-innovation-to-deployment-turing-2.html
http://inverseprobability.com/talks/notes/from-innovation-to-deployment-turing-2.htmlnotesPerspectives on AIMon, 02 Dec 2019 00:00:00 +0000
http://inverseprobability.com/talks/notes/perspectives-on-ai.html
http://inverseprobability.com/talks/notes/perspectives-on-ai.htmlnotesNaive DaysMon, 02 Dec 2019 00:00:00 +0000
http://inverseprobability.com/talks/notes/guest-lecture.html
http://inverseprobability.com/talks/notes/guest-lecture.htmlnotesReal World Machine Learning ChallengesMachine learning solutions, in particular those based on deep learning methods, form an underpinning of the current revolution in “artificial intelligence” that has dominated popular press headlines and is having a significant influence on the wider tech agenda. In this talk I will give an overview of where we are now with machine learning solutions, and what challenges we face both in the near and far future. These include practical application of existing algorithms in the face of the need to explain decision making, mechanisms for improving the quality and availability of data, dealing with large unstructured datasets.Thu, 28 Nov 2019 00:00:00 +0000
http://inverseprobability.com/talks/notes/real-world-machine-learning-challenges.html
http://inverseprobability.com/talks/notes/real-world-machine-learning-challenges.htmlnotesDebating IBM’s Project DebaterThu, 21 Nov 2019 00:00:00 +0000
http://inverseprobability.com/talks/notes/debating-project-debater.html
http://inverseprobability.com/talks/notes/debating-project-debater.htmlnotesPost Digital TransformationArtificial intelligence promises automated decision making that will alleviate and revolutionise the nature of work. In practice, we know from previous technological solutions, new technologies often take time to percolate through to productivity. Robert Solow’s paradox saw “computers everywhere, except in the productivity statistics”. This session will equip attendees with an understanding of how to establish best practices around automated decision making. In particular, we will focus on the raw material of the AI revolution: the data.Tue, 19 Nov 2019 00:00:00 +0000
http://inverseprobability.com/talks/notes/post-digital-transformation.html
http://inverseprobability.com/talks/notes/post-digital-transformation.htmlnotesR250: GP IntroIn this talk we give an introduction to Gaussian processes for students who are interested in working with GPs for the the R250 module.Thu, 14 Nov 2019 00:00:00 +0000
http://inverseprobability.com/talks/notes/r250-gp-intro.html
http://inverseprobability.com/talks/notes/r250-gp-intro.htmlnotesWhat is Artificial Intelligence?In this talk we give an introduction to what artificial intelligence technologies are doing today and how they are influencing business and society.Mon, 11 Nov 2019 00:00:00 +0000
http://inverseprobability.com/talks/notes/what-is-artificial-intelligence.html
http://inverseprobability.com/talks/notes/what-is-artificial-intelligence.htmlnotesData First CultureArtificial intelligence promises automated decision making that will alleviate and revolutionise the nature of work. In practice, we know from previous technological solutions, new technologies often take time to percolate through to productivity. Robert Solow’s paradox saw “computers everywhere, except in the productivity statistics”. This session will equip attendees with an understanding of how to establish best practices around automated decision making. In particular, we will focus on the raw material of the AI revolution: the data.Thu, 07 Nov 2019 00:00:00 +0000
http://inverseprobability.com/talks/notes/data-first-culture.html
http://inverseprobability.com/talks/notes/data-first-culture.htmlnotesMachine Learning Systems DesignMachine learning solutions, in particular those based on deep learning methods, form an underpinning of the current revolution in “artificial intelligence” that has dominated popular press headlines and is having a significant influence on the wider tech agenda. In this talk I will give an overview of where we are now with machine learning solutions, and what challenges we face both in the near and far future. These include practical application of existing algorithms in the face of the need to explain decision-making, mechanisms for improving the quality and availability of data, dealing with large unstructured datasets.Tue, 05 Nov 2019 00:00:00 +0000
http://inverseprobability.com/talks/notes/machine-learning-systems-design-cambridge-ai-group-seminar.html
http://inverseprobability.com/talks/notes/machine-learning-systems-design-cambridge-ai-group-seminar.htmlnotesAutoAIDeployed artificial intelligence solutions consist of interacting components often trained as the result of <em>supervised machine learning</em>. Automatic training of these sub-components is known as AutoML. But the real world challenges of deployment consist of the monitoring of system performance in the real world, in terms of accuracy but also for fairness and bias. To make such systems easily maintainable there is a need for automation of the process of monitoring and redeploying models as well as checking the quality of the overall system decomposition. In contrast to AutoML, we call this system-wide approach “Auto AI”. This is the subject of my Turing FellowshipWed, 30 Oct 2019 00:00:00 +0000
http://inverseprobability.com/talks/notes/auto-ai.html
http://inverseprobability.com/talks/notes/auto-ai.htmlnotesFrom Data Subject to Data CitizenResolutely complementary to top-down regulation, bottom-up data trusts aim to ‘give a voice’ to data subjects whose choices when it comes to data governance are often reduced to binary, ill-informed consent. While the rights granted by instruments like the GDPR can be used as tools in a bit to shape possible data-reliant futures - such as better use of natural resources, medical care etc., their exercise is both demanding and unlikely to be as impactful when leveraged individually. The power that stems from aggregated data should be returned to individuals through the legal mechanism of trusts.Mon, 28 Oct 2019 00:00:00 +0000
http://inverseprobability.com/talks/notes/from-data-subject-to-data-citizen.html
http://inverseprobability.com/talks/notes/from-data-subject-to-data-citizen.htmlnotesFrom Innovation to DeploymentIn this talk we introduce a five year project funded by the UK’s Turing Institute to shift the focus from developing AI systems to deploying AI systems that are safe and reliable. The AI systems we are developing and deploying are based on interconnected machine learning components. There is a need for AI-assisted design and monitoring of these systems to ensure they perform robustly, safely and accurately in their deployed environment. We address the entire pipeline of AI system development, from data acquisition to decision making. Data Oriented Architectures are an ecosystem that includes system monitoring for performance, interpretability and fairness. The will enable us to move from individual component optimisation to full system monitoring and optimisation.Thu, 24 Oct 2019 00:00:00 +0000
http://inverseprobability.com/talks/notes/from-innovation-to-deployment.html
http://inverseprobability.com/talks/notes/from-innovation-to-deployment.htmlnotesWhat is Machine Learning?In this talk we will introduce the fundamental ideas in machine learning. We’ll develop our exposition around the ideas of prediction function and the objective function. We don’t so much focus on the derivation of particular algorithms, but more the general principles involved to give an idea of the machine learning <em>landscape</em>.Mon, 21 Oct 2019 00:00:00 +0000
http://inverseprobability.com/talks/notes/what-is-machine-learning-ashesi.html
http://inverseprobability.com/talks/notes/what-is-machine-learning-ashesi.htmlnotesThe Future of AIWaves of automation have driven human advance, and each wave requires humans to The promise of AI is to launch new systems of automated intellectual endeavour that will be the first systems to adapt to us. In reality, the systems we have will not achieve this, and it is the biological sciences that teach us this lesson most starkly. In this talk I will review some of the successes and challenges of AI and its deployment and propose practical visions for the future based on approaches that have worked in the past.Thu, 26 Sep 2019 00:00:00 +0000
http://inverseprobability.com/talks/health/the-future-of-ai.html
http://inverseprobability.com/talks/health/the-future-of-ai.htmlhealthMachine Learning Systems DesignMachine learning solutions, in particular those based on deep learning methods, form an underpinning of the current revolution in “artificial intelligence” that has dominated popular press headlines and is having a significant influence on the wider tech agenda. In this talk I will give an overview of where we are now with machine learning solutions, and what challenges we face both in the near and far future. These include practical application of existing algorithms in the face of the need to explain decision-making, mechanisms for improving the quality and availability of data, dealing with large unstructured datasets.Fri, 20 Sep 2019 00:00:00 +0000
http://inverseprobability.com/talks/notes/machine-learning-systems-design.html
http://inverseprobability.com/talks/notes/machine-learning-systems-design.htmlnotesIntroduction to Deep Gaussian ProcessesIn this talk we introduce deep Gaussian processes, describe what they are and what they are good for. Deep Gaussian process models make use of stochastic process composition to combine Gaussian processes together to form new models which are non-Gaussian in structure. They serve both as a theoretical model for deep learning and a functional model for regression, classification and unsupervised learning. The challenge in these models is propagating the uncertainty through the process.Tue, 10 Sep 2019 00:00:00 +0000
http://inverseprobability.com/talks/notes/introduction-to-deep-gps.html
http://inverseprobability.com/talks/notes/introduction-to-deep-gps.htmlnotesInterpretable End-to-End LearningPractical artificial intelligence systems can be seen as algorithmic decision makers. The fractal nature of decision making implies that this involves interacting systems of components where decisions are made multiple times across different time frames. This affects the decomposability of an artificial intelligence system. Classical systems design relies on decomposability for efficient maintenance and deployment of machine learning systems, in this talk we consider the challenges of optimizing and maintaining such systems.Wed, 26 Jun 2019 00:00:00 +0000
http://inverseprobability.com/talks/notes/interpretable-end-to-end-learning.html
http://inverseprobability.com/talks/notes/interpretable-end-to-end-learning.htmlnotesMachine Learning and Data ScienceMachine learning solutions, in particular those based on deep learning methods, form an underpinning of the current revolution in “artificial intelligence” that has dominated popular press headlines and is having a significant influence on the wider tech agenda. In this talk I will give an overview of where we are now with machine learning solutions, and what challenges we face both in the near and far future. These include practical application of existing algorithms in the face of the need to explain decision making, mechanisms for improving the quality and availability of data, dealing with large unstructured datasets.Wed, 19 Jun 2019 00:00:00 +0000
http://inverseprobability.com/talks/notes/machine-learning-and-data-science.html
http://inverseprobability.com/talks/notes/machine-learning-and-data-science.htmlnotesMachine Learning Systems DesignMachine learning solutions, in particular those based on deep learning methods, form an underpinning of the current revolution in “artificial intelligence” that has dominated popular press headlines and is having a significant influence on the wider tech agenda. In this talk I will give an overview of where we are now with machine learning solutions, and what challenges we face both in the near and far future. These include practical application of existing algorithms in the face of the need to explain decision-making, mechanisms for improving the quality and availability of data, dealing with large unstructured datasets.Thu, 06 Jun 2019 00:00:00 +0000
http://inverseprobability.com/talks/notes/the-three-ds-of-machine-learning.html
http://inverseprobability.com/talks/notes/the-three-ds-of-machine-learning.htmlnotesWhat is Machine Learning?In this talk we will introduce the fundamental ideas in machine learning. We’ll develop our exposition around the ideas of prediction function and the objective function. We don’t so much focus on the derivation of particular algorithms, but more the general principles involved to give an idea of the machine learning <em>landscape</em>.Mon, 03 Jun 2019 00:00:00 +0000
http://inverseprobability.com/talks/notes/what-is-machine-learning.html
http://inverseprobability.com/talks/notes/what-is-machine-learning.htmlnotesNarrowing the Intelligence GapHow are we making computers do the things we used to associate only with humans? Have we made a breakthrough in understanding human intelligence? While recent achievements might give the sense that the answer is yes, the short answer is that we are nowhere near. All we’ve achieved for the moment is a breakthrough in emulating intelligence. In this talk we discuss two differences between the artificial intelligence we’ve deployed and the natural intelligence we exhibit. Resolving one is a challenge of changing the way we do systems design, the other, we argue, is a more fundamental difference that may never be overcome.Thu, 30 May 2019 00:00:00 +0000
http://inverseprobability.com/talks/notes/narrowing-the-intelligence-gap.html
http://inverseprobability.com/talks/notes/narrowing-the-intelligence-gap.htmlnotesMeta-Modelling and Deploying ML SoftwareData is not so much the new oil, it is the new software. Data driven algorithms are increasingly present in continuously deployed production software. What challenges does this present and how can the mathematical sciences help?Thu, 23 May 2019 00:00:00 +0000
http://inverseprobability.com/talks/notes/meta-modelling-and-deploying-ml-software.html
http://inverseprobability.com/talks/notes/meta-modelling-and-deploying-ml-software.htmlnotesModern Data Oriented ProgrammingThere has been a great deal of interest in probabilistic programs: placing modeling at the heart of programming language. In this talk we set the scene for data oriented programming. Data is a fundamental component of machine learning, yet the availability, quality and discoverability of data are often ignored in formal computer science. While languages for data manipulation exist (for example SQL), they are not suitable for the modern world of machine learning data. Modern data oriented languages should place data at the center of modern digital systems design and provide an infrastructure in which monitoring of data quality and model decision making are automaticaly available. We provide the context for Modern Data Oriented Programming, and give some insight into our initial ideas in this space.Tue, 21 May 2019 00:00:00 +0000
http://inverseprobability.com/talks/notes/modern-data-oriented-programming.html
http://inverseprobability.com/talks/notes/modern-data-oriented-programming.htmlnotesWhat is AI and What are the Implications of Advances in AI for Religion?What is artificial intelligence and what are the implications of advances in artificial intelligence for religion? How does artificial intelligences differ from natural intelligences. We consider these ideas from the perspective of information theory. In the context of these differences we then consider parallels between the perspectives on religion and AI both in today’s popular culture, but also with a more optimistic perspective looking forward.Fri, 17 May 2019 00:00:00 +0000
http://inverseprobability.com/talks/notes/what-is-ai-and-what-are-the-implications-of-advances-in-ai-for-religion.html
http://inverseprobability.com/talks/notes/what-is-ai-and-what-are-the-implications-of-advances-in-ai-for-religion.htmlnotesTowards Machine Learning Systems DesignMachine learning solutions, in particular those based on deep learning methods, form an underpinning for the modern artificial intelligence revolution that has dominated popular press headlines and is having a strong influence on the wider tech agenda. In this talk I will give an overview of where we are now with machine learning solutions, and what challenges we face both in the near and far future. Many of these lessons were first formed in computational biology, throughout the talk I’ll highlight connections I see, emphasizing the relevance of biological data analysis to real world data analysis.Tue, 14 May 2019 00:00:00 +0000
http://inverseprobability.com/talks/genomics/towards-ml-systems-design-lessons-from-comp-bio.html
http://inverseprobability.com/talks/genomics/towards-ml-systems-design-lessons-from-comp-bio.htmlgenomicsDigital DisruptionWe look towards the future of digital disruption by considering the past of disruption, with a particular focus on the production and movement of goods. We introduce the notion of the ‘smith’, and consider how, by localizing the provision, or supply, a ‘smith’ can ensure high added value for their skills. Using analogies from <em>pull</em> and <em>push</em> supply chains, We argue that our future economy needs to include an environment where <em>smiths</em> prosper. From craft coffee to craft software, to add value in a global marketplace we argue that we need to exploit localization.Mon, 13 May 2019 00:00:00 +0000
http://inverseprobability.com/talks/notes/digital-disruption.html
http://inverseprobability.com/talks/notes/digital-disruption.htmlnotesData Readiness LevelsIn this talk we consider data readiness levels and how they may be deployed.Wed, 01 May 2019 00:00:00 +0000
http://inverseprobability.com/talks/notes/data-readiness-levels.html
http://inverseprobability.com/talks/notes/data-readiness-levels.htmlnotesFaith and AI: Introduction to Machine LearningWhat is artificial intelligence and what are the implications of advances in artificial intelligence for religion? In this talk we give a short introduction to the technology that's underpinning advances in artificial intelligence, machine learning. We then develop those ideas with a particular focus on how artificial intelligences differ from <em>natural</em> intelligences. Next, we consider parallels between the perspectives on religion and AI in popular culture, initially with a 'cartoon view', but then diving deeper and reflecting on the shared drive for introspection that a mature approach to artificial intelligence and religion might bring.Fri, 29 Mar 2019 00:00:00 +0000
http://inverseprobability.com/talks/notes/faith-and-ai-introduction-to-machine-learning.html
http://inverseprobability.com/talks/notes/faith-and-ai-introduction-to-machine-learning.htmlnotesData Readiness Levels<p>In this brief talk we motivate Data Readiness Levels, an attempt to develop a language around data quality that can bridge the gap between technical solutions and decision makers such as managers and project planners.</p>Mon, 25 Feb 2019 00:00:00 +0000
http://inverseprobability.com/talks/notes/data-readiness-levels.html
http://inverseprobability.com/talks/notes/data-readiness-levels.htmlnotesTowards Machine Learning Systems DesignMachine learning solutions, in particular those based on deep learning methods, form an underpinning for the modern artificial intelligence revolution that has dominated popular press headlines and is having a strong influence on the wider tech agenda. In this talk I will give an overview of where we are now with machine learning solutions, and what challenges we face both in the near and far future. These include practical application of existing algorithms in the face of the need to explain decision making, mechanisms for improving the quality and availability of data, and dealing with large unstructured datasets.Fri, 22 Feb 2019 00:00:00 +0000
http://inverseprobability.com/talks/glasgow2019/towards-ml-systems-design.html
http://inverseprobability.com/talks/glasgow2019/towards-ml-systems-design.htmlglasgow2019Data Science and Digital SystemsMachine learning solutions, in particular those based on deep learning methods, form an underpinning of the current revolution in “artificial intelligence” that has dominated popular press headlines and is having a significant influence on the wider tech agenda. In this talk I will give an overview of where we are now with machine learning solutions, and what challenges we face both in the near and far future. These include practical application of existing algorithms in the face of the need to explain decision making, mechanisms for improving the quality and availability of data, dealing with large unstructured datasets.Tue, 19 Feb 2019 00:00:00 +0000
http://inverseprobability.com/talks/notes/data-science-and-digital-systems.html
http://inverseprobability.com/talks/notes/data-science-and-digital-systems.htmlnotesDeep Gaussian Processes<p>Gaussian process models provide a flexible, non-parametric approach to modelling that sustains uncertainty about the function. However, computational demands and the joint Gaussian assumption make them inappropriate for some applications. In this talk we review low rank approximations for Gaussian processes and use stochastic process composition to create non-Gaussian processes. We illustrate the models on simple regression tasks to give a sense of how uncertainty propagates through the model. We end will demonstrations on unsupervised learning of digits and motion capture data.</p>Fri, 11 Jan 2019 00:00:00 +0000
http://inverseprobability.com/talks/notes/deep-gaussian-processes.html
http://inverseprobability.com/talks/notes/deep-gaussian-processes.htmlnotesGaussian Processes<p>Classical machine learning and statistical approaches to learning, such as neural networks and linear regression, assume a parametric form for functions. Gaussian process models are an alternative approach that assumes a probabilistic prior over functions. This brings benefits, in that uncertainty of function estimation is sustained throughout inference, and some challenges: algorithms for fitting Gaussian processes tend to be more complex than parametric models. In this sessions I will introduce Gaussian processes and explain why sustaining uncertainty is important.</p>Wed, 09 Jan 2019 00:00:00 +0000
http://inverseprobability.com/talks/notes/gaussian-processes.html
http://inverseprobability.com/talks/notes/gaussian-processes.htmlnotesMachine Learning and the Physical World<p>Machine learning is a data driven endeavour, but real world systems are physical and mechanistic. In this talk we will review approaches to integrating machine learning with real world systems. Our focus will be on emulation (otherwise known as surrogate modeling).</p>Mon, 10 Dec 2018 00:00:00 +0000
http://inverseprobability.com/talks/notes/machine-learning-and-the-physical-world.html
http://inverseprobability.com/talks/notes/machine-learning-and-the-physical-world.htmlnotesData Science and Digital SystemsMachine learning solutions, in particular those based on deep learning methods, form an underpinning of the current revolution in “artificial intelligence” that has dominated popular press headlines and is having a significant influence on the wider tech agenda. In this talk I will give an overview of where we are now with machine learning solutions, and what challenges we face both in the near and far future. These include practical application of existing algorithms in the face of the need to explain decision making, mechanisms for improving the quality and availability of data, dealing with large unstructured datasets.Fri, 30 Nov 2018 00:00:00 +0000
http://inverseprobability.com/talks/notes/data-science-and-digital-systems.html
http://inverseprobability.com/talks/notes/data-science-and-digital-systems.htmlnotesThe Three Ds of Machine Learning<p>Machine learning solutions, in particular those based on deep learning methods, form an underpinning of the current revolution in “artificial intelligence” that has dominated popular press headlines and is having a significant influence on the wider tech agenda. In this talk I will give an overview of where we are now with machine learning solutions, and what challenges we face both in the near and far future. These include practical application of existing algorithms in the face of the need to explain decision making, mechanisms for improving the quality and availability of data, dealing with large unstructured datasets.</p>Thu, 15 Nov 2018 00:00:00 +0000
http://inverseprobability.com/talks/notes/the-three-ds-of-machine-learning.html
http://inverseprobability.com/talks/notes/the-three-ds-of-machine-learning.htmlnotesBayesian MethodsIn this session we review the <em>probabilistic</em> approach to machine learning. We start with a review of probability, and introduce the concepts of probabilistic modelling. We then apply the approach in practice to Naive Bayesian classification. In this session we review the probabilistic formulation of a classification model, reviewing initially maximum likelihood and the naive Bayes model.Wed, 14 Nov 2018 00:00:00 +0000
http://inverseprobability.com/talks/notes/bayesian-methods-abuja.html
http://inverseprobability.com/talks/notes/bayesian-methods-abuja.htmlnotesFairness and Diversity of Decision Making<p>Mathematical definitions of fairness insist on clearly categorized groups and clear mathematical interpretations of fairness. In law this arises through the concept of <em>unlawful</em> descrimination. There is no such thing as a correct model. We must accept that our predictions will sometimes be wrong. In the face of this certainty we have a choice: how we should be wrong. We can choose to be wrong by over-simplifying or we can choose to be wrong by over-complicating (given the available data). In machine learning this is known as the bias-variance dilemma. In this talk we consider the implications of the bias-variance dilemma for fairness of decision making.</p>Thu, 08 Nov 2018 00:00:00 +0000
http://inverseprobability.com/talks/notes/fairness-and-diversity-of-decision-making.html
http://inverseprobability.com/talks/notes/fairness-and-diversity-of-decision-making.htmlnotesMachine Learning and the Physical World<p>Machine learning is a data driven endeavour, but real world systems are physical and mechanistic. In this talk we will review approaches to integrating machine learning with real world systems. Our focus will be on emulation (otherwise known as surrogate modeling).</p>Tue, 06 Nov 2018 00:00:00 +0000
http://inverseprobability.com/talks/notes/machine-learning-and-the-physical-world.html
http://inverseprobability.com/talks/notes/machine-learning-and-the-physical-world.htmlnotesMind and Machine IntelligenceWhat is the nature of machine intelligence and how does it differ from humans? In this talk we introduce embodiment factors. They represent the extent to which our intelligence is locked inside us. The locked in nature of our intelligence makes us fundamentally different from the machine intelligences we are creating around us. Having summarized these differences we consider the Three Ds of machine learning system design: a set of considerations to take into acount when building machine intelligences.Tue, 30 Oct 2018 00:00:00 +0000
http://inverseprobability.com/talks/notes/mind-and-machine-intelligence.html
http://inverseprobability.com/talks/notes/mind-and-machine-intelligence.htmlnotesYou and AI Panel DebateSun, 28 Oct 2018 00:00:00 +0000
http://inverseprobability.com/talks/you-and-ai.html
http://inverseprobability.com/talks/you-and-ai.htmlNatural and Artificial IntelligenceWhat is the nature of machine intelligence and how does it differ from humans? In this talk we explore embodiment factors, the extent to which our intelligence is locked in and how this makes us fundamentally different form the machine intelligences we are creating around us.Thu, 18 Oct 2018 00:00:00 +0000
http://inverseprobability.com/talks/notes/natural-and-artificial-intelligence.html
http://inverseprobability.com/talks/notes/natural-and-artificial-intelligence.htmlnotesfAIthWhat is artificial intelligence and what are the implications of advances in artificial intelligence for society? In this talk we give a short introduction to the technology that’s underpinning advances in artificial intelligence, machine learning. We then develop those ideas with a particular focus on how artificial intelligences differ from <em>natural</em> intelligences. Finally, we reflect on what the existence of different intelligences might mean for our experiences as humans.Wed, 12 Sep 2018 00:00:00 +0000
http://inverseprobability.com/talks/notes/faith.html
http://inverseprobability.com/talks/notes/faith.htmlnotesData Science and the ProfessionsMachine learning methods and software are becoming widely deployed. But as we deploy algorithms that operate on individual data, how do we account for their effect on society? In terms of the practice of data science, we seem to be at a similar point today as software engineering was in the early 1980s. Best practice is not widely understood or deployed. One aspect of professions is trust. How can we bring trust to the data-sphere?Wed, 05 Sep 2018 00:00:00 +0000
http://inverseprobability.com/talks/notes/data-science-and-the-professions.html
http://inverseprobability.com/talks/notes/data-science-and-the-professions.htmlnotesIntroduction to Gaussian Processes<p>In this talk we introduce Gaussian process models. Motivating the representation of uncertainty through probability distributions we review Laplace's approach to understanding uncertainty and how uncertainty in functions can be represented through a multivariate Gaussian density.</p>Mon, 03 Sep 2018 00:00:00 +0000
http://inverseprobability.com/talks/notes/gpss-session-1.html
http://inverseprobability.com/talks/notes/gpss-session-1.htmlnotesProbabilistic Machine LearningIn this talk we review the <em>probabilistic</em> approach to machine learning. We start with a review of probability, and introduce the concepts of probabilistic modelling. We then apply the approach in practice to Naive Bayesian classification. In this session we review the Bayesian formalism in the context of linear models, reviewing initially maximum likelihood and introducing basis functions as a way of driving non-linearity in the model.Sat, 25 Aug 2018 00:00:00 +0000
http://inverseprobability.com/talks/notes/probabilistic-machine-learning.html
http://inverseprobability.com/talks/notes/probabilistic-machine-learning.htmlnotesMon, 04 Jun 2018 00:00:00 +0000
http://inverseprobability.com/talks/notes/bayesian-methods.html
http://inverseprobability.com/talks/notes/bayesian-methods.htmlnotesFaith and AIWhat is artificial intelligence and what are the implications of advances in artificial intelligence for religion? In this talk we give a short introduction to the technology that’s underpinning advances in artificial intelligence, machine learning. We then develop those ideas with a particular focus on how artificial intelligences differ from <em>natural</em> intelligences. Next, we consider parallels between the perspectives on religion and AI in popular culture, initially with a ‘cartoon view’, but then diving deeper and reflecting on the shared drive for introspection that a mature approach to artificial intelligence and religion might bring.Thu, 31 May 2018 00:00:00 +0000
http://inverseprobability.com/talks/notes/faith-and-ai.html
http://inverseprobability.com/talks/notes/faith-and-ai.htmlnotesUncertainty in Loss FunctionsBayesian formalisms deal with uncertainty in parameters, frequentist formalisms deal with the <em>risk</em> of a data set, uncertainty in the data sample. In this talk, we consider uncertainty in the <em>loss function</em>. Uncertainty in the loss function. We introduce uncertainty through linear weightings of terms in the loss function and show how a distribution over the loss can be maintained through the <em>maximum entropy principle</em>. This allows us minimize the expected loss under our maximum entropy distribution of the loss function. We recover weighted least squares and a LOESS-like regression from the formalism.Tue, 29 May 2018 00:00:00 +0000
http://inverseprobability.com/talks/notes/uncertainty-in-loss-functions.html
http://inverseprobability.com/talks/notes/uncertainty-in-loss-functions.htmlnotesOutlook for UK AI and Machine LearningMachine learning solutions, in particular those based on deep learning methods, form an underpinning of the current revolution in “artificial intelligence” that has dominated popular press headlines and is having a significant influence on the wider tech agenda. In this talk I will give an overview of where we are now with machine learning solutions, and what challenges we face both in the near and far future. These include practical application of existing algorithms in the face of the need to explain decision making, mechanisms for improving the quality and availability of data, dealing with large unstructured datasets.Fri, 11 May 2018 00:00:00 +0000
http://inverseprobability.com/talks/notes/outlook-for-uk-ai-and-ml.html
http://inverseprobability.com/talks/notes/outlook-for-uk-ai-and-ml.htmlnotesTowards Machine Learning Systems DesignWed, 02 May 2018 00:00:00 +0000
http://inverseprobability.com/talks/towards-machine-learning-systems-design.html
http://inverseprobability.com/talks/towards-machine-learning-systems-design.htmlDecision Making and DiversityMon, 30 Apr 2018 00:00:00 +0000
http://inverseprobability.com/talks/notes/decision-making-and-diversity.html
http://inverseprobability.com/talks/notes/decision-making-and-diversity.htmlnotesChallenges for Data Science in HealthcareWed, 18 Apr 2018 00:00:00 +0000
http://inverseprobability.com/talks/challenges-for-data-science-in-healthcare.html
http://inverseprobability.com/talks/challenges-for-data-science-in-healthcare.htmlNatural and Artificial IntelligenceWhat is the nature of machine intelligence and how does it differ from humans? In this talk we explore some of the differences between natural and machine intelligence.Thu, 29 Mar 2018 00:00:00 +0000
http://inverseprobability.com/talks/notes/on-natural-and-artificial-intelligence.html
http://inverseprobability.com/talks/notes/on-natural-and-artificial-intelligence.htmlnotesData Science: Time for Professionalisation?Machine learning methods and software are becoming widely deployed. But how are we sharing expertise about bottlenecks and pain points in deploying solutions? In terms of the practice of data science, we seem to be at a similar point today as software engineering was in the early 1980s. Best practice is not widely understood or deployed. In this talk we will focus on two particular components of data science solutions: the preparation of data and the deployment of machine learning systems.
Tue, 27 Mar 2018 00:00:00 +0000
http://inverseprobability.com/talks/data-science-time-for-professionalisation.html
http://inverseprobability.com/talks/data-science-time-for-professionalisation.htmlMachine Learning and Data Readiness LevelsIn this talk we will look at the challenges facing deployment of machine learning, with a particular focus on the reuse of data and data quality. We suggest data readiness levels as a mechanism for monitoring data quality.Thu, 25 Jan 2018 00:00:00 +0000
http://inverseprobability.com/talks/notes/machine-learning-and-data-readiness-levels.html
http://inverseprobability.com/talks/notes/machine-learning-and-data-readiness-levels.htmlnotesReal World Machine Learning Challenges: Present and FutureMachine learning solutions, in particular those based on deep learning methods, form an underpinning of the current revolution in “artificial intelligence” that has dominated popular press headlines and is having a significant influence on the wider tech agenda.
In this talk I will give an overview of where we are now with machine learning solutions, and what challenges we face both in the near and far future. These include practical application of existing algorithms in the face of the need to explain decision making, mechanisms for improving the quality and availability of data, dealing with large unstructured datasets.
Tue, 23 Jan 2018 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-cms18/real-world-machine-learning-challenges-present-and-future.html
http://inverseprobability.com/talks/lawrence-cms18/real-world-machine-learning-challenges-present-and-future.htmlLawrence-cms18Deep Probabilistic Modelling with with Gaussian Processes<p>Neural network models are algorithmically simple, but mathematically complex. Gaussian process models are mathematically simple, but algorithmically complex. In this tutorial we will explore Deep Gaussian Process models. They bring advantages in their mathematical simplicity but are challenging in their algorithmic complexity. We will give an overview of Gaussian processes and highlight the algorithmic approximations that allow us to stack Gaussian process models: they are based on variational methods. In the last part of the tutorial will explore a use case exemplar: uncertainty quantification. We end with open questions.</p>Mon, 04 Dec 2017 00:00:00 +0000
http://inverseprobability.com/talks/notes/deep-probabilistic-modelling-with-gaussian-processes.html
http://inverseprobability.com/talks/notes/deep-probabilistic-modelling-with-gaussian-processes.htmlnotesPersonalized Health: Challenges in Data ScienceThe promise of personalized health is driven by the wide availability of data, but we don't need to talk so much about where we want to be, rather how we should get there. What are the challenges that need to be bridged technologically to unlock the potential in the much greater availability of data we now have? In this talk we'll consider three challenges of data science in the context of personalized health, the three challenges each need to be bridged to bring the era of true precision, or personalized, medicine within the reach of an affordable health care service.
Thu, 23 Nov 2017 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-cwiml17/personalized-health-challenges-in-data-science.html
http://inverseprobability.com/talks/lawrence-cwiml17/personalized-health-challenges-in-data-science.htmlLawrence-cwiml17Embodiment Factors and PrivacyIn this talk we will explore a fundamental limitation of human intelligence which, we argue, makes privacy absolutely critical. We will relate this to our machine intelligences and speculate about how there may be challenges at the interace. Finally we propose Data Trusts as a solution for these challenges.
Thu, 26 Oct 2017 00:00:00 +0000
http://inverseprobability.com/talks/embodiment-factors-and-privacy.html
http://inverseprobability.com/talks/embodiment-factors-and-privacy.htmlEmbodiment Factors and PrivacyIn this talk we will explore a fundamental limitation of human intelligence which, we argue, makes privacy absolutely critical. We will relate this to our machine intelligences and speculate about how there may be challenges at the interace. Finally we propose Data Trusts as a solution for these challenges.Thu, 26 Oct 2017 00:00:00 +0000
http://inverseprobability.com/talks/notes/embodiment-factors-and-privacy.html
http://inverseprobability.com/talks/notes/embodiment-factors-and-privacy.htmlnotesData Science: Time for Professionalisation?Machine learning methods and software are becoming widely deployed. But how are we sharing expertise about bottlenecks and pain points in deploying solutions? In terms of the practice of data science, we seem to be at a similar point today as software engineering was in the early 1980s. Best practice is not widely understood or deployed. In this talk we will focus on two particular components of data science solutions: the preparation of data and the deployment of machine learning systems.
Fri, 13 Oct 2017 00:00:00 +0000
http://inverseprobability.com/talks/data-science-time-for-professionalisation.html
http://inverseprobability.com/talks/data-science-time-for-professionalisation.htmlLiving TogetherWhat is the nature of machine intelligence and how does it differ from humans? In this talk we explore embodiment factors, the extent to which our intelligence is locked in and how this makes us fundamentally different form the machine intelligences we are creating around us.
Fri, 06 Oct 2017 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-tedx17/living-together.html
http://inverseprobability.com/talks/lawrence-tedx17/living-together.htmlLawrence-tedx17Where Next for AI?Our current generation of artificial intelligence techniques are driven by data. But also we expect to be able to deploy artificial intelligence techniques on data. What does that mean, is it a contradiction? How will this effect the wider technology landscape? Is it simply a matter of refining deep neural nets? Or are more disruptive technologies needed? What will be the challenges of deploying AI systems?Tue, 03 Oct 2017 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-cwtec17/where-next-for-ai.html
http://inverseprobability.com/talks/lawrence-cwtec17/where-next-for-ai.htmlLawrence-cwtec17Introduction to Gaussian ProcessesIn this talk I will give a brief and intuitive introduction to Gaussian process models.
Mon, 11 Sep 2017 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-gpss17/gpss-session-1.html
http://inverseprobability.com/talks/lawrence-gpss17/gpss-session-1.htmlLawrence-gpss17Cloaking Functions: Differential Privacy with Gaussian Processes<p>Processing of personally sensitive information should respect an individual's privacy. One promising framework is Differential Privacy (DP). In this talk I'll present work led by Michael Smith at the University of Sheffield on the use of cloaking functions to make Gaussian process (GP) predictions differentially private. Gaussian process models are flexible models with particular advantages in handling missing and noisy data. Our hope is that advances in DP for GPs will make it easier to 'learn without looking', i.e. gain the advantages of prediction from patient data without impinging on their privacy. Joint work with <strong>Michael T. Smith</strong>, Max Zwiessele and Mauricio Alvarez</p>Wed, 30 Aug 2017 00:00:00 +0000
http://inverseprobability.com/talks/notes/cloaking-functions.html
http://inverseprobability.com/talks/notes/cloaking-functions.htmlnotesWhat is Machine Learning?In this talk we provide an introduction and an overview to the field of machine learning. We emphasise the importance of data and the nature of modelling we carry out in machine learning. We briefly review the different challenges such as supervised, unsupervised and reinforcement learning.
Mon, 17 Jul 2017 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-dsa17/what-is-machine-learning.html
http://inverseprobability.com/talks/lawrence-dsa17/what-is-machine-learning.htmlLawrence-dsa17Once Upon a Universal Standard Time: Embodiment and AI NarrativesIn this talk we consider a fundamental difference between human and machine intelligence, a ratio between their ability to compute and their ability to communicate we refer to as the embodiment factor. Having suggested why this makes us fundamentally different we speculate on implications for developing <em>narrative</em> structure from data.Thu, 13 Jul 2017 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-cfi17/once-upon-a-universal-standard-time.html
http://inverseprobability.com/talks/lawrence-cfi17/once-upon-a-universal-standard-time.htmlLawrence-cfi17Data Analytics Perspectives: Machine LearningIn this talk we will firstly set the landscape of machine learning, artificial intelligence and data science by describing what characteristics they share, and how they differ. We'll then shift focus to the promise and challenges associated with both Data Science and Artficial Intelligence, with particular attention paid to the potential for a "data crisis" and challenges in "machine learning systems design".
Thu, 29 Jun 2017 00:00:00 +0000
http://inverseprobability.com/talks/data-analytics-perspectives.html
http://inverseprobability.com/talks/data-analytics-perspectives.htmlMachine Learning, Technology and the Future of IntelligenceThe Leverhulme Centre for the Future of Intelligence is a fulcrum around which debate in intelligence technology can be joined across the wide range of intereted experts. In this talk I'll give some perspectives on machine learning and my interactions with CFI.
Mon, 26 Jun 2017 00:00:00 +0000
http://inverseprobability.com/talks/machine-learning-technology-and-the-future-of-intelligence.html
http://inverseprobability.com/talks/machine-learning-technology-and-the-future-of-intelligence.htmlProbabilistic Dimensionality ReductionTue, 06 Jun 2017 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-peppercorns17/probabilistic-dimensionality-reduction.html
http://inverseprobability.com/talks/lawrence-peppercorns17/probabilistic-dimensionality-reduction.htmlLawrence-peppercorns17Peppercorns and Machine Learning System DesignMachine learning is fundamental to two important technological domains, artificial intelligence and data science. In this talk we will attempt to make a simple definition to distinguish between the two, then we will focus on the challenges of machine learning in <em>application</em> to artificial intelligence particularly from the perspective of systems design. We expect a particular challenge to be the deployment of such systems in real environment, where unforeseen consequences of interaction with real world environments will produce embarrassing failures. Because these failures are not bugs, in that the system will be performing as designed, but failures of imagination of the designers we introduce a new term for them: ‘peppercorns’.Fri, 02 Jun 2017 00:00:00 +0000
http://inverseprobability.com/talks/notes/peppercorns-and-machine-learning-systems-design.html
http://inverseprobability.com/talks/notes/peppercorns-and-machine-learning-systems-design.htmlnotesPeppercorns and Machine Learning System DesignMachine learning is fundamental to two important technological domains, artificial intelligence and data science. In this talk we will attempt to make a simple definition to distinguish between the two, then we will focus on the challenges of machine learning in *application* to artificial intelligence particularly from the perspective of systems design. We expect a particular challenge to be the deployment of such systems in real environment, where unforeseen consequences of interaction with real world environments will produce embarrassing failures. Because these failures are not bugs, in that the system will be performing as designed, but failures of imagination of the designers we introduce a new term for them: 'peppercorns'.Fri, 02 Jun 2017 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-peppercorns17/peppercorns-and-machine-learning-system-design.html
http://inverseprobability.com/talks/lawrence-peppercorns17/peppercorns-and-machine-learning-system-design.htmlLawrence-peppercorns17The Data Science ProcessIn this talk we will focus on challenges in facilitating the data science pipeline. Drawing on experience from projects in computational biology, the developing world and Amazon I’ll propose different ideas for facilitating the data science process including analogies that help software engineers understand the challenges for data science and formalizations, such as data readiness levels, which allow management to reason about the obstacles in the process.Wed, 10 May 2017 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-dsp17/the-data-science-process.html
http://inverseprobability.com/talks/lawrence-dsp17/the-data-science-process.htmlLawrence-dsp17The Data Science ProcessIn this talk we will focus on challenges in facilitating the data science pipeline. Drawing on experience from projects in computational biology, the developing world and Amazon I’ll propose different ideas for facilitating the data science process including analogies that help software engineers understand the challenges for data science and formalizations, such as data readiness levels, which allow management to reason about the obstacles in the process.Tue, 18 Apr 2017 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-dsp17/the-data-science-process.html
http://inverseprobability.com/talks/lawrence-dsp17/the-data-science-process.htmlLawrence-dsp17Machine Learning and the Data Science ProcessThe current generation of machine learning technologies is powering new applications in artificial intelligence. This is presenting challenges and opportunities. In this talk we will focus on the challenge of constructing and deploying machine learning algorithms with a particular focus on two aspects: machine learning systems design and data readiness. We will also discuss implications and opportunities, with speculative thoughts on the nature of artificial intelligence in future devices and what new opportunities and challenges this may present.Thu, 30 Mar 2017 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-oxwasp17/machine-learning-and-the-data-science-process.html
http://inverseprobability.com/talks/lawrence-oxwasp17/machine-learning-and-the-data-science-process.htmlLawrence-oxwasp17The rise of the algorithm - artificial intelligence, ethics, trust and tech developmentHow do notions of human trust extend to the digital world? And what challenges could that present for our society?Thu, 16 Mar 2017 00:00:00 +0000
http://inverseprobability.com/talks/the-rise-of-the-algorithm.html
http://inverseprobability.com/talks/the-rise-of-the-algorithm.htmlEthics, Computer Systems and the ProfessionsA discussion with Sylvie Delacroix, Jonathan Price, Burkhard Schafer hosted by Anthony Finkelstein.Wed, 15 Mar 2017 00:00:00 +0000
http://inverseprobability.com/talks/ethics-computer-systems-and-the-professions.html
http://inverseprobability.com/talks/ethics-computer-systems-and-the-professions.htmlChallenges and Opportunities in Machine Learning and Artificial IntelligenceThe current generation of machine learning technologies is powering new applications in artificial intelligence. This is presenting challenges and opportunities. In this talk we will focus on the challenge of constructing and deploying machine learning algorithms with a particular focus on two aspects: machine learning systems design and data readiness. We will also discuss implications and opportunities, with speculative thoughts on the nature of artificial intelligence in future devices and what new opportunities and challenges this may present. Mon, 13 Mar 2017 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-arm17/challenges-in-ml-and-data-science.html
http://inverseprobability.com/talks/lawrence-arm17/challenges-in-ml-and-data-science.htmlLawrence-arm17Challenges for Delivering Machine Learning in HealthThe wealth of data availability presents new opportunities
in health but also challenges. In this talk we will focus on
challenges for machine learning in health: 1. Paradoxes of the Data
Society, 2. Quantifying the Value of Data, 3. Privacy, loss of
control, marginalization. Each of these challenges has particular
implications for machine learning. The paradoxes relate to our
evolving relationship with data and our changing
expectations. Quantifying value is vital for accounting for the
influence of data in our new digital economies and issues of privacy
and loss of control are fundamental to how our pre-existing rights
evolve as the digital world encroaches more closely on the
physical. One of the goals of research community should be to
provide the technological tooling to address these challenges ensure
that we are empowered to avoid the pitfalls of the data driven
society, allowing us to reap the benefits of machine learning in
applications from personalized health to health in the developing
world.
Tue, 28 Feb 2017 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-manchester17/challenges-for-delivering-machine-learning-in-health.html
http://inverseprobability.com/talks/lawrence-manchester17/challenges-for-delivering-machine-learning-in-health.htmlLawrence-manchester17Three Challenges in Data ScienceData science presents new opportunities but also new challenges. In this talk we will focus on three separate challenges for data science: 1. Paradoxes of the Data Society, 2. Quantifying the Value of Data, 3. Privacy, loss of control, marginalization. Each of these challenges has particular implications for data science. The paradoxes relate to our evolving relationship with data and our changing expectations. Quantifying value is vital for accounting for the influence of data in our new digital economies and issues of privacy and loss of control are fundamental to how our pre-existing rights evolve as the digital world encroaches more closely on the physical. One of the goals of open data science should be to address these challenges to ensure that we can avoid the pitfalls of the data driven society, allowing us to reap the benefits of data science in applications from personalized health to the developing world.
Tue, 21 Feb 2017 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-manchester17/data-science-challenges.html
http://inverseprobability.com/talks/lawrence-manchester17/data-science-challenges.htmlLawrence-manchester17Latent Variable Models with Gaussian ProcessesGaussian process models are flexible non parametric probabilistic models for functions. In this talk we will show how they can be incorporated into latent variable models to form probabilistic latent variable models. The resulting approaches have some unusual properties. In particular, they express conditional independencies across features, rather than data. This implies that rather than a curse of dimensionality they exhibit a blessing of dimensionality. We will give background of the model and show some exemplar applications.Mon, 06 Feb 2017 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-gplvm17/latent-variable-models-with-gaussian-processes.html
http://inverseprobability.com/talks/lawrence-gplvm17/latent-variable-models-with-gaussian-processes.htmlLawrence-gplvm17Introduction to Gaussian ProcessesIn this master class we will give a short introduction to Gaussian process models, and then explore their use in the domain of Bayesian Optimization. Gaussian process models are flexible models which allow us to place probability distributions over functions. In Bayesian Optimization, the Gaussian process is used as a surrogate for the process of interest. Rather than directly optimizing the process, the surrogate is optimized. This leads to an efficient approach for improving efficiency in a wide range of physical systems. The seminar will introduce lab classes which will make use of the python software GPy and GPyOpt (https://github.com/sheffieldml/GPy, https://github.com/sheffieldml/GPyOpt).
This first talk will be an introduction to Gaussian process models that will assume knowledge of probability, linear algebra and the multivariate Gaussian.Mon, 06 Feb 2017 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-gpbo17/introduction-to-gaussian-processes.html
http://inverseprobability.com/talks/lawrence-gpbo17/introduction-to-gaussian-processes.htmlLawrence-gpbo17Covariance Functions and the Marginal LikelihoodIn this master class we will give a short introduction to Gaussian process models, and then explore their use in the domain of Bayesian Optimization. Gaussian process models are flexible models which allow us to place probability distributions over functions. In Bayesian Optimization, the Gaussian process is used as a surrogate for the process of interest. Rather than directly optimizing the process, the surrogate is optimized. This leads to an efficient approach for improving efficiency in a wide range of physical systems. The seminar will introduce lab classes which will make use of the python software GPy and GPyOpt (https://github.com/sheffieldml/GPy, https://github.com/sheffieldml/GPyOpt).
This talk will develop the idea of the covariance function and give intutions as to how the marginal likelihood can be maximized. Given time we willl also develop the idea of multiple output Gaussian process models.Mon, 06 Feb 2017 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-gpbo17b/covariance-functions-and-the-marginal-likelihood.html
http://inverseprobability.com/talks/lawrence-gpbo17b/covariance-functions-and-the-marginal-likelihood.htmlLawrence-gpbo17bPersonalized Health: Challenges in Data ScienceThe promise of personalized health is driven by the wide availability of data, but we don't need to talk so much about where we want to be, rather how we should get there. What are the challenges that need to be bridged technologically to unlock the potential in the much greater availability of data we now have? In this talk we'll consider three challenges of data science in the context of personalized health, the three challenges each need to be bridged to bring the era of true precision, or personalized, medicine within the reach of an affordable health care service.
Thu, 12 Jan 2017 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-smpgd17/personalized-health-challenges-in-data-science.html
http://inverseprobability.com/talks/lawrence-smpgd17/personalized-health-challenges-in-data-science.htmlLawrence-smpgd17The Data LandscapeIn this talk I'll give an overview of the challenges in the data landscape, both institutional and societal.Thu, 15 Dec 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-defra16/the-data-landscape.html
http://inverseprobability.com/talks/lawrence-defra16/the-data-landscape.htmlLawrence-defra16Personalized Health: Challenges in Data ScienceThe promise of personalized health is driven by the wide availability of data, but we don't need to talk so much about where we want to be, rather how we should get there. What are the challenges that need to be bridged technologically to unlock the potential in the much greater availability of data we now have? In this talk we'll consider three challenges of data science in the context of personalized health, the three challenges each need to be bridged to bring the era of true precision, or personalized, medicine within the reach of an affordable health care service.Fri, 09 Dec 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-ml4hc16b/personalized-health-challenges-in-data-science.html
http://inverseprobability.com/talks/lawrence-ml4hc16b/personalized-health-challenges-in-data-science.htmlLawrence-ml4hc16bComputational Perspectives: Fairness and Awareness in the Analysis of DataWhat is data science? An new name for something old perhaps. Nevertheless there is something new happening. Data is being acquired in ways that coudl never have been envisaged 100 years ago. This is presenting new challenges, and ones that no single field is equipped to face. As well as the need for new methodologies and theoretical underpinnings, modern data processing is having a direct effect on our citizens in real time. In this talk I’ll suggest that data science provides a banner under which the computational and statistical sciences can unite to provide an unified response.Thu, 27 Oct 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-rss16b/computational-perspectives-fairness-and-awareness-in-the-analysis-of-data.html
http://inverseprobability.com/talks/lawrence-rss16b/computational-perspectives-fairness-and-awareness-in-the-analysis-of-data.htmlLawrence-rss16bThree Challenges for Open Data ScienceData science presents new opportunities but also new challenges. In this talk we will focus on three separate challenges for data science: 1. Paradoxes of the Data Society, 2. Quantifying the Value of Data, 3. Privacy, loss of control, marginalization. Each of these challenges has particular implications for data science. The paradoxes relate to our evolving relationship with data and our changing expectations. Quantifying value is vital for accounting for the influence of data in our new digital economies and issues of privacy and loss of control are fundamental to how our pre-existing rights evolve as the digital world encroaches more closely on the physical. One of the goals of open data science should be to address these challenges to ensure that we can avoid the pitfalls of the data driven society, allowing us to reap the benefits of data science in applications from personalized health to the developing world.
Sat, 08 Oct 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-osdc16/three-challenges-for-open-data-science.html
http://inverseprobability.com/talks/lawrence-osdc16/three-challenges-for-open-data-science.htmlLawrence-osdc16The Data DelusionThe widespread success of deep learning in a variety of domains is being hailed as a new revolution in artificial intelligence. It has taken 20 years to go from defeating Kasparov at Chess to Lee Sedol at Go. But what have the real advances been across this time? The fundamental change has been in terms of data availability and compute availability. The underlying technology has not changed much in the last 20 years. So what does that mean for areas like medicine and health? Significant challenges remain, improving the data efficiency of these algorithms and retaining the balance between individual privacy and predictive power of the models. In this talk we will review these challenges and propose some ways forward. Bio: Neil Lawrence is a Professor of Machine Learning and Computational Biology at the University of Sheffield. His main research interest is machine learning through probabilistic models. He focuses on both the algorithmic side of these models and their application. He has a particular focus on applications in personalized health and applications in the developing world. He is well known for his work with Gaussian processes, and has proposed Gaussian process variants of many of the succesful deep learning architectures. He is highly active in the machine learning community, most recently Program Chairing the NIPS conference in 2014 and General Chairing (alongside Corinna Cortes) in 2015.Thu, 22 Sep 2016 00:00:00 +0000
http://inverseprobability.com/talks/notes/the-data-delusion-democratising.html
http://inverseprobability.com/talks/notes/the-data-delusion-democratising.htmlnotesThe Challenges of Data ScienceData science presents new opportunities but also new challenges. In this talk we will focus on three separate challenges for data science: 1. Paradoxes of the Data Society, 2. Quantifying the Value of Data, 3. Privacy, loss of control, marginalization. Each of these challenges has particular implications for data science. The paradoxes relate to our evolving relationship with data and our changing expectations. Quantifying value is vital for accounting for the influence of data in our new digital economies and issues of privacy and loss of control are fundamental to how our pre-existing rights evolve as the digital world encroaches more closely on the physical. By addressing these challenges now we can ensure that the pitfalls of the data driven society are overcome allowing us to reap the benefits of data science in applications from personalized health to the developing world.Wed, 14 Sep 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-enbis16/the-challenges-of-data-science.html
http://inverseprobability.com/talks/lawrence-enbis16/the-challenges-of-data-science.htmlLawrence-enbis16Fitting Covariance and Multioutput Gaussian ProcessesIn this second session we will talk about fitting covariance matrices and look at multiple output processes.Tue, 13 Sep 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-gpss16b/fitting-covariance-and-multioutput-gaussian-processes.html
http://inverseprobability.com/talks/lawrence-gpss16b/fitting-covariance-and-multioutput-gaussian-processes.htmlLawrence-gpss16bIntroduction to Gaussian ProcessesIn this first session we will introduce Gaussian process models, non parametric Bayesian models that allow for principled propagation of uncertainty in regression analysis. We will assume a background in parametric models, linear algebra and probability.Mon, 12 Sep 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-gpss16a/introduction-to-gaussian-processes.html
http://inverseprobability.com/talks/lawrence-gpss16a/introduction-to-gaussian-processes.htmlLawrence-gpss16aData Science: Where Computation and Statistics Meet?What is data science? An new name for something old perhaps. Nevertheless there is something new happening. Data is being acquired in ways that could never have been envisaged 100 years ago. This is presenting new challenges, and ones that no single field is equipped to face. In this talk we will focus on three separate challenges for data science: 1. Paradoxes of the Data Society, 2. Quantifying the Value of Data, 3. Privacy, loss of control, marginalization. Each of these challenges has particular implications for data science and the interface between computation and statistics. By addressing these challenges now we can ensure that the pitfalls of the data driven society are overcome allowing to reap the benefits.Tue, 06 Sep 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-rss16a/data-science-where-computation-and-statistics-meet.html
http://inverseprobability.com/talks/lawrence-rss16a/data-science-where-computation-and-statistics-meet.htmlLawrence-rss16aCommunicating Machine LearningAs machine learning approaches become more widely adopted their societal impact is increasing. This raises issues in public understanding of science. In this talk I will give an overview of my own approach to addressing this challenge, mixing thoughts and experience into an approach to communicating machine learning.Wed, 31 Aug 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-edinburgh16/communicating-machine-learning.html
http://inverseprobability.com/talks/lawrence-edinburgh16/communicating-machine-learning.htmlLawrence-edinburgh16Variational Compression and Deep <span>G</span>aussian ProcessesIn this fourth sesssion we describe how deep neural networks can be modified to produce deep Gaussian process models. The framework of deep Gaussian processes allow for unsupervised learning, transfer learning, semi-supervised learning, multi-task learning and principled handling of different data types (count data, binary data, heavy tailed noise distributions). The main challenge is to handle the intractabilities. In this talk we review the variational bounds that are used under the framework of variational compression and give some initial results of deep Gaussian process models.Thu, 04 Aug 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-mlss16biv/variational-compression-and-deep-gaussian-processes.html
http://inverseprobability.com/talks/lawrence-mlss16biv/variational-compression-and-deep-gaussian-processes.htmlLawrence-mlss16bIVProbabilistic Dimensionality Reduction with Gaussian ProcessesIn the third session we will look at latent variable models from a Gaussian process perspective with a particular focus on dimensionality reduction.Wed, 03 Aug 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-mlss16biii/probabilistic-dimensionality-reduction-with-gaussian-processes.html
http://inverseprobability.com/talks/lawrence-mlss16biii/probabilistic-dimensionality-reduction-with-gaussian-processes.htmlLawrence-mlss16bIIIIntroduction to Gaussian ProcessesIn this first session we will introduce Gaussian process models, non parametric Bayesian models that allow for principled propagation of uncertainty in regression analysis. We will assume a background in parametric models, linear algebra and probability.Tue, 02 Aug 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-mlss16bi/introduction-to-gaussian-processes.html
http://inverseprobability.com/talks/lawrence-mlss16bi/introduction-to-gaussian-processes.htmlLawrence-mlss16bIIntroduction to Gaussian Processes IIIn the second session we will look at how Gaussian process models are related to Kalman filters and how they may be extended to deal with multiple outputs and mechanistic models.Tue, 02 Aug 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-mlss16bii/introduction-to-gaussian-processes-ii.html
http://inverseprobability.com/talks/lawrence-mlss16bii/introduction-to-gaussian-processes-ii.htmlLawrence-mlss16bIIPrivacy and LearningAbsolute security of information locks it down and exposes it to only those who are granted access. Social privacy can be seen as a continuum where we expose different information to different parties according to levels of trust. In this talk we will briefly introduce our efforts on integrating privacy into learning algorithms to ensure a more equitable and free data society.Thu, 14 Jul 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-security16/privacy-and-learning.html
http://inverseprobability.com/talks/lawrence-security16/privacy-and-learning.htmlLawrence-security16Machine Learning and the ProfessionsAs part of the Royal Society Working Group on Machine Learning this talk is a short introduction to machine learning for members of the professions followed by a provocation on what machine learning might mean for the future of the professions.Wed, 13 Jul 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-professions16/machine-learning-and-the-professions.html
http://inverseprobability.com/talks/lawrence-professions16/machine-learning-and-the-professions.htmlLawrence-professions16New Directions in Data ScienceData science presents new opportunities for Africa but also new challenges. In this talk we will focus on three separate challenges for data science: 1. Paradoxes of the Data Society, 2. Quantifying the Value of Data, 3. Privacy, loss of control, marginalization. Each of these challenges has particular implications for data science in the developing world. By addressing these challenges now we can ensure that the pitfalls of the data driven society are overcome allowing to reap the benefits.Fri, 01 Jul 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-dsa16b/new-directions-in-data-science.html
http://inverseprobability.com/talks/lawrence-dsa16b/new-directions-in-data-science.htmlLawrence-dsa16bIntroduction to Data Science and Machine LearningMon, 27 Jun 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-dsa16a/introduction-to-data-science-and-machine-learning.html
http://inverseprobability.com/talks/lawrence-dsa16a/introduction-to-data-science-and-machine-learning.htmlLawrence-dsa16aSystem Zero: What Kind of AI Have We Created?Machine learning technologies have evolved to the extent that they are now considered the principle underlying technology for our advances in artificial intelligence. Artificial intelligence is an emotive term, given the implications for replacing qualities that humans consider specific to ourselves. In this talk we’ll consider what kind of artificial intelligence we’ve created and what possible implications are for our society.Thu, 09 Jun 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-futureofhumanity16/system-zero-what-kind-of-ai-have-we-created.html
http://inverseprobability.com/talks/lawrence-futureofhumanity16/system-zero-what-kind-of-ai-have-we-created.htmlLawrence-futureofhumanity16Machine Learning and the Future of WorkMachine learning technologies have evolved to the extent that they are now considered the principle underlying technology for our advances in artificial intelligence. Artificial intelligence is an emotive term, given the implications for replacing qualities that humans consider specific to ourselves. As always new technology has a significant disruptive effect on existing markets, jobs and economies. In this talk we’ll explore where the advances are coming from and speculate about how our machine learning future is likely to pan out with a particular focus on work.Fri, 27 May 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-futureofwork16/machine-learning-and-the-future-of-work.html
http://inverseprobability.com/talks/lawrence-futureofwork16/machine-learning-and-the-future-of-work.htmlLawrence-futureofwork16What Kind of AI Have We Created?There have been fears voiced by Elon Musk and Stephen Hawking about the direction of artificial intelligent research. They worry about the creation of a sentient AI, one that might outwit us. However, the nature of the AI we have actually created is a long way distant from this. In this talk we will try and relate our models of artificial intelligence to models that have been proposed for the way humans think. The AI that Hawking and Musk fear is not yet here, but is the AI we have actually developed more or less disturbing than the vision they project?Tue, 24 May 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-pintofscience16/what-kind-of-ai-have-we-created.html
http://inverseprobability.com/talks/lawrence-pintofscience16/what-kind-of-ai-have-we-created.htmlLawrence-pintofscience16Data Efficiency and Machine LearningEntropy is a key component of information and probability, and may provide the key to *data efficient* learning. While we’ve seen great success with the AlphaGo computer program and strides forward in image and speech recognition our current machine learning systems are incredibly data inefficient. Better understanding of entropy with in these systems may provide the key to data efficient learning.Mon, 23 May 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-entropyday16/data-efficiency-and-machine-learning.html
http://inverseprobability.com/talks/lawrence-entropyday16/data-efficiency-and-machine-learning.htmlLawrence-entropyday16Introduction to Gaussian Processes IIFri, 13 May 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-mlss16ii/introduction-to-gaussian-processes-ii.html
http://inverseprobability.com/talks/lawrence-mlss16ii/introduction-to-gaussian-processes-ii.htmlLawrence-mlss16IIIntroduction to Gaussian ProcessesThu, 12 May 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-mlss16i/introduction-to-gaussian-processes.html
http://inverseprobability.com/talks/lawrence-mlss16i/introduction-to-gaussian-processes.htmlLawrence-mlss16IBeyond Backpropagation: Uncertainty PropagationDeep learning is founded on composable functions that are structured to capture regularities in data and can have their parameters optimized by backpropagation (differentiation via the chain rule). Their recent success is founded on the increased availability of data and computational power. However, they are not very data efficient. In low data regimes parameters are not well determined and severe overfitting can occur. The solution is to explicitly handle the indeterminacy by converting it to parameter uncertainty and propagating it through the model. Uncertainty propagation is more involved than backpropagation because it involves convolving the composite functions with probability distributions and integration is more challenging than differentiation. We will present one approach to fitting such models using Gaussian processes. The resulting models perform very well in both supervised and unsupervised learning on small data sets. The remaining challenge is to scale the algorithms to much larger data.Tue, 03 May 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-iclr16/beyond-backpropagation-uncertainty-propagation.html
http://inverseprobability.com/talks/lawrence-iclr16/beyond-backpropagation-uncertainty-propagation.htmlLawrence-iclr16Machine Learning with Gaussian ProcessesGaussian processes (GPs) provide a principled probabilistic approach to prior probability distributions for functions. In this talk we will give an overview of some uses of GPs and their extensions. In particular we will introduce mechanistic models alongside GPs and also use GPs within a structured framework of latent variable models.Thu, 28 Apr 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-amazon16/machine-learning-with-gaussian-processes.html
http://inverseprobability.com/talks/lawrence-amazon16/machine-learning-with-gaussian-processes.htmlLawrence-amazon16Beyond Backpropagation: Uncertainty PropagationDeep learning is founded on composable functions that are structured to capture regularities in data and can have their parameters optimized by backpropagation (differentiation via the chain rule). Their recent success is founded on the increased availability of data and computational power. However, they are not very data efficient. In low data regimes parameters are not well determined and severe overfitting can occur. The solution is to explicitly handle the indeterminacy by converting it to parameter uncertainty and propagating it through the model. Uncertainty propagation is more involved than backpropagation because it involves convolving the composite functions with probability distributions and integration is more challenging than differentiation. We will present one approach to fitting such models using Gaussian processes. The resulting models perform very well in both supervised and unsupervised learning on small data sets. The remaining challenge is to scale the algorithms to much larger data.Tue, 26 Apr 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-msrne16b/beyond-backpropagation-uncertainty-propagation.html
http://inverseprobability.com/talks/lawrence-msrne16b/beyond-backpropagation-uncertainty-propagation.htmlLawrence-msrne16bVariational Inference in Deep GPsThu, 21 Apr 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-msrne16a/variational-inference-in-deep-gps.html
http://inverseprobability.com/talks/lawrence-msrne16a/variational-inference-in-deep-gps.htmlLawrence-msrne16aProbabilistic Dimensionality ReductionIn this talk I give a quick overview of probabilistic interpretations of dimensionality reduction, starting with probabilistic principal component analysis and generalising to non-linear approaches such as the Gaussian Process Latent variable model.Thu, 14 Apr 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-facebook16/probabilistic-dimensionality-reduction.html
http://inverseprobability.com/talks/lawrence-facebook16/probabilistic-dimensionality-reduction.htmlLawrence-facebook16The Data Delusion: Challenges for Democratising Deep LearningThe widespread success of deep learning in a variety of domains is being hailed as a new revolution in artificial intelligence. It has taken 20 years to go from defeating Kasparov at Chess to Lee Sedol at Go. But what have the real advances been across this time? The fundamental change has been in terms of data availability and compute availability. The underlying technology has not changed much in the last 20 years. So what does that mean for areas like medicine and health? Significant challenges remain, improving the data efficiency of these algorithms and retaining the balance between individual privacy and predictive power of the models. In this talk we will review these challenges and propose some ways forward. Bio: Neil Lawrence is a Professor of Machine Learning and Computational Biology at the University of Sheffield. His main research interest is machine learning through probabilistic models. He focuses on both the algorithmic side of these models and their application. He has a particular focus on applications in personalized health and applications in the developing world. He is well known for his work with Gaussian processes, and has proposed Gaussian process variants of many of the succesful deep learning architectures. He is highly active in the machine learning community, most recently Program Chairing the NIPS conference in 2014 and General Chairing (alongside Corinna Cortes) in 2015.Thu, 07 Apr 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-deepsummit16/the-data-delusion-challenges-for-democratising-deep-learning.html
http://inverseprobability.com/talks/lawrence-deepsummit16/the-data-delusion-challenges-for-democratising-deep-learning.htmlLawrence-deepSummit16The Data DelusionThe race on to develop the next generation of artificially intelligent algorithms, recent successes in hitherto unmanageable problems have somewhat blinded us to our own capabilities. Despite the commercial success of the current generation of learning algorithms, the time has come for the academic community to take stock. Have we really got the tools in place to solve the next generation of learning problems? Or is our current confidence in our toolsets misplaced? In this talk we’ll develop at least one direction where our capabilities are lacking.Mon, 21 Mar 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-mars16/the-data-delusion.html
http://inverseprobability.com/talks/lawrence-mars16/the-data-delusion.htmlLawrence-mars16What Kind of AI have we Created?There have been fears voiced by Elon Musk and Stephen Hawking about the direction of artificial intelligent research. They worry about the creation of a sentient AI, one that might outwit us. However, the nature of the AI we have actually created is a long way distant from this. In this talk we will try and relate our models of artificial intelligence to models that have been proposed for the way humans think. The AI that Hawking and Musk fear is not yet here, but is the AI we have actually developed more or less disturbing than the vision they project?Thu, 17 Mar 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-notre15/what-kind-of-ai-public.html
http://inverseprobability.com/talks/lawrence-notre15/what-kind-of-ai-public.htmlLawrence-notre15What Kind of AI have we Created?There have been fears voiced by Elon Musk and Stephen Hawking about the direction of artificial intelligent research. They worry about the creation of a sentient AI, one that might outwit us. However, the nature of the AI we have actually created is a long way distant from this. In this talk we will try and relate our models of artificial intelligence to models that have been proposed for the way humans think. The AI that Hawking and Musk fear is not yet here, but is the AI we have actually developed more or less disturbing than the vision they project?Thu, 10 Mar 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-birley15/what-kind-of-ai-have-we-created.html
http://inverseprobability.com/talks/lawrence-birley15/what-kind-of-ai-have-we-created.htmlLawrence-birley15Future Debates: This House Believes an Artificial Intelligence will Benefit SocietyThe British Science Association hosts a series of debates to encourage constructive debate about science’s role in people’s lives, economy and the UK’s future. This debate was hosted by the Sheffield association and was focussed on artificial intelligence. The debate was led by two speakers from Sheffield’s Debating Society with Tony Dodd supporting the ’against’ and myself supporting the ’for’.Mon, 29 Feb 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-futuredebates16/future-debates-this-house-believes-an-artificial-intelligence-will-benefit-society.html
http://inverseprobability.com/talks/lawrence-futuredebates16/future-debates-this-house-believes-an-artificial-intelligence-will-benefit-society.htmlLawrence-futuredebates16Machine Learning with Gaussian ProcessesGaussian processes (GPs) provide a principled probabilistic approach to prior probability distributions for functions. In this talk we will give an overview of some uses of GPs and their extensions. In particular we will introduce mechanistic models alongside GPs and also use GPs within the framework of latent variable models.Fri, 29 Jan 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-oxwasp16/machine-learning-with-gaussian-processes.html
http://inverseprobability.com/talks/lawrence-oxwasp16/machine-learning-with-gaussian-processes.htmlLawrence-oxwasp16What kind of AI have we created?The media is full of concerns about our data and how algorithms are affecting us. We worry about personal information becoming public, we worry about what intelligent machines have in store for us. This talk will be about the state of the art in terms of Artificial Intelligence. It will consider what it can do and what it can’t do. We are a long way away from implementing a ’sentient intelligence’, but what do we have in its place? This talk will explore current technology and speculate on what futures it may lead to.Tue, 26 Jan 2016 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-phd16/what-kind-of-ai-have-we-created.html
http://inverseprobability.com/talks/lawrence-phd16/what-kind-of-ai-have-we-created.htmlLawrence-phd16The Open Data Science InitiativeThe Open Data Science Initiative is founded on the idea that there are a set of core principles that are restricting our ability, as a society, to exploit the large quantity of data we are now generating. In this talk we identify the challenges across the range of industry, science, health and the developing world. We then review the principles of open data science which we hope will address these challenges.Wed, 16 Dec 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-odsi15/the-open-data-science-initiative.html
http://inverseprobability.com/talks/lawrence-odsi15/the-open-data-science-initiative.htmlLawrence-odsi15Special Topics: Gaussian ProcessesTue, 15 Dec 2015 00:00:00 +0000
http://inverseprobability.com/talks/notes/gaussian-processes.html
http://inverseprobability.com/talks/notes/gaussian-processes.htmlnotesThe Mechanistic Fallacy and Modelling How We ThinkIn this talk we will discuss how our current set of modelling solutions relates to dual process models from psychology. By analogising with layered models of networks we first address the danger of focussing purely on mechanism (or biological plausibility) when discussion modelling in the brain. We term this idea the mechanistic fallacy. In an attempt to operate at a higher level of abstraction, we then take a conceptual approach and attempt to map the broader domain of mechanistic and phenomological models to dual process ideas from psychology. it seems that System 1 is closer to phenomological and System 2 is closer to mechanistic ideas. We will draw connections to surrogate modelling (also known as emmulation) and speculate that one role of System 2 may be to provide additional simulation data for System 1.Fri, 11 Dec 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-mechanistic15/the-mechanistic-fallacy-and-modelling-how-we-think.html
http://inverseprobability.com/talks/lawrence-mechanistic15/the-mechanistic-fallacy-and-modelling-how-we-think.htmlLawrence-mechanistic15Logistic Regression and GLMsNaive Bayes assumptions allow us to specify class conditional densities through assuming that the data are conditionally independent given parameters. A logistic regression is an approach to classification which extends the linear basis function models we’ve already explored. Rather than modeling the output of the function directly the assumption is that we model the <em>log-odds</em> with the basis functions.Tue, 01 Dec 2015 00:00:00 +0000
http://inverseprobability.com/talks/notes/logistic-and-glm.html
http://inverseprobability.com/talks/notes/logistic-and-glm.htmlnotesProbabilistic Classification: Naive BayesIn the last lecture we looked at unsupervised learning. We introduced latent variables, dimensionality reduction and clustering. In this lecture we’re going to look at clustering, specifically the probabilistic approach to clustering. We’ll focus on a simple but often effective algorithm known as <em>naive Bayes</em>.Tue, 24 Nov 2015 00:00:00 +0000
http://inverseprobability.com/talks/notes/naive-bayes.html
http://inverseprobability.com/talks/notes/naive-bayes.htmlnotesInformation Infrastructure for HealthIn this talk we will address challenges in information infrastructure for health. Personalized health care is one of the promises of the information revolution. However, there are major challenges in the curation, collection and management of the data. These are not currently being properly addressed. The care.data fiasco demonstrated the high sensitivity of the public to this data regime. Data leaks from the Pentagon, TalkTalk, Carphone Warehouse have demonstrated the inability of major institutions to keep our data secure. healthcare data is purportedly worth ten times credit card information on international black markets. Machine learning techniques are currently part of the problem, not the solution, they require centralised assimilation of data in a repository that can be easily accessed. A more robust information infrastructure would distribute data and contain afar greater degree of patient control over access. Such user-centric models may offer greater opportunity in terms of obtaining the necessary data-liquidity to fulfill the full potential of personalized health in effecting individuals’ health outcomes.Wed, 18 Nov 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-atiscope15/information-infrastructure-for-health.html
http://inverseprobability.com/talks/lawrence-atiscope15/information-infrastructure-for-health.htmlLawrence-atiscope15Bayesian RegressionBayesian formalisms deal with uncertainty in parameters,Tue, 03 Nov 2015 00:00:00 +0000
http://inverseprobability.com/talks/notes/bayesian-regression.html
http://inverseprobability.com/talks/notes/bayesian-regression.htmlnotesGeneralization: Model ValidationGeneralization is the main objective of a machine learning algorithm. The models we design should work on data they have not seen before. Confirming whether a model generalizes well or not is the domain of <em>model validation</em>. In this lecture we introduce approaches to model validation such as hold out validation and cross validation.Tue, 27 Oct 2015 00:00:00 +0000
http://inverseprobability.com/talks/notes/generalization.html
http://inverseprobability.com/talks/notes/generalization.htmlnotesWhat Kind of Artificial Intelligence are we Creating?The media is full of concerns about our data and how algorithms are affecting us. We worry about personal information becoming public, we worry about what intelligent machines have in store for us. This talk will be about the state of the art in terms of Artificial Intelligence. It will consider what it can do and what it can’t do. We are a long way away from implementing a ’sentient intelligence’, but what do we have in its place? This talk will explore current technology and speculate on what futures it may lead to.Fri, 23 Oct 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-rise15a/what-kind-of-artificial-intelligence-are-we-creating.html
http://inverseprobability.com/talks/lawrence-rise15a/what-kind-of-artificial-intelligence-are-we-creating.htmlLawrence-rise15aMachine Learning Tutorial: Probabilistic Dimensionality Reduction <span>II</span>In the second part of this tutorial we will develop non linear approaches to dimensionality reduction from the probabilistic perspective. Firstly we will briefly review a probabilistic perspectives on spectral approaches, and then we will build on the non-linear approaches we derived using Gaussian processes in the first part of the tutorial.Wed, 21 Oct 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-imperial15b/machine-learning-tutorial-probabilistic-dimensionality-reduction-span-ii-span.html
http://inverseprobability.com/talks/lawrence-imperial15b/machine-learning-tutorial-probabilistic-dimensionality-reduction-span-ii-span.htmlLawrence-imperial15bWhat Kind of Artificial Intelligence have we Created?The media is full of concerns about our data and how algorithms are affecting us. We worry about personal information becoming public, we worry about what intelligent machines have in store for us. This talk will be about the state of the art in terms of Artificial Intelligence. It will consider what it can do and what it can’t do. We are a long way away from implementing a ’sentient intelligence’, but what do we have in its place? This talk will explore current technology and speculate on what futures it may lead to.Tue, 20 Oct 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-rise15/what-kind-of-artificial-intelligence-have-we-created.html
http://inverseprobability.com/talks/lawrence-rise15/what-kind-of-artificial-intelligence-have-we-created.htmlLawrence-rise15Basis Functions<p>In the last session we explored least squares for univariate and multivariate <em>regression</em>. We introduced <em>matrices</em>, <em>linear algebra</em> and <em>derivatives</em>.</p> In this session we will introduce <em>basis functions</em> which allow us to implement <em>non-linear regression models</em>.Tue, 20 Oct 2015 00:00:00 +0000
http://inverseprobability.com/talks/notes/basis-functions.html
http://inverseprobability.com/talks/notes/basis-functions.htmlnotesPersonalised Health and <span>Gaussian</span> ProcessesWed, 14 Oct 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-benevolent15/personalised-health-and-gaussian-processes.html
http://inverseprobability.com/talks/lawrence-benevolent15/personalised-health-and-gaussian-processes.htmlLawrence-benevolent15Linear Algebra and Linear RegressionIn this session we combine the objective function perspective and the probabilistic perspective on <em>linear regression</em>. We motivate the importance of <em>linear algebra</em> by showing how much faster we can complete a linear regression using linear algebra.Tue, 13 Oct 2015 00:00:00 +0000
http://inverseprobability.com/talks/notes/linear-regression.html
http://inverseprobability.com/talks/notes/linear-regression.htmlnotesThe Digital Oligarchy: Information, Knowledge and the Internet EraThe data revolution is among us and the technical press is filled with stories of big data and artificial intelligence. What is driving this progress? In this talk we will argue that collection of data on its own is of little utility, it is interconnection of data that allows information to become knowledge. Businesses need to place data at the core of what they do to benefit from these techniques. The talk will be grounded in academic ideas of what information, knowledge and data are. But these concepts have practical utility that can influence decision making on where data sits within an organisation.Thu, 08 Oct 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-impact15/the-digital-oligarchy-information-knowledge-and-the-internet-era.html
http://inverseprobability.com/talks/lawrence-impact15/the-digital-oligarchy-information-knowledge-and-the-internet-era.htmlLawrence-impact15Objective Functions: A Simple Example with Matrix FactorisationIn this session we introduce the notion of objective functions and show how they can be used in a simple recommender system based on <em>matrix factorisation</em>.Tue, 06 Oct 2015 00:00:00 +0000
http://inverseprobability.com/talks/notes/matrix-factorization.html
http://inverseprobability.com/talks/notes/matrix-factorization.htmlnotesProbability and an Introduction to Jupyter, Python and PandasIn this first session we will introduce <em>machine learning</em>, review <em>probability</em> and begin familiarization with the Jupyter notebook, python and pandas.Tue, 29 Sep 2015 00:00:00 +0000
http://inverseprobability.com/talks/notes/intro-probability.html
http://inverseprobability.com/talks/notes/intro-probability.htmlnotesPeer Review and The NIPS ExperimentThe peer review process can be difficult to navigate for newcomers. In this informal talk we will review the results of the NIPS experiment, an experiment on the repeatability of peer review conducted for the 2014 conference. We will try to keep the presentation information to ensure questions can be asked. With luck it will give more insight into the processes that a program committee goes through when selecting papers.Mon, 21 Sep 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-peer15/peer-review-and-the-nips-experiment.html
http://inverseprobability.com/talks/lawrence-peer15/peer-review-and-the-nips-experiment.htmlLawrence-peer15Deep <span>G</span>aussian ProcessesThu, 20 Aug 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-harvard15/deep-span-g-span-aussian-processes.html
http://inverseprobability.com/talks/lawrence-harvard15/deep-span-g-span-aussian-processes.htmlLawrence-harvard15Personalized Health with <span>G</span>aussian ProcessesModern data connectivity gives us different views of the patient which need to be unified for truly personalized health care. I’ll give an personal perspective on the type of methodological and social challenges we expect to arise in this this domain and motivate Gaussian process models as one approach to dealing with the explosion of data.Wed, 19 Aug 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-msrne15/personalized-health-with-span-g-span-aussian-processes.html
http://inverseprobability.com/talks/lawrence-msrne15/personalized-health-with-span-g-span-aussian-processes.htmlLawrence-msrne15Latent Force Models: Bridging the Divide between Mechanistic and Data Modelling ParadigmsTue, 21 Jul 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-mlss15bc/latent-force-models-bridging-the-divide-between-mechanistic-and-data-modelling-paradigm.html
http://inverseprobability.com/talks/lawrence-mlss15bc/latent-force-models-bridging-the-divide-between-mechanistic-and-data-modelling-paradigm.htmlLawrence-mlss15bc<span>G</span>aussian Processes (Part III)Sat, 18 Jul 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-mlss15biii/gaussian-processes-part-iii.html
http://inverseprobability.com/talks/lawrence-mlss15biii/gaussian-processes-part-iii.htmlLawrence-mlss15bIII<span>G</span>aussian Processes (Part II)Fri, 17 Jul 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-mlss15bii/gaussian-processes-part-ii.html
http://inverseprobability.com/talks/lawrence-mlss15bii/gaussian-processes-part-ii.htmlLawrence-mlss15bII<span>G</span>aussian Processes (Part I)Thu, 16 Jul 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-mlss15bi/gaussian-processes-part-i.html
http://inverseprobability.com/talks/lawrence-mlss15bi/gaussian-processes-part-i.htmlLawrence-mlss15bIPanel DiscussionSat, 11 Jul 2015 00:00:00 +0000
http://inverseprobability.com/talks/welling-deeppanel15/panel-discussion.html
http://inverseprobability.com/talks/welling-deeppanel15/panel-discussion.htmlWelling-deeppanel15Large Scale Learning in <span>G</span>aussian ProcessesGaussian process models view the kernel matrix as representing the covariance between data points. In a Gaussian process, the RKHS function is a mean of a posterior distribution over possible functions. Gaussian processes sustain uncertainty around this means and this leads to a posterior \*covariance\* function (or kernel) associated with the process. A complication for large scale Gaussian process models is the need to sustain the estimate for this covariance function. In this talk we’ll review how this can be done probabilistically through a variational approach we know as ’variational compression’.Sat, 11 Jul 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-largeicml15/large-scale-learning-in-span-g-span-aussian-processes.html
http://inverseprobability.com/talks/lawrence-largeicml15/large-scale-learning-in-span-g-span-aussian-processes.htmlLawrence-largeicml15Deep Gaussian ProcessesIn this talk we describe how deep neural networks can be modified to produce deep Gaussian process models. The framework of deep Gaussian processes allow for unsupervised learning, transfer learning, semi-supervised learning, multi-task learning and principled handling of different data types (count data, binary data, heavy tailed noise distributions). The main challenge is to handle the intractabilities. In this talk we review the variational bounds that are used under the framework of variational compression and give some initial results of deep Gaussian process models.Sat, 11 Jul 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-deepicml15/deep-gaussian-processes.html
http://inverseprobability.com/talks/lawrence-deepicml15/deep-gaussian-processes.htmlLawrence-deepicml15Personalized HealthThu, 18 Jun 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-nyeri15c/personalized-health.html
http://inverseprobability.com/talks/lawrence-nyeri15c/personalized-health.htmlLawrence-nyeri15cRegressionMon, 15 Jun 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-nyeri15b/regression.html
http://inverseprobability.com/talks/lawrence-nyeri15b/regression.htmlLawrence-nyeri15bIntroduction to Machine Learning and Data ScienceMon, 15 Jun 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-nyeri15a/introduction-to-machine-learning-and-data-science.html
http://inverseprobability.com/talks/lawrence-nyeri15a/introduction-to-machine-learning-and-data-science.htmlLawrence-nyeri15aDeep <span>G</span>aussian ProcessesIn this talk we describe how deep neural networks can be modified to produce deep Gaussian process models. The framework of deep Gaussian processes allow for unsupervised learning, transfer learning, semi-supervised learning, multi-task learning and principled handling of different data types (count data, binary data, heavy tailed noise distributions). The main challenge is to handle the intractabilities. In this talk we review the variational bounds that are used under the framework of variational compression and give some initial results of deep Gaussian process models.Tue, 09 Jun 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-edinburgh15/deep-span-g-span-aussian-processes.html
http://inverseprobability.com/talks/lawrence-edinburgh15/deep-span-g-span-aussian-processes.htmlLawrence-edinburgh15Deep <span>G</span>aussian ProcessesIn this talk we describe how deep neural networks can be modified to produce deep Gaussian process models. The framework of deep Gaussian processes allow for unsupervised learning, transfer learning, semi-supervised learning, multi-task learning and principled handling of different data types (count data, binary data, heavy tailed noise distributions). The main challenge is to handle the intractabilities. In this talk we review the variational bounds that are used under the framework of variational compression and give some initial results of deep Gaussian process models.Mon, 11 May 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-nyu15/deep-span-g-span-aussian-processes.html
http://inverseprobability.com/talks/lawrence-nyu15/deep-span-g-span-aussian-processes.htmlLawrence-nyu15Deep <span>G</span>aussian ProcessesIn this talk we describe how deep neural networks can be modified to produce deep Gaussian process models. The framework of deep Gaussian processes allow for unsupervised learning, transfer learning, semi-supervised learning, multi-task learning and principled handling of different data types (count data, binary data, heavy tailed noise distributions). The main challenge is to handle the intractabilities. In this talk we review the variational bounds that are used under the framework of variational compression and give some initial results of deep Gaussian process models.Thu, 30 Apr 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-kth15/deep-gaussian-processes.html
http://inverseprobability.com/talks/lawrence-kth15/deep-gaussian-processes.htmlLawrence-kth15Deep <span>G</span>aussian ProcessesIn this talk we describe how deep neural networks can be modified to produce deep Gaussian process models. The framework of deep Gaussian processes allow for unsupervised learning, transfer learning, semi-supervised learning, multi-task learning and principled handling of different data types (count data, binary data, heavy tailed noise distributions). The main challenge is to handle the intractabilities. In this talk we review the variational bounds that are used under the framework of variational compression and give some initial results of deep Gaussian process models.Wed, 29 Apr 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-linkoping15/deep-gaussian-processes.html
http://inverseprobability.com/talks/lawrence-linkoping15/deep-gaussian-processes.htmlLawrence-linkoping15Deep Gaussian ProcessesIn this talk we describe how deep neural networks can be modified to produce deep Gaussian process models. The framework of deep Gaussian processes allow for unsupervised learning, transfer learning, semi-supervised learning, multi-task learning and principled handling of different data types (count data, binary data, heavy tailed noise distributions). The main challenge is to handle the intractabilities. In this talk we review the variational bounds that are used under the framework of variational compression and give some initial results of deep Gaussian process models.Wed, 08 Apr 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-mascotnum15/deep-gaussian-processes.html
http://inverseprobability.com/talks/lawrence-mascotnum15/deep-gaussian-processes.htmlLawrence-mascotnum15Modelling in the Context of Massively Missing DataIn the age of large streaming data it seems appropriate to revisit the foundations of what we think of as data modelling. In this talk I’ll argue that traditional statistical approaches based on parametric models and i.i.d. assumptions are inappropriate for the type of large scale machine learning we need to do in the age of massive streaming data sets. Particularly when we realise that regardless of the size of data we have, it pales in comparison to the data we could have. This is the domain of *massively missing data*. I’ll be arguing for flexible non-parametric models as the answer. This presents a particular challenge, non parametric models require data storage of the entire data set, which presents problems for massive, streaming data. I will present a potential solution, but perhaps end with more questions than we started with.Wed, 18 Mar 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-mpi15/modelling-in-the-context-of-massively-missing-data.html
http://inverseprobability.com/talks/lawrence-mpi15/modelling-in-the-context-of-massively-missing-data.htmlLawrence-mpi15The Data FarmLike Hansel and Gretel’s breadcrumbs into the forest we leave a data trail of data-crumbs wherever we go: social networks, mobile
phones, hospital visits, credit cards and loyalty cards. Our every move is being watched! The data-crumbs are seeds of information but
what results from them... is it a jungle with dangers lurking or a productive farmyard? And if our data is being farmed, where does all
the produce go?
<p>This edition of the talk was given to an age group between 8 and 10.Fri, 13 Mar 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-datafarm15a/the-data-farm.html
http://inverseprobability.com/talks/lawrence-datafarm15a/the-data-farm.htmlLawrence-datafarm15aData Science: A New Field or Just a Rebadging Exercise?Scientific fields don’t necessarily emerge because fundamental new knowledge is being generated, but often because a shift in the key questions that are facing us, and the tools that we have to answer them. The current information revolution is causing us to reassess our approach to data. Our mathematical and computational toolsets are co-evolving. The potential of very large interconnected data is placing urgent demands on our methodologies. In this talk, inspired by these challenges, I will give a personal perspective on what this means for those of us at the interface of Computer Science/Mathematics and Statistics. I’ll attempt to do this not only in the context of modelling and analysis, but also in the context of how we deploy our conclusions for the benefit of wider society. Many of our current suite of methodologies were motivated by different needs, and I’ll argue that it may now be time to return to the fundamental ideas from where these methodologies were inspired, but with a contemporary slant on the nature of data. My own perspective is that if what I describe \*is\* data science, then it does not stand as a field alone, but it represents a new and pressing set of questions that bridge the computational and mathematical sciences. Regardless of its phylogeny, exploring this interface through these questions will be mutually beneficial.Thu, 12 Mar 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-nottingham15/data-science-a-new-field-or-just-a-rebadging-exercise.html
http://inverseprobability.com/talks/lawrence-nottingham15/data-science-a-new-field-or-just-a-rebadging-exercise.htmlLawrence-nottingham15Machine Learning Tutorial: Probabilistic Dimensionality ReductionIn this tutorial we will present probabilistic approaches to dimensionality reduction based on latent variable models. We will motivate dimensionality reduction and then start with principal component analysis and extend it to include non linear approaches to reducing the dimension of data.Wed, 11 Mar 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-imperial15/machine-learning-tutorial-probabilistic-dimensionality-reduction.html
http://inverseprobability.com/talks/lawrence-imperial15/machine-learning-tutorial-probabilistic-dimensionality-reduction.htmlLawrence-imperial15The Data FarmLike Hansel and Gretel’s breadcrumbs into the forest we leave a data trail of data-crumbs wherever we go: social networks, mobile phones, hospital visits, credit cards and loyalty cards. Our every move is being watched! The data-crumbs are seeds of information but what results from them... is it a jungle with dangers lurking or a productive farmyard? And if our data is being farmed, where does all the produce go?Thu, 05 Mar 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-fest15/the-data-farm.html
http://inverseprobability.com/talks/lawrence-fest15/the-data-farm.htmlLawrence-fest15Introduction to <span>G</span>aussian ProcessesSat, 21 Feb 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-mlss15/introduction-to-span-g-span-aussian-processes.html
http://inverseprobability.com/talks/lawrence-mlss15/introduction-to-span-g-span-aussian-processes.htmlLawrence-mlss15The NIPS ExperimentThe peer review process can be difficult to navigate for newcomers. In this informal talk we will review the results of the NIPS experiment, an experiment on the repeatability of peer review conducted for the 2014 conference. We will try to keep the presentation information to ensure questions can be asked. With luck it will give more insight into the processes that a program committee goes through when selecting papers.Fri, 30 Jan 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-radiant15/the-nips-experiment-examining-the-repeatability-of-peer-review.html
http://inverseprobability.com/talks/lawrence-radiant15/the-nips-experiment-examining-the-repeatability-of-peer-review.htmlLawrence-radiant15Deep Gaussian ProcessesIn this talk we describe how deep neural networks can be modified to produce deep Gaussian process models. The framework of deep Gaussian processes allow for unsupervised learning, transfer learning, semi-supervised learning, multi-task learning and principled handling of different data types (count data, binary data, heavy tailed noise distributions). The main challenge is to solve these models efficiently for massive data sets. That challenge is in reach through a new class of variational approximations known as variational compression. The underlying variational bounds are very similar to the objective functions for deep neural networks, giving the promise of efficient approaches to deep learning that are constructed from components with very well understood analytical properties.Fri, 23 Jan 2015 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-iit15/deep-gaussian-processes.html
http://inverseprobability.com/talks/lawrence-iit15/deep-gaussian-processes.htmlLawrence-iit15Data Science: A New Field or Just a Rebadging Exercise?Scientific fields don’t necessarily emerge because fundamental new knowledge is being generated, but often because a shift in the key questions that are facing us, and the tools that we have to answer them. The current information revolution is causing us to reassess our approach to data. Our mathematical and computational toolsets are co-evolving. The potential of very large interconnected data is placing urgent demands on our methodologies. In this talk, inspired by these challenges, I will give a personal perspective on what this means for those of us at the interface of Computer Science/Mathematics and Statistics. I’ll attempt to do this not only in the context of modelling and analysis, but also in the context of how we deploy our conclusions for the benefit of wider society. Many of our current suite of methodologies were motivated by different needs, and I’ll argue that it may now be time to return to the fundamental ideas from where these methodologies were inspired, but with a contemporary slant on the nature of data. My own perspective is that if what I describe \*is\* data science, then it does not stand as a field alone, but it represents a new and pressing set of questions that bridge the computational and mathematical sciences. Regardless of its phylogeny, exploring this interface through these questions will be mutually beneficial.Wed, 26 Nov 2014 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-warwick14/data-science-a-new-field-or-just-a-rebadging-exercise.html
http://inverseprobability.com/talks/lawrence-warwick14/data-science-a-new-field-or-just-a-rebadging-exercise.htmlLawrence-warwick14Statistical Computing: PythonFri, 21 Nov 2014 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-rss14/statistical-computing-python.html
http://inverseprobability.com/talks/lawrence-rss14/statistical-computing-python.htmlLawrence-rss14Approximate Inference in Deep GPsIn this talk we will review deep Gaussian process models and relate them to neural network models. We will then consider the details of how variational inference may be performed in these models. The approach is centred on 'variational compression', an approach to variational inference that compresses information into an augmented variable space. The aim of the deep Gaussian process framework is to enable probabilistic learning of multi-modal data. We will therefore end by highlighting directions for future research and discussing application of these models in domains such as personalised health.Thu, 23 Oct 2014 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-ucl14c/approximate-inference-in-deep-gps.html
http://inverseprobability.com/talks/lawrence-ucl14c/approximate-inference-in-deep-gps.htmlLawrence-ucl14cDeep Gaussian ProcessesIn this talk we describe how deep neural networks can be modified to produce
deep Gaussian process models. The framework of deep Gaussian processes allow for
unsupervised learning, transfer learning, semi-supervised learning, multi-task learning
and principled handling of different data types (count data, binary data, heavy
tailed noise distributions). The main challenge is to solve these models efficiently
for massive data sets. That challenge is in reach through a new class of variational
approximations known as variational compression. The underlying variational bounds
are very similar to the objective functions for deep neural networks, giving the
promise of efficient approaches to deep learning that are constructed from components
with very well understood analytical properties.
Thu, 04 Sep 2014 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-ucl14b/deep-gaussian-processes.html
http://inverseprobability.com/talks/lawrence-ucl14b/deep-gaussian-processes.htmlLawrence-ucl14bBig Data and Open Data ScienceIn this talk we will focus on the challenges that are arising through big data and focussing on potential solutions, both from a methodological side, but also in terms of the way that statistics and computer science need to respond to the challenges culturally.Wed, 02 Jul 2014 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-uclid14/big-data-and-open-data-science.html
http://inverseprobability.com/talks/lawrence-uclid14/big-data-and-open-data-science.htmlLawrence-uclid14Flexible Parametric Representations of Non Parametric ModelsIn the age of large streaming data it seems appropriate to revisit the foundations of what we think of as data modelling. In this talk I’ll argue that traditional statistical approaches based on parametric models and i.i.d. assumptions are inappropriate for the type of large scale machine learning we need to do in the age of massive streaming data sets. I’ll be arguing for flexible non-parametric models as the answer. This presents a particular challenge, non parametric models require data storage of the entire data set, which presents problems for massive, streaming data. I’ll argue that recently proposed variational approximations allow us to retain the advantages of both non-parametric and parametric models within a consistent framework that performs an optimal compression of our data from an information gain perspective.Mon, 19 May 2014 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-edinburgh14/flexible-parametric-representations-of-non-parametric-models.html
http://inverseprobability.com/talks/lawrence-edinburgh14/flexible-parametric-representations-of-non-parametric-models.htmlLawrence-edinburgh14Visualizing Biological Data with <span>G</span>aussian ProcessesTue, 13 May 2014 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-ebi14b/visualizing-biological-data-with-span-g-span-aussian-processes.html
http://inverseprobability.com/talks/lawrence-ebi14b/visualizing-biological-data-with-span-g-span-aussian-processes.htmlLawrence-ebi14bGaussian Processes for Dynamic ModellingTue, 13 May 2014 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-ebi14a/gaussian-processes-for-dynamic-modelling.html
http://inverseprobability.com/talks/lawrence-ebi14a/gaussian-processes-for-dynamic-modelling.htmlLawrence-ebi14aWhat is Machine Learning? A Probabilistic Perspective (Part II)Sat, 26 Apr 2014 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-mlss14b/what-is-machine-learning-a-probabilistic-perspective-part-ii.html
http://inverseprobability.com/talks/lawrence-mlss14b/what-is-machine-learning-a-probabilistic-perspective-part-ii.htmlLawrence-mlss14bWhat is Machine Learning? A Probabilistic Perspective (Part I)Sat, 26 Apr 2014 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-mlss14/what-is-machine-learning-a-probabilistic-perspective-part-i.html
http://inverseprobability.com/talks/lawrence-mlss14/what-is-machine-learning-a-probabilistic-perspective-part-i.htmlLawrence-mlss14Flexible Parametric Representations of Non Parametric ModelsIn the age of large streaming data it seems appropriate to revisit the foundations of what we think of as data modelling. In this talk I’ll argue that traditional statistical approaches based on parametric models and i.i.d. assumptions are inappropriate for the type of large scale machine learning we need to do in the age of massive streaming data sets. I’ll be arguing for flexible non-parametric models as the answer. This presents a particular challenge, non parametric models require data storage of the entire data set, which presents problems for massive, streaming data. I’ll argue that recently proposed variational approximations allow us to retain the advantages of both non-parametric and parametric models within a consistent framework that performs an optimal compression of our data from an information gain perspective.Thu, 03 Apr 2014 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-smile14/flexible-parametric-representations-of-non-parametric-models.html
http://inverseprobability.com/talks/lawrence-smile14/flexible-parametric-representations-of-non-parametric-models.htmlLawrence-smile14Applications of <span>G</span>aussian Processes in Computational BiologyIn this talk we will give a brief overview of Gaussian processes and a quick review of how they can be applied to solve questions in computational biology. In particular we will show how we can construct covariance functions to solve simple tasks (like differential expression) or more complex tasks (like unpicking regulatory networks).Thu, 03 Apr 2014 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-curie14/applications-of-span-g-span-aussian-processes-in-computational-biology.html
http://inverseprobability.com/talks/lawrence-curie14/applications-of-span-g-span-aussian-processes-in-computational-biology.htmlLawrence-curie14Modelling with Massively Missing DataSupervised deep learning techniques now dominate in terms of performance for complex classification tasks such as ImageNet. For these, the set of inputs (features) and targets (labels) are typically well defined in advance. However, for many tasks in artificial intelligence the questions that need to be answered evolve, alongside the features that we can acquire. For example, imagine we wish to infer the health status of individuals by building population scale models based on clinical data. For most people in the population most of the data will be missing because clinical tests are not applied to patients as a matter of course. Indeed, some of the features we may wish to use in our model may not even exist when our model is first designed (e.g. emerging clinical tests and treatments). We refer to this scenario as ’massively missing data’. It is a scenario humans are faced with every day. Almost all of the time we are missing almost all of the data. And yet we have no difficulty assimilating disparate pieces of information from a wide range of sources to draw inferences about our world. Implementing machine learning systems that can replicate this characteristic requires model architectures that can be adapted at ’runtime’ as the data evolves, we don’t want to be limited by decisions made at ’design time’ when perhaps a more limited feature set existed. This poses particular challenges that we will address in this talk.Thu, 20 Mar 2014 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-facebook14/modelling-with-massively-missing-data.html
http://inverseprobability.com/talks/lawrence-facebook14/modelling-with-massively-missing-data.htmlLawrence-facebook14Flexible Parametric Representations of Non Parametric ModelsIn the age of large streaming data it seems appropriate to revisit the foundations of what we think of as data modelling. In this talk I’ll argue that traditional statistical approaches based on parametric models and i.i.d. assumptions are inappropriate for the type of large scale machine learning we need to do in the age of massive streaming data sets. I’ll be arguing for flexible non-parametric models as the answer. This presents a particular challenge, non parametric models require data storage of the entire data set, which presents problems for massive, streaming data. I’ll argue that recently proposed variational approximations allow us to retain the advantages of both non-parametric and parametric models within a consistent framework that performs an optimal compression of our data from an information gain perspective.Wed, 26 Feb 2014 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-ucl14/flexible-parametric-representations-of-non-parametric-models.html
http://inverseprobability.com/talks/lawrence-ucl14/flexible-parametric-representations-of-non-parametric-models.htmlLawrence-ucl14Personalized Health with <span>G</span>aussian ProcessesModern data connectivity gives us different views of the patient which need to be unified for truly personalized health care. I’ll give an personal perspective on the type of methodological challenges we expect to arise in this this domain and motivate Gaussian process models as one approach to dealing with the explosion of data.Wed, 19 Feb 2014 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-manizales14/personalized-health-with-span-g-span-aussian-processes.html
http://inverseprobability.com/talks/lawrence-manizales14/personalized-health-with-span-g-span-aussian-processes.htmlLawrence-manizales14Deep Gaussian ProcessesIn this talk we will introduce deep Gaussian process (GP) models. Deep GPs are a deep probabilistic model based on Gaussian process mappings. The data is modelled as the output of a multivariate GP. The inputs to that Gaussian process are then governed by another GP. A single layer model is equivalent to a standard GP or the GP latent variable model (GPLVM). We will motivate these models by considering applications in personalized health.
We perform inference in the model by approximate variational marginalization. This results in a strict lower bound on the marginal likelihood of the model which we use for model selection (number of layers and nodes per layer). Deep belief networks are typically applied to relatively large data sets using stochastic gradient descent for optimization. Our fully Bayesian treatment allows for the application of deep models even when data is scarce. Model selection by our variational bound shows that a five layer hierarchy is justified even when modelling a digit data set containing only 150 examples. In the seminar we will briefly review dimensionality reduction via Gaussian processes, before showing how this framework can be extended to build deep models.
Thu, 06 Feb 2014 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-oxford14/deep-gaussian-processes.html
http://inverseprobability.com/talks/lawrence-oxford14/deep-gaussian-processes.htmlLawrence-oxford14New Perspectives on Variational Approximations in <span>G</span>aussian Processes: Modelling DataIn this talk I’ll introduce new perspectives on variational approximations. Many of the ideas may be widely applicable, but we will try to instantiate them in the context of Gaussian process models.\
\
Although the variational material itself is reasonably technical, I’ll try and start the talk by making general statements about data modelling. Then, in an effort to make the talk seem coherent, I’ll make claims that the technical material which follows was inspired by the wider perspective I’ve given. Of course in practice, the technical material really emerged across a number of years during discussions with many people, and the general perspective has been retrofitted. Still, I’ll be giving the talk amongst friends, so no one will mind too much if the story doesn’t really fit together, and in fact it might be a good trigger for discussion. Speaking of which, I’ll be looking forward to lots of audience participation, and such participation may take the talk in previously unplanned directions.\
\
The talk will be given without the use of electronic aids.Tue, 21 Jan 2014 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-cued14/new-perspectives-on-variational-approximations-in-span-g-span-aussian-processes-modelli.html
http://inverseprobability.com/talks/lawrence-cued14/new-perspectives-on-variational-approximations-in-span-g-span-aussian-processes-modelli.htmlLawrence-cued14Latent Variable Models with Gaussian ProcessesWed, 15 Jan 2014 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-gpwsthree14/latent-variable-models-with-gaussian-processes.html
http://inverseprobability.com/talks/lawrence-gpwsthree14/latent-variable-models-with-gaussian-processes.htmlLawrence-gpwsThree14Fitting Covariance and Multi-output <span>G</span>aussian ProcessesTue, 14 Jan 2014 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-gpwstwo14/fitting-covariance-and-multi-output-span-g-span-aussian-processes.html
http://inverseprobability.com/talks/lawrence-gpwstwo14/fitting-covariance-and-multi-output-span-g-span-aussian-processes.htmlLawrence-gpwsTwo14Introduction to <span>G</span>aussian ProcessesMon, 13 Jan 2014 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-gpwsone14/introduction-to-span-g-span-aussian-processes.html
http://inverseprobability.com/talks/lawrence-gpwsone14/introduction-to-span-g-span-aussian-processes.htmlLawrence-gpwsOne14Unravelling the Big Data RevolutionModern data connectivity gives us massive uncurated data sets which present enormous challenges for modelling and inference. I’ll review where I think this is taking mathematics and speculate on the methodological and social challenges that this revolution will entail with some final reflections on how it might effect the teaching curriculum.Wed, 18 Dec 2013 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-leeds13/unravelling-the-big-data-revolution.html
http://inverseprobability.com/talks/lawrence-leeds13/unravelling-the-big-data-revolution.htmlLawrence-leeds13Unravelling the Data Revolution with Machine LearningModern data connectivity gives us different views of the patient which need to be unified for truly personalized health care. I’ll review where I think this is taking medicine and speculate on the methodological and social challenges that this revolution will entail.Thu, 14 Nov 2013 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-necs13/unravelling-the-data-revolution-with-machine-learning.html
http://inverseprobability.com/talks/lawrence-necs13/unravelling-the-data-revolution-with-machine-learning.htmlLawrence-necs13Personalized Health with <span>G</span>aussian ProcessesModern data connectivity gives us different views of the patient which need to be unified for truly personalized health care. I’ll give an personal perspective on the type of methodological challenges we expect to arise in this this domain and motivate Gaussian process models as one approach to dealing with the explosion of data.Mon, 04 Nov 2013 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-leahurst13/personalized-health-with-span-g-span-aussian-processes.html
http://inverseprobability.com/talks/lawrence-leahurst13/personalized-health-with-span-g-span-aussian-processes.htmlLawrence-leahurst13Deep Health: Machine Learning for Personalized MedicineI’ll give an overview of the methodological challenges we see arising in personalized medicine. These are associated with the explosion of data giving us different views of the patient which need to be unified for truly personalized health care.Thu, 03 Oct 2013 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-e4l13/deep-health-machine-learning-for-personalized-medicine.html
http://inverseprobability.com/talks/lawrence-e4l13/deep-health-machine-learning-for-personalized-medicine.htmlLawrence-e4l13Probabilistic Approaches for Computational Biology and MedicineIn this talk I’ll discuss some of the challenges in personalized medicine and consider some of the implications for machine learning models. I’ll introduce the probabilistic approach to machine learning, with a particular focus on Gaussian models. Giving some examples of applications I’ll discuss Bayesian approaches to regression modelling and lead into Gaussian process models.Wed, 25 Sep 2013 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-mlpm13/probabilistic-approaches-for-computational-biology-and-medicine.html
http://inverseprobability.com/talks/lawrence-mlpm13/probabilistic-approaches-for-computational-biology-and-medicine.htmlLawrence-mlpm13A Unifying Probabilistic Perspective on Spectral Approaches to Dimensionality ReductionSpectral approaches to dimensionality reduction typically reduce the dimensionality of a data set through taking the eigenvectors of a Laplacian or a similarity matrix. Classical multidimensional scaling also makes use of the eigenvectors of a similarity matrix. In this talk we introduce a maximum entropy approach to designing this similarity matrix. The approach is closely related to maximum variance unfolding. Other spectral approaches such as locally linear embeddings and Laplacian eigenmaps also turn out to be closely related. Each method can be seen as a sparse Gaussian graphical model where correlations between data points (rather than across data features) are specified in the graph. This also suggests optimization via sparse inverse covariance techniques such as the graphical LASSO. The hope is that this unifying perspective will allow the relationships between these methods to be better understood and will also provide the groundwork for further research.Thu, 05 Sep 2013 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-msr13b/a-unifying-probabilistic-perspective-on-spectral-approaches-to-dimensionality-reduction.html
http://inverseprobability.com/talks/lawrence-msr13b/a-unifying-probabilistic-perspective-on-spectral-approaches-to-dimensionality-reduction.htmlLawrence-msr13bDeep <span>Gaussian</span> ProcessesIn this talk we will introduce deep Gaussian process (GP) models. Deep GPs are a deep belief network based on Gaussian process mappings. The data is modeled as the output of a multivariate GP. The inputs to that Gaussian process are then governed by another GP. A single layer model is equivalent to a standard GP or the GP latent variable model (GPLVM). We perform inference in the model by approximate variational marginalization. This results in a strict lower bound on the marginal likelihood of the model which we use for model selection (number of layers and nodes per layer). Deep belief networks are typically applied to relatively large data sets using stochastic gradient descent for optimization. Our fully Bayesian treatment allows for the application of deep models even when data is scarce. Model selection by our variational bound shows that a five layer hierarchy is justified even when modelling a digit data set containing only 150 examples. In the seminar we will briefly review dimensionality reduction via Gaussian processes, before showing how this framework can be extended to build deep models.Tue, 03 Sep 2013 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-msr13a/deep-span-gaussian-span-processes.html
http://inverseprobability.com/talks/lawrence-msr13a/deep-span-gaussian-span-processes.htmlLawrence-msr13aDeep <span>Gaussian</span> ProcessesIn this talk we will introduce deep Gaussian process (GP) models. Deep GPs are a deep belief network based on Gaussian process mappings. The data is modeled as the output of a multivariate GP. The inputs to that Gaussian process are then governed by another GP. A single layer model is equivalent to a standard GP or the GP latent variable model (GPLVM). We perform inference in the model by approximate variational marginalization. This results in a strict lower bound on the marginal likelihood of the model which we use for model selection (number of layers and nodes per layer). Deep belief networks are typically applied to relatively large data sets using stochastic gradient descent for optimization. Our fully Bayesian treatment allows for the application of deep models even when data is scarce. Model selection by our variational bound shows that a five layer hierarchy is justified even when modelling a digit data set containing only 150 examples. In the seminar we will briefly review dimensionality reduction via Gaussian processes, before showing how this framework can be extended to build deep models.Thu, 04 Jul 2013 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-ncaf13/deep-span-gaussian-span-processes.html
http://inverseprobability.com/talks/lawrence-ncaf13/deep-span-gaussian-span-processes.htmlLawrence-ncaf13Deep HealthMon, 17 Jun 2013 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-manchester13/deep-health.html
http://inverseprobability.com/talks/lawrence-manchester13/deep-health.htmlLawrence-manchester13Latent Force Models: IntroductionThu, 13 Jun 2013 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-lfmintro13/latent-force-models-introduction.html
http://inverseprobability.com/talks/lawrence-lfmintro13/latent-force-models-introduction.htmlLawrence-lfmIntro13Unsupervised Learning with <span>G</span>aussian ProcessesWed, 12 Jun 2013 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-gpssthree13/unsupervised-learning-with-span-g-span-aussian-processes.html
http://inverseprobability.com/talks/lawrence-gpssthree13/unsupervised-learning-with-span-g-span-aussian-processes.htmlLawrence-gpssThree13Multioutput <span>G</span>aussian ProcessesTue, 11 Jun 2013 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-gpsstwo13/multioutput-span-g-span-aussian-processes.html
http://inverseprobability.com/talks/lawrence-gpsstwo13/multioutput-span-g-span-aussian-processes.htmlLawrence-gpssTwo13Introduction to <span>G</span>aussian ProcessesMon, 10 Jun 2013 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-gpssone13/introduction-to-span-g-span-aussian-processes.html
http://inverseprobability.com/talks/lawrence-gpssone13/introduction-to-span-g-span-aussian-processes.htmlLawrence-gpssOne13Deep <span>Gaussian</span> ProcessesIn this talk we will introduce deep Gaussian process (GP) models. Deep GPs are a deep belief network based on Gaussian process mappings. The data is modeled as the output of a multivariate GP. The inputs to that Gaussian process are then governed by another GP. A single layer model is equivalent to a standard GP or the GP latent variable model (GPLVM). We perform inference in the model by approximate variational marginalization. This results in a strict lower bound on the marginal likelihood of the model which we use for model selection (number of layers and nodes per layer). Deep belief networks are typically applied to relatively large data sets using stochastic gradient descent for optimization. Our fully Bayesian treatment allows for the application of deep models even when data is scarce. Model selection by our variational bound shows that a five layer hierarchy is justified even when modelling a digit data set containing only 150 examples. In the seminar we will briefly review dimensionality reduction via Gaussian processes, before showing how this framework can be extended to build deep models.Wed, 01 May 2013 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-cambridge13/deep-span-gaussian-span-processes.html
http://inverseprobability.com/talks/lawrence-cambridge13/deep-span-gaussian-span-processes.htmlLawrence-cambridge13How the Planets Affect Our Daily Lives: A Brief History of UncertaintyWithin the last 400 years scientists became able to predict the future. Crystal balls were replaced with computation. Uncertainty met mathematics. This talk gives a brief history of uncertainty and prediction. You will find out how planets affect who your Facebook friends are.Thu, 21 Mar 2013 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-scienceweek_edwards13/how-the-planets-affect-our-daily-lives-a-brief-history-of-uncertainty.html
http://inverseprobability.com/talks/lawrence-scienceweek_edwards13/how-the-planets-affect-our-daily-lives-a-brief-history-of-uncertainty.htmlLawrence-scienceweek_edwards13Deep Learning: What is it and What are We doing About it?In November last year, deep learning algorithms made the front page of the New York Times. What’s special about these learning
algorithms? What are they being used for and how are we using them in Sheffield? In this talk I’ll explain what deep learning is, why
it’s considered exciting, and what the success stories are. I’ll also explain what the problems with these learning systems and how we
are trying to address these problems with our own class of deep architectures been developed in our group in Sheffield.Wed, 20 Mar 2013 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-sheffield13/deep-learning-what-is-it-and-what-are-we-doing-about-it.html
http://inverseprobability.com/talks/lawrence-sheffield13/deep-learning-what-is-it-and-what-are-we-doing-about-it.htmlLawrence-sheffield13How the Planets Affect Our Daily Lives: A Brief History of UncertaintyWithin the last 400 years scientists became able to predict the future. Crystal balls were replaced with computation. Uncertainty met mathematics. This talk gives a brief history of uncertainty and prediction. You will find out how planets affect who your Facebook friends are.Tue, 19 Mar 2013 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-scienceweek_wilfrids13/how-the-planets-affect-our-daily-lives-a-brief-history-of-uncertainty.html
http://inverseprobability.com/talks/lawrence-scienceweek_wilfrids13/how-the-planets-affect-our-daily-lives-a-brief-history-of-uncertainty.htmlLawrence-scienceweek_wilfrids13How the Planets Affect Our Daily Lives: A Brief History of UncertaintyWithin the last 400 years scientists became able to predict the future. Crystal balls were replaced with computation. Uncertainty met mathematics. This talk gives a brief history of uncertainty and prediction. You will find out how planets affect who your Facebook friends are.Mon, 18 Mar 2013 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-scienceweek_birley13/how-the-planets-affect-our-daily-lives-a-brief-history-of-uncertainty.html
http://inverseprobability.com/talks/lawrence-scienceweek_birley13/how-the-planets-affect-our-daily-lives-a-brief-history-of-uncertainty.htmlLawrence-scienceweek_birley13Variational <span>Gaussian</span> ProcessesIn this talk we will review the variational approximation to Gaussian processes which enables Bayesian learning of latent variables. We will focus in particular on a new explanation of the variational approach that also leads to stochastic variational inference for GPs.Mon, 11 Mar 2013 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-tuebingen_var13/variational-span-gaussian-span-processes.html
http://inverseprobability.com/talks/lawrence-tuebingen_var13/variational-span-gaussian-span-processes.htmlLawrence-tuebingen_var13Deep <span>Gaussian</span> ProcessesIn this talk we will introduce deep Gaussian process (GP) models. Deep GPs are a deep belief network based on Gaussian process mappings. The data is modeled as the output of a multivariate GP. The inputs to that Gaussian process are then governed by another GP. A single layer model is equivalent to a standard GP or the GP latent variable model (GPLVM). We perform inference in the model by approximate variational marginalization. This results in a strict lower bound on the marginal likelihood of the model which we use for model selection (number of layers and nodes per layer). Deep belief networks are typically applied to relatively large data sets using stochastic gradient descent for optimization. Our fully Bayesian treatment allows for the application of deep models even when data is scarce. Model selection by our variational bound shows that a five layer hierarchy is justified even when modelling a digit data set containing only 150 examples. In the seminar we will briefly review dimensionality reduction via Gaussian processes, before showing how this framework can be extended to build deep models.Mon, 11 Mar 2013 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-tuebingen13/deep-span-gaussian-span-processes.html
http://inverseprobability.com/talks/lawrence-tuebingen13/deep-span-gaussian-span-processes.htmlLawrence-tuebingen13Deep <span>Gaussian</span> ProcessesIn this talk we will introduce deep Gaussian process (GP) models. Deep GPs are a deep belief network based on Gaussian process mappings. The data is modeled as the output of a multivariate GP. The inputs to that Gaussian process are then governed by another GP. A single layer model is equivalent to a standard GP or the GP latent variable model (GPLVM). We perform inference in the model by approximate variational marginalization. This results in a strict lower bound on the marginal likelihood of the model which we use for model selection (number of layers and nodes per layer). Deep belief networks are typically applied to relatively large data sets using stochastic gradient descent for optimization. Our fully Bayesian treatment allows for the application of deep models even when data is scarce. Model selection by our variational bound shows that a five layer hierarchy is justified even when modelling a digit data set containing only 150 examples. In the seminar we will briefly review dimensionality reduction via Gaussian processes, before showing how this framework can be extended to build deep models.Wed, 30 Jan 2013 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-ucl13/deep-gaussian-processes.html
http://inverseprobability.com/talks/lawrence-ucl13/deep-gaussian-processes.htmlLawrence-ucl13Deep <span>Gaussian</span> ProcessesIn this talk we will introduce deep Gaussian process (GP) models. Deep GPs are a deep belief network based on Gaussian process mappings. The data is modeled as the output of a multivariate GP. The inputs to that Gaussian process are then governed by another GP. A single layer model is equivalent to a standard GP or the GP latent variable model (GPLVM). We perform inference in the model by approximate variational marginalization. This results in a strict lower bound on the marginal likelihood of the model which we use for model selection (number of layers and nodes per layer). Deep belief networks are typically applied to relatively large data sets using stochastic gradient descent for optimization. Our fully Bayesian treatment allows for the application of deep models even when data is scarce. Model selection by our variational bound shows that a five layer hierarchy is justified even when modelling a digit data set containing only 150 examples. In the seminar we will first review dimensionality reduction via Gaussian processes, before showing how this framework can be extended to build deep models.Thu, 24 Jan 2013 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-aalto13/deep-span-gaussian-span-processes.html
http://inverseprobability.com/talks/lawrence-aalto13/deep-span-gaussian-span-processes.htmlLawrence-aalto13Reproducible Research: <span>Lessons</span> from Machine LearningTue, 15 Jan 2013 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-reproducible13/reproducible-research-span-lessons-span-from-machine-learning.html
http://inverseprobability.com/talks/lawrence-reproducible13/reproducible-research-span-lessons-span-from-machine-learning.htmlLawrence-reproducible13Machine Learning and the Life Sciences: from Modelling to MedicineFri, 11 Jan 2013 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-infection13/machine-learning-and-the-life-sciences-from-modelling-to-medicine.html
http://inverseprobability.com/talks/lawrence-infection13/machine-learning-and-the-life-sciences-from-modelling-to-medicine.htmlLawrence-infection13Life, the Universe and Machine LearningWhat is Machine Learning? Why is it useful for us? Machine learning algorithms are the engines that are driving forward an intelligent internet. They are allowing us to uncover the causes of cancer and helping us understand the way the universe is put together. They are suggesting who your friends are on facebook, enabling driverless cars and causing flagging potentially fraudulent transactions on your credit card. To put it simply, machine learning is about understanding data. In this lecture I will try and give a sense of the challenges we face in machine learning, with a particular focus on those that have inspired my research. We will look at applications of data modelling from the early 18th century to the present, and see how they relate to modern machine learning. There will be a particular focus on dealing with <i>uncertainty</i>: something humans are good at, but an area where computers have typically struggled. We will emphasize the role of uncertainty in data modelling and hope to persuade the audience that correct handling of uncertainty may be one of the keys to intelligent systems.Thu, 06 Sep 2012 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-inaugural12/life-the-universe-and-machine-learning.html
http://inverseprobability.com/talks/lawrence-inaugural12/life-the-universe-and-machine-learning.htmlLawrence-inaugural12Model Based Target Identification from Expression DataA simple approach to target identification through gene expression studies has been to cluster the expression profiles and look for coregulated genes within clusters. Within systems biology mechanistic models of gene expression are typically constructed through differential equations. mRNA’s production is taken to be proportional to transcription factor activity (with the proportionality given by the sensitivity) and the mRNA is assumed to decay at a particular rate. The assumption that coregulated genes have similar profiles is equivalent to assuming both the decay and the sensitivity are high.\
\
Typically researchers either use a data driven approach (such as clustering) or a model based approach (such as differential equations). In this talk we advocate hybrid techniques which have aspects of the mechanistic and data driven models. We combine simple differential equation models with Gaussian process priors to make probabilistic models with mechanistic underpinnings. We show applications in target identification from mRNA measurements.Fri, 27 Jul 2012 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-ucla12b/model-based-target-identification-from-expression-data.html
http://inverseprobability.com/talks/lawrence-ucla12b/model-based-target-identification-from-expression-data.htmlLawrence-ucla12bA Brief Introduction to <span>G</span>aussian ProcessesGaussian processes are non-parametric probabilistic models for function representation. In this tutorial we give a brief introduction to Gaussian process models. Using simple examples we show how, with particular choices for covariance functions (analagous to a kernel matrix in kernel methods), we can perform inference about functions using only data sampled from those functions. We give overview how the probabilistic interpretation allows us to fit the parameters of the covariance function.Fri, 27 Jul 2012 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-ucla12a/a-brief-introduction-to-span-g-span-aussian-processes.html
http://inverseprobability.com/talks/lawrence-ucla12a/a-brief-introduction-to-span-g-span-aussian-processes.htmlLawrence-ucla12aBridging the Gap Between Computational Biology and Systems BiologyWed, 04 Jul 2012 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-pathsoc12/bridging-the-gap-between-computational-biology-and-systems-biology.html
http://inverseprobability.com/talks/lawrence-pathsoc12/bridging-the-gap-between-computational-biology-and-systems-biology.htmlLawrence-pathSoc12Kernels for Vector Valued FunctionsIn this talk we review kernels for vector valued functions from the perspective of Gaussian processes. Deriving a multiple output Gaussian process from the perspective of a linear dynamical system (Kalman Filter) we introduce the Intrinsic Coregionalization Model and the Linear Model of Coregionalization. We discuss how they relate to multi-task learning with GPs and the Semi Parametric Latent Factor model. Finally, we will introduce convolutional process models from the perspective of the latent force model.Sat, 30 Jun 2012 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-icmlvector12/kernels-for-vector-valued-functions.html
http://inverseprobability.com/talks/lawrence-icmlvector12/kernels-for-vector-valued-functions.htmlLawrence-icmlVector12Everything You Want to Know About <span>G</span>aussian Processes: <span>G</span>aussian Process RegressionSat, 16 Jun 2012 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-cvpr12_1/everything-you-want-to-know-about-span-g-span-aussian-processes-span-g-span-aussian-pro.html
http://inverseprobability.com/talks/lawrence-cvpr12_1/everything-you-want-to-know-about-span-g-span-aussian-processes-span-g-span-aussian-pro.htmlLawrence-cvpr12_1Everything You Want to Know About <span>G</span>aussian Processes: Multioutput Covariances and Mechanistic ModelsSat, 16 Jun 2012 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-cvpr12_2/everything-you-want-to-know-about-span-g-span-aussian-processes-multioutput-covariances.html
http://inverseprobability.com/talks/lawrence-cvpr12_2/everything-you-want-to-know-about-span-g-span-aussian-processes-multioutput-covariances.htmlLawrence-cvpr12_2<span>G</span>aussian Processes in Computational Biology Tutorial: Session 2Tue, 12 Jun 2012 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-biopredyn12_2/span-g-span-aussian-processes-in-computational-biology-tutorial-session-2.html
http://inverseprobability.com/talks/lawrence-biopredyn12_2/span-g-span-aussian-processes-in-computational-biology-tutorial-session-2.htmlLawrence-biopredyn12_2<span>G</span>aussian Processes in Computational Biology Tutorial: Multioutput <span>G</span>aussian Processes and Mechanistic ModelsTue, 12 Jun 2012 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-biopredyn12_1/span-g-span-aussian-processes-in-computational-biology-tutorial-multioutput-span-g-spa.html
http://inverseprobability.com/talks/lawrence-biopredyn12_1/span-g-span-aussian-processes-in-computational-biology-tutorial-multioutput-span-g-spa.htmlLawrence-biopredyn12_1Latent Force Models: Bridging the Divide between Mechanistic and Data Modelling ParadigmsPhysics based approaches to data modeling involve constructing an accurate mechanistic model of data, often based on differential equations. Machine learning and statistical approaches are typically data driven—perhaps through regularized function approximation. These two approaches to data modeling are often seen as polar opposites, but in reality they are two different ends to a spectrum of approaches we might take. In this talk we introduce latent force models. Latent force models are a new approach to data representation that model data through unknown forcing functions that drive differential equation models. By treating the unknown forcing functions with Gaussian process priors we can create probabilistic models that exhibit particular physical characteristics of interest, for example, in dynamical systems resonance and inertia. This allows us to perform a synthesis of the data driven and physical modeling paradigms.Wed, 02 May 2012 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-liverpool12/latent-force-models-bridging-the-divide-between-mechanistic-and-data-modelling-paradigm.html
http://inverseprobability.com/talks/lawrence-liverpool12/latent-force-models-bridging-the-divide-between-mechanistic-and-data-modelling-paradigm.htmlLawrence-liverpool12What is Machine Learning?Sun, 15 Apr 2012 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-mlssfour12/what-is-machine-learning.html
http://inverseprobability.com/talks/lawrence-mlssfour12/what-is-machine-learning.htmlLawrence-mlssFour12Nonlinear Probabilistic Dimensionality ReductionFri, 13 Apr 2012 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-mlssthree12/nonlinear-probabilistic-dimensionality-reduction.html
http://inverseprobability.com/talks/lawrence-mlssthree12/nonlinear-probabilistic-dimensionality-reduction.htmlLawrence-mlssThree12Spectral Approaches to Dimensionality ReductionThu, 12 Apr 2012 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-mlsstwo12/spectral-approaches-to-dimensionality-reduction.html
http://inverseprobability.com/talks/lawrence-mlsstwo12/spectral-approaches-to-dimensionality-reduction.htmlLawrence-mlssTwo12Dimensionality Reduction: Motivation and Linear ModelsWed, 11 Apr 2012 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-mlssone12/dimensionality-reduction-motivation-and-linear-models.html
http://inverseprobability.com/talks/lawrence-mlssone12/dimensionality-reduction-motivation-and-linear-models.htmlLawrence-mlssOne12Latent Force Models: Combining the Mechanistic and Data Driven Modelling ParadigmsThe main focus of machine learning is to combine data with assumptions that reflect our belief about the regularity of the world. This, then, allows us to generalize and make new predictions for ‘test data’. Relative to other modelling paradigms such as those found in physics that are based on mechanistic understandings of the world, models in machine learning typically make only weak assumptions about data.\
\
In this talk, we argue that these weak assumptions are also mechanistic in nature. In particular, a very common assumption is smoothness, which can arise through the heat equation or other models of diffusion. Our assumption of smoothness reflects our belief in an underlying physical world in which smoothness is the norm. Strong mechanistic models, such as those used in computational fluid dynamics, climate etc. typically impose much more rigid constraints on the data and are often inappropriate for machine learning tasks where the model needs to be adaptive and should still perform well even when our mechanistic assumptions are not completely fulfilled. These strong mechanistic frameworks can, however, incorporate regularities beyond smoothness. Systems with inertia exhibit resonance and oscillation and these can be easily incorporated with strong mechanistic assumptions.\
\
We believe that the area between the strong and weak mechanistic paradigms should be a focus for much more research. For many interesting datasets we need adaptive models which include mechanistic assumptions. The latent force modeling paradigm is one way of approaching this which relies on the combination of differential equation systems which are driven, or have their initial or boundary conditions set, by Gaussian processes. The Gaussian processes provide the necessary adaptability and the differential equation encodes mechanistic assumptions. In this talk we introduce the model and demonstrate results in motion capture date and, given time, computational biology.Wed, 28 Mar 2012 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-rank12/latent-force-models-combining-the-mechanistic-and-data-driven-modelling-paradigms.html
http://inverseprobability.com/talks/lawrence-rank12/latent-force-models-combining-the-mechanistic-and-data-driven-modelling-paradigms.htmlLawrence-rank12Latent force models: Combining Probabilistic and Mechanistic ModellingPhysics based approaches to data modeling involve constructing an accurate mechanistic model of data, often based on differential equations. Statistical and machine learning approaches are typically data driven-perhaps through regularized function approximation.\
\
These two approaches to data modeling are often seen as polar opposites, but in reality they are two different ends to a spectrum of approaches we might take. Physics based approaches can be seen as strongly mechanistic, the mechanistic assumptions are hard encoded into the model. Data-driven approaches do incorporate assumptions that might be seen as being derived from some underlying mechanism, such as smoothness. In this sense they are weakly mechanistic.\
\
In this talk we introduce latent force models. Latent force models are a new approach to data representation that model data through unknown forcing functions that drive differential equation models. By treating the unknown forcing functions with Gaussian process priors we can create probabilistic models that exhibit particular physical characteristics of interest, for example, in dynamical systems resonance and inertia. This allows us to perform a synthesis of the data driven and physical modeling paradigms. A moderately mechanistic approach. We show an application in modelling of human motion capture data.Mon, 13 Feb 2012 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-oxfordlatent12/latent-force-models-combining-probabilistic-and-mechanistic-modelling.html
http://inverseprobability.com/talks/lawrence-oxfordlatent12/latent-force-models-combining-probabilistic-and-mechanistic-modelling.htmlLawrence-oxfordLatent12A Unifying Review of Spectral Methods for Dimensionality ReductionIn this tutorial we will review spectral approaches to dimensionality reduction, introducing a unifying probabilistic perspective. Our unifying perspective is based on the maximum entropy principle and the resulting probabilistic models are based on GRFs. We will review maximum variance unfolding, Laplacian eigenmaps, locally linear embeddings and Isomap. Under the framework, these approaches can be divided into those that preserve local distances and those that don’t. For two small data sets we show that local distance preserving methods tend to perform better. Finally we use the unifying framework to relate these approaches to the Gaussian process latent variable model.Mon, 13 Feb 2012 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-oxfordunify12/a-unifying-review-of-spectral-methods-for-dimensionality-reduction.html
http://inverseprobability.com/talks/lawrence-oxfordunify12/a-unifying-review-of-spectral-methods-for-dimensionality-reduction.htmlLawrence-oxfordUnify12Model Based Target Identification from Expression DataA simple approach to target identification through gene expression studies has been to cluster the expression profiles and look for coregulated genes within clusters. Within systems biology mechanistic models of gene expression are typically constructed through differential equations. mRNA’s production is taken to be proportional to transcription factor activity (with the proportionality given by the sensitivity) and the mRNA is assumed to decay at a particular rate. The assumption that coregulated genes have similar profiles is equivalent to assuming both the decay and the sensitivity are high.\
\
Typically researchers either use a data driven approach (such as clustering) or a model based approach (such as differential equations). In this talk we advocate hybrid techniques which have aspects of the mechanistic and data driven models. We combine simple differential equation models with Gaussian process priors to make probabilistic models with mechanistic underpinnings. We show applications in target identification from mRNA measurements.Mon, 06 Feb 2012 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-cruk12/model-based-target-identification-from-expression-data.html
http://inverseprobability.com/talks/lawrence-cruk12/model-based-target-identification-from-expression-data.htmlLawrence-cruk12Mon, 12 Dec 2011 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-biopredyn11/.html
http://inverseprobability.com/talks/lawrence-biopredyn11/.htmlLawrence-BioPreDyn11A Maximum Entropy Perspective on Spectral Dimensionality ReductionSpectral approaches to dimensionality reduction typically reduce the dimensionality of a data set through taking the eigenvectors of a Laplacian or a similarity matrix. Classical multidimensional scaling also makes use of the eigenvectors of a similarity matrix. In this talk we introduce a maximum entropy approach to designing this similarity matrix. The approach is closely related to maximum variance unfolding. Other spectral approaches, e.g. locally linear embeddings, turn out to also be closely related. These methods can be seen as a sparse Gaussian graphical model where correlations between data points (rather than across data features) are specified in the graph. The hope is that this unifying perspective will allow the relationships between these methods to be better understood and will also provide the groundwork for further research.Wed, 16 Nov 2011 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-cambridge11/a-maximum-entropy-perspective-on-spectral-dimensionality-reduction.html
http://inverseprobability.com/talks/lawrence-cambridge11/a-maximum-entropy-perspective-on-spectral-dimensionality-reduction.htmlLawrence-cambridge11Model Based Target Identification from Expression DataA simple approach to target identification through gene expression studies has been to cluster the expression profiles and look for coregulated genes within clusters. Within systems biology mechanistic models of gene expression are typically constructed through differential equations. mRNA’s production is taken to be proportional to transcription factor activity (with the proportionality given by the sensitivity) and the mRNA is assumed to decay at a particular rate. The assumption that coregulated genes have similar profiles is equivalent to assuming both the decay and the sensitivity are high.\
\
Typically researchers either use a data driven approach (such as clustering) or a model based approach (such as differential equations). In this talk we advocate hybrid techniques which have aspects of the mechanistic and data driven models. We combine simple differential equation models with Gaussian process priors to make probabilistic models with mechanistic underpinnings. We show applications in target identification from mRNA measurements.Wed, 12 Oct 2011 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-liverpool11/model-based-target-identification-from-expression-data.html
http://inverseprobability.com/talks/lawrence-liverpool11/model-based-target-identification-from-expression-data.htmlLawrence-liverpool11Between Systems and Data-driven Modeling for Computational Biology: Target Identification with <span>Gaussian</span> ProcessesA simple approach to target identification through gene expression studies has been to cluster the expression profiles and look for coregulated genes within clusters. Within systems biology mechanistic models of gene expression are typically constructed through differential equations. mRNA’s production is taken to be proportional to transcription factor activity (with the proportionality given by the sensitivity) and the mRNA is assumed to decay at a particular rate. The assumption that coregulated genes have similar profiles is equivalent to assuming both the decay and the sensitivity are high.\
\
Typically researchers either use a data driven approach (such as clustering) or a model based approach (such as differential equations). In this talk we advocate hybrid techniques which have aspects of the mechanistic and data driven models. We combine simple differential equation models with Gaussian process priors to make probabilistic models with mechanistic underpinnings. We show applications in target identification from mRNA measurements.Sat, 10 Sep 2011 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-abcd11/between-systems-and-data-driven-modeling-for-computational-biology-target-identificatio.html
http://inverseprobability.com/talks/lawrence-abcd11/between-systems-and-data-driven-modeling-for-computational-biology-target-identificatio.htmlLawrence-abcd11Latent Force ModelsPhysics based approaches to data modeling involve constructing an accurate mechanistic model of data, often based on differential equations. Statistical and machine learning approaches are typically data driven—perhaps through regularized function approximation.\
\
These two approaches to data modeling are often seen as polar opposites, but in reality they are two different ends to a spectrum of approaches we might take. Physics based approaches can be seen as *strongly mechanistic*, the mechanistic assumptions are hard encoded into the model. Data-driven approaches do incorporate assumptions that might be seen as being derived from some underlying mechanism, such as smoothness. In this sense they are *weakly mechanistic*.\
\
In this talk we introduce latent force models. Latent force models are a new approach to data representation that model data through unknown forcing functions that drive differential equation models. By treating the unknown forcing functions with Gaussian process priors we can create probabilistic models that exhibit particular physical characteristics of interest, for example, in dynamical systems resonance and inertia. This allows us to perform a synthesis of the data driven and physical modeling paradigms. A *moderately mechanistic* approach. We show an application in modelling of human motion capture data.Tue, 06 Sep 2011 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-bayes11/latent-force-models.html
http://inverseprobability.com/talks/lawrence-bayes11/latent-force-models.htmlLawrence-bayes11Gaussian Processes and Probabilistic Models for Dimensionality ReductionIn this talk we present an overview of probabilistic approaches to dimensionality reduction and probabilistic interpretations of dimensionality reduction. We start by reviewing spectral methods and then turn to probabilistic PCA and the Gaussian process latent variable model.Thu, 25 Aug 2011 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-dagstuhl11/gaussian-processes-and-probabilistic-models-for-dimensionality-reduction.html
http://inverseprobability.com/talks/lawrence-dagstuhl11/gaussian-processes-and-probabilistic-models-for-dimensionality-reduction.htmlLawrence-dagstuhl11Model Based Target Identification from Expression DataA simple approach to target identification through gene expression studies has been to cluster the expression profiles and look for coregulated genes within clusters. Within systems biology mechanistic models of gene expression are typically constructed through differential equations. mRNA’s production is taken to be proportional to transcription factor activity (with the proportionality given by the sensitivity) and the mRNA is assumed to decay at a particular rate. The assumption that coregulated genes have similar profiles is equivalent to assuming both the decay and the sensitivity are high.\
\
Typically researchers either use a data driven approach (such as clustering) or a model based approach (such as differential equations). In this talk we advocate hybrid techniques which have aspects of the mechanistic and data driven models. We combine simple differential equation models with Gaussian process priors to make probabilistic models with mechanistic underpinnings. We show applications in target identification from mRNA measurements.Tue, 07 Jun 2011 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-krebs11/model-based-target-identification-from-expression-data.html
http://inverseprobability.com/talks/lawrence-krebs11/model-based-target-identification-from-expression-data.htmlLawrence-krebs11A unifying probabilistic perspective on spectral approaches to dimensionality reductionSpectral approaches to dimensionality reduction typically reduce the dimensionality of a data set through taking the eigenvectors of a Laplacian or a similarity matrix. Classical multidimensional scaling also makes use of the eigenvectors of a similarity matrix. In this talk we introduce a maximum entropy approach to designing this similarity matrix. The approach is closely related to maximum variance unfolding. Other spectral approaches such as locally linear embeddings and Laplacian eigenmaps also turn out to be closely related. Each method can be seen as a sparse Gaussian graphical model where correlations between data points (rather than across data features) are specified in the graph. This also suggests optimization via sparse inverse covariance techniques such as the graphical LASSO. The hope is that this unifying perspective will allow the relationships between these methods to be better understood and will also provide the groundwork for further research.Tue, 31 May 2011 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-bonn11/a-unifying-probabilistic-perspective-on-spectral-approaches-to-dimensionality-reduction.html
http://inverseprobability.com/talks/lawrence-bonn11/a-unifying-probabilistic-perspective-on-spectral-approaches-to-dimensionality-reduction.htmlLawrence-bonn11Advanced Use of <span>G</span>aussian ProcessesThu, 07 Apr 2011 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-siena11b/advanced-use-of-span-g-span-aussian-processes.html
http://inverseprobability.com/talks/lawrence-siena11b/advanced-use-of-span-g-span-aussian-processes.htmlLawrence-siena11bIntroduction to <span>G</span>aussian ProcessesWed, 06 Apr 2011 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-siena11a/introduction-to-span-g-span-aussian-processes.html
http://inverseprobability.com/talks/lawrence-siena11a/introduction-to-span-g-span-aussian-processes.htmlLawrence-siena11aLatent Force ModelsPhysics based approaches to data modeling involve constructing an accurate mechanistic model of data, often based on differential equations. Machine learning approaches are typically data driven—perhaps through regularized function approximation. These two approaches to data modeling are often seen as polar opposites, but in reality they are two different ends to a spectrum of approaches we might take. In this talk we introduce latent force models. Latent force models are a new approach to data representation that model data through unknown forcing functions that drive differential equation models. By treating the unknown forcing functions with Gaussian process priors we can create probabilistic models that exhibit particular physical characteristics of interest, for example, in dynamical systems resonance and inertia. This allows us to perform a synthesis of the data driven and physical modeling paradigms. We will show applications of these models in systems biology and (given time) modelling of human motion capture data.Wed, 16 Mar 2011 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-exeter11/latent-force-models.html
http://inverseprobability.com/talks/lawrence-exeter11/latent-force-models.htmlLawrence-exeter11Probabilistic Dimensional Reduction with the <span>G</span>aussian Process Latent Variable ModelDensity modelling in high dimensions is a very difficult problem. Traditional approaches, such as mixtures of Gaussians, typically fail to capture the structure of data sets in high dimensional spaces. In this talk we will argue that for many data sets of interest, the data can be represented as a lower dimensional manifold immersed in the higher dimensional space. We will then present the Gaussian Process Latent Variable Model (GP-LVM), a non-linear probabilistic variant of principal component analysis (PCA) which implicitly assumes that the data lies on a lower dimensional space. Having introduced the GP-LVM we will review extensions to the algorithm. Given time we will review dynamical extensions, Bayesian approaches to dimensionality determination, learning of large data sets. We will demonstrate the application of the model and its extensions to a range of data sets, including human motion data, speech data and video.Wed, 09 Mar 2011 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-loughborough11/probabilistic-dimensional-reduction-with-the-span-g-span-aussian-process-latent-variabl.html
http://inverseprobability.com/talks/lawrence-loughborough11/probabilistic-dimensional-reduction-with-the-span-g-span-aussian-process-latent-variabl.htmlLawrence-loughborough11A Unifying Probabilistic Perspective on Spectral Approaches to Dimensionality ReductionSpectral approaches to dimensionality reduction typically reduce the dimensionality of a data set through taking the eigenvectors of a Laplacian or a similarity matrix. Classical multidimensional scaling also makes use of the eigenvectors of a similarity matrix. In this talk we introduce a maximum entropy approach to designing this similarity matrix. The approach is closely related to maximum variance unfolding. Other spectral approaches such as locally linear embeddings and Laplacian eigenmaps also turn out to be closely related. Each method can be seen as a sparse Gaussian graphical model where correlations between data points (rather than across data features) are specified in the graph. This also suggests optimization via sparse inverse covariance techniques such as the graphical LASSO. The hope is that this unifying perspective will allow the relationships between these methods to be better understood and will also provide the groundwork for further research.Tue, 01 Mar 2011 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-edinburgh11/a-unifying-probabilistic-perspective-on-spectral-approaches-to-dimensionality-reduction.html
http://inverseprobability.com/talks/lawrence-edinburgh11/a-unifying-probabilistic-perspective-on-spectral-approaches-to-dimensionality-reduction.htmlLawrence-edinburgh11Between Systems and Data-driven Modeling for Computational Biology: Target Identification with <span>Gaussian</span> ProcessesA simple approach to target identification through gene expression studies has been to cluster the expression profiles and look for coregulated genes within clusters. Within systems biology mechanistic models of gene expression are typically constructed through differential equations. mRNA’s production is taken to be proportional to transcription factor activity (with the proportionality given by the sensitivity) and the mRNA is assumed to decay at a particular rate. The assumption that coregulated genes have similar profiles is equivalent to assuming both the decay and the sensitivity are high.\
\
Typically researchers either use a data driven approach (such as clustering) or a model based approach (such as differential equations). In this talk we advocate hybrid techniques which have aspects of the mechanistic and data driven models. We combine simple differential equation models with Gaussian process priors to make probabilistic models with mechanistic underpinnings. We show applications in target identification from mRNA measurements.Thu, 27 Jan 2011 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-smpgd11/between-systems-and-data-driven-modeling-for-computational-biology-target-identificatio.html
http://inverseprobability.com/talks/lawrence-smpgd11/between-systems-and-data-driven-modeling-for-computational-biology-target-identificatio.htmlLawrence-smpgd11A Probabilistic Perspective on Spectral Dimensionality ReductionSpectral approaches to dimensionality reduction typically reduce the dimensionality of a data set through taking the eigenvectors of a Laplacian or a similarity matrix. Classical multidimensional scaling also makes use of the eigenvectors of a similarity matrix. In this talk we introduce a maximum entropy approach to designing this similarity matrix. The approach is closely related to maximum variance unfolding and other spectral approaches such as locally linear embeddings and Laplacian eigenmaps also turn out to be closely related. Each method can be seen as a sparse Gaussian graphical model where correlations between data points (rather than across data features) are specified in the graph. This also suggests optimization via sparse inverse covariance techniques such as the graphical LASSO. The hope is that this unifying perspective will allow the relationships between these methods to be better understood and will also provide the groundwork for further research.Sat, 11 Dec 2010 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-nipsw10/a-probabilistic-perspective-on-spectral-dimensionality-reduction.html
http://inverseprobability.com/talks/lawrence-nipsw10/a-probabilistic-perspective-on-spectral-dimensionality-reduction.htmlLawrence-nipsw10A Probabilistic Perspective on Spectral Dimensionality ReductionSpectral approaches to dimensionality reduction typically reduce the dimensionality of a data set through taking the eigenvectors of a Laplacian or a similarity matrix. Classical multidimensional scaling also makes use of the eigenvectors of a similarity matrix. In this talk we introduce a maximum entropy approach to designing this similarity matrix. The approach is closely related to maximum variance unfolding and other spectral approaches such as locally linear embeddings and Laplacian eigenmaps also turn out to be closely related. Each method can be seen as a sparse Gaussian graphical model where correlations between data points (rather than across data features) are specified in the graph. This also suggests optimization via sparse inverse covariance techniques such as the graphical LASSO. The hope is that this unifying perspective will allow the relationships between these methods to be better understood and will also provide the groundwork for further research.Thu, 11 Nov 2010 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-aaai10/a-probabilistic-perspective-on-spectral-dimensionality-reduction.html
http://inverseprobability.com/talks/lawrence-aaai10/a-probabilistic-perspective-on-spectral-dimensionality-reduction.htmlLawrence-aaai10Latent Force ModelsPhysics based approaches to data modeling involve constructing an accurate mechanistic model of data, often based on differential equations. Machine learning approaches are typically data driven—perhaps through regularized function approximation. These two approaches to data modeling are often seen as polar opposites, but in reality they are two different ends to a spectrum of approaches we might take. In this talk we introduce latent force models. Latent force models are a new approach to data representation that model data through unknown forcing functions that drive differential equation models. By treating the unknown forcing functions with Gaussian process priors we can create probabilistic models that exhibit particular physical characteristics of interest, for example, in dynamical systems resonance and inertia. This allows us to perform a synthesis of the data driven and physical modeling paradigms. We will show applications of these models in systems biology and (given time) modelling of human motion capture data.Thu, 04 Nov 2010 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-lfmsheffield10/latent-force-models.html
http://inverseprobability.com/talks/lawrence-lfmsheffield10/latent-force-models.htmlLawrence-lfmSheffield10A Probabilistic Perspective on Spectral Dimensionality ReductionSpectral approaches to dimensionality reduction typically reduce the dimensionality of a data set through taking the eigenvectors of a Laplacian or a similarity matrix. Classical multidimensional scaling also makes use of the eigenvectors of a similarity matrix. In this talk we introduce a maximum entropy approach to designing this similarity matrix. The approach is closely related to maximum variance unfolding and other spectral approaches such as locally linear embeddings and Laplacian eigenmaps also turn out to be closely related. Each method can be seen as a sparse Gaussian graphical model where correlations between data points (rather than across data features) are specified in the graph. This also suggests optimization via sparse inverse covariance techniques such as the graphical LASSO. The hope is that this unifying perspective will allow the relationships between these methods to be better understood and will also provide the groundwork for further research.Wed, 20 Oct 2010 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-aalto10/a-probabilistic-perspective-on-spectral-dimensionality-reduction.html
http://inverseprobability.com/talks/lawrence-aalto10/a-probabilistic-perspective-on-spectral-dimensionality-reduction.htmlLawrence-aalto10Bayesian approaches to Transcription Factor Target IdentificationA simple approach to target identification through gene expression studies has been to cluster the expression profiles and look for coregulated genes within clusters. Within systems biology mechanistic models of gene expression are typically constructed through differential equations. mRNA’s production is taken to be proportional to transcription factor activity (with the proportionality given by the sensitivity) and the mRNA is assumed to decay at a particular rate. The assumption that coregulated genes have similar profiles is equivalent to assuming both the decay and the sensitivity are high.\
\
In this lecture we introduce Bayesian approaches to target identification which make use of sampling approaches to rank candidate lists of targets. We will begin with an introduction to the target identification problem and an overview of the power of Bayesian approaches in solving it. We will then consider how probabilistic models such as Gaussian processes can be used for ranking potential targets of a transcription factor. These models are simple enough to allow genome wide target identification, but rich enough to encode dynamical behavior that, allowing us to identify putative targets even when decay rates are low.Sun, 10 Oct 2010 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-eurogene10/bayesian-approaches-to-transcription-factor-target-identification.html
http://inverseprobability.com/talks/lawrence-eurogene10/bayesian-approaches-to-transcription-factor-target-identification.htmlLawrence-eurogene10Making Implementations Available for the Research CommunityMachine learning research is either inspired by a particular application, or by a general desire to make technology more “inteligent”. In modern machine learning most methodological development is mathematically inspired and results in an algorithm for optimization or fitting of a model to data. Design choices in implementation of an algorithm can have a significant effect on the quality of results. Decisions such as model initializaiton and data pre-processing are all part of the implementation. Necessarily, space constraints sometimes mean that such details are not included in the associated paper. It seems clear that the paper only tells part of the story. Implementations need to be made available at the time of submission of the paper, so that the full story may be followed. In our research group we have done this since 2001. In this talk I will make the arguments in favour of doing this universally and give personal experiences of the results.Wed, 06 Oct 2010 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-validation10/making-implementations-available-for-the-research-community.html
http://inverseprobability.com/talks/lawrence-validation10/making-implementations-available-for-the-research-community.htmlLawrence-validation10Latent Force ModelsPhysics based approaches to data modeling involve constructing an accurate mechanistic model of data, often based on differential equations. Machine learning approaches are typically data driven—perhaps through regularized function approximation. These two approaches to data modeling are often seen as polar opposites, but in reality they are two different ends to a spectrum of approaches we might take. In this talk we introduce latent force models. Latent force models are a new approach to data representation that model data through unknown forcing functions that drive differential equation models. By treating the unknown forcing functions with Gaussian process priors we can create probabilistic models that exhibit particular physical characteristics of interest, for example, in dynamical systems resonance and inertia. This allows us to perform a synthesis of the data driven and physical modeling paradigms. We will show applications of these models in systems biology and (given time) modelling of human motion capture data.Mon, 27 Sep 2010 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-phylogenetics10/latent-force-models.html
http://inverseprobability.com/talks/lawrence-phylogenetics10/latent-force-models.htmlLawrence-phylogenetics10<span>PRIB</span> Tutorial: <span>G</span>aussian Processes and Gene RegulationComputational biology models are often missing information, such as the concentration of biochemical species of interest. One approach to dealing with this missing information is to place a probabilistic prior over the missing data. One possible choice for such a prior is a Gaussian process.\
\
In this tutorial we will give an introduction to Gaussian processes. We will give simple examples of Gaussian processes in regression and interpolation. We will then show how Gaussian processes can be incorporated with differential equation models to give probabilistic models for transcription. Such models can then be used to rank potential targets of given transcription factors.Wed, 22 Sep 2010 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-tutorialprib10/span-prib-span-tutorial-span-g-span-aussian-processes-and-gene-regulation.html
http://inverseprobability.com/talks/lawrence-tutorialprib10/span-prib-span-tutorial-span-g-span-aussian-processes-and-gene-regulation.htmlLawrence-tutorialPRIB10Between Systems and Data-driven Modeling for Computational Biology: Target Identification with <span>Gaussian</span> ProcessesA simple approach to target identification through gene expression studies has been to cluster the expression profiles and look for coregulated genes within clusters. Within systems biology mechanistic models of gene expression are typically constructed through differential equations. mRNA’s production is taken to be proportional to transcription factor activity (with the proportionality given by the sensitivity) and the mRNA is assumed to decay at a particular rate. The assumption that coregulated genes have similar profiles is equivalent to assuming both the decay and the sensitivity are high. Typically researchers either use a data driven approach (such as clustering) or a model based approach (such as differential equations). In this talk we advocate hybrid techniques which have aspects of the mechanistic and data driven models. We combine simple differential equation models with Gaussian process priors to make probabilistic models with mechanistic underpinnings. We show applications in target identification from mRNA measurements.Tue, 27 Jul 2010 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-ibsb10/between-systems-and-data-driven-modeling-for-computational-biology-target-identificatio.html
http://inverseprobability.com/talks/lawrence-ibsb10/between-systems-and-data-driven-modeling-for-computational-biology-target-identificatio.htmlLawrence-ibsb10Latent Force ModelsPhysics based approaches to data modeling involve constructing an accurate mechanistic model of data, often based on differential equations. Machine learning approaches are typically data driven— perhaps through regularized function approximation.\
\
These two approaches to data modeling are often seen as polar opposites, but in reality they are two different ends to a spectrum of approaches we might take.\
\
In this talk we introduce latent force models. Latent force models are a new approach to data representation that model data through unknown forcing functions that drive differential equation models. By treating the unknown forcing functions with Gaussian process priors we can create probabilistic models that exhibit particular physical characteristics of interest, for example, in dynamical systems resonance and inertia. This allows us to perform a synthesis of the data driven and physical modeling paradigms. We will show applications of these models in systems biology and modelling of human motion capture data.Mon, 01 Mar 2010 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-inference10/latent-force-models.html
http://inverseprobability.com/talks/lawrence-inference10/latent-force-models.htmlLawrence-inference10Transfer Learning and Multiple Output Kernel FunctionsA standard Bayesian approach to transfer learning is to construct hierarchical probabilistic models. Learning tasks are typically related in the model through conditional independencies of the variables/parameters. Many of the variables are unobserved. Marginalization of the unobserved variables and Bayesian treatment of parameters induces structure and correlations between the tasks. Gaussian processes are prior distributions over functions: kernel functions are the covariances associated with these priors. A Gaussian process can be set up to have multiple outputs. However, for these outputs to have correlation between them a covariance function that models correlations between outputs is required. Equivalently we need to develop multiple output kernel functions (also known as multitask kernel functions, or structured output kernels). In this talk we will briefly review work in creating multiple output kernels before focusing on models represented by a convolution processes. We will arrive at convolutional processes through physical interpretations of our models. We will try to illustrate these models with a range of real world examples of both transfer learning and other applications.Sat, 12 Dec 2009 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-tlsd09/transfer-learning-and-multiple-output-kernel-functions.html
http://inverseprobability.com/talks/lawrence-tlsd09/transfer-learning-and-multiple-output-kernel-functions.htmlLawrence-tlsd09Latent Force ModelsPhysics based approaches to data modeling involve constructing an accurate mechanistic model of data, often based on differential equations. Machine learning approaches are typically data driven— perhaps through regularized function approximation.\
\
These two approaches to data modeling are often seen as polar opposites, but in reality they are two different ends to a spectrum of approaches we might take.\
\
In this talk we introduce latent force models. Latent force models are a new approach to data representation that model data through unknown forcing functions that drive differential equation models. By treating the unknown forcing functions with Gaussian process priors we can create probabilistic models that exhibit particular physical characteristics of interest, for example, in dynamical systems resonance and inertia. This allows us to perform a synthesis of the data driven and physical modeling paradigms. We will show applications of these models in systems biology and modelling of human motion capture data.Wed, 25 Nov 2009 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-kcl09/latent-force-models.html
http://inverseprobability.com/talks/lawrence-kcl09/latent-force-models.htmlLawrence-kcl09Nonlinear Response in <span>G</span>aussian Process Models of TranscriptionThu, 29 Oct 2009 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-tigem09/nonlinear-response-in-span-g-span-aussian-process-models-of-transcription.html
http://inverseprobability.com/talks/lawrence-tigem09/nonlinear-response-in-span-g-span-aussian-process-models-of-transcription.htmlLawrence-tigem09Model Based Target Identification from Gene Expression with <span>G</span>aussian ProcessesA simple approach to target identification through gene expression studies has been to cluster the expression profiles and look for coregulated genes within clusters. Within systems biology mechanistic models of gene expression are typically constructed through differential equations. mRNA’s production is taken to be proportional to transcription factor activity (with the proportionality given by the sensitivity) and the mRNA is assumed to decay at a particular rate. The assumption that coregulated genes have similar profiles is equivalent to assuming both the decay and the sensitivity are high. In this talk we advocate model based target identification. We develop a simple probabilistic models of transcription (and translation) which encode mRNA (or Transcription Factor) production and decay. Our models are simple enough to allow genome wide target identification, but rich enough to encode dynamical behavior that, allowing us to identify putative targets even when decay rates are low.Wed, 28 Oct 2009 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-napoli09/model-based-target-identification-from-gene-expression-with-span-g-span-aussian-process.html
http://inverseprobability.com/talks/lawrence-napoli09/model-based-target-identification-from-gene-expression-with-span-g-span-aussian-process.htmlLawrence-napoli09Latent Force ModelsPhysics based approaches to data modeling involve constructing an accurate mechanistic model of data, often based on differential equations. Machine learning approaches are typically data driven— perhaps through regularized function approximation.\
\
These two approaches to data modeling are often seen as polar opposites, but in reality they are two different ends to a spectrum of approaches we might take.\
\
In this talk we introduce latent force models. Latent force models are a new approach to data representation that model data through unknown forcing functions that drive differential equation models. By treating the unknown forcing functions with Gaussian process priors we can create probabilistic models that exhibit particular physical characteristics of interest, for example, in dynamical systems resonance and inertia. This allows us to perform a synthesis of the data driven and physical modeling paradigms. We will show applications of these models in systems biology and modelling of human motion capture data.Fri, 23 Oct 2009 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-nyu09/latent-force-models.html
http://inverseprobability.com/talks/lawrence-nyu09/latent-force-models.htmlLawrence-nyu09Latent Force ModelsPhysics based approaches to data modeling involve constructing an accurate mechanistic model of data, often based on differential equations. Machine learning approaches are typically data driven— perhaps through regularized function approximation.\
\
These two approaches to data modeling are often seen as polar opposites, but in reality they are two different ends to a spectrum of approaches we might take.\
\
In this talk we introduce latent force models. Latent force models are a new approach to data representation that model data through unknown forcing functions that drive differential equation models. By treating the unknown forcing functions with Gaussian process priors we can create probabilistic models that exhibit particular physical characteristics of interest, for example, in dynamical systems resonance and inertia. This allows us to perform a synthesis of the data driven and physical modeling paradigms. We will show applications of these models in systems biology and modelling of human motion capture data.Wed, 21 Oct 2009 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-google09/latent-force-models.html
http://inverseprobability.com/talks/lawrence-google09/latent-force-models.htmlLawrence-google09Model Based Target Identification from Gene Expression with <span>G</span>aussian ProcessesA simple approach to target identification through gene expression studies has been to cluster the expression profiles and look for coregulated genes within clusters. Within systems biology mechanistic models of gene expression are typically constructed through differential equations. mRNA’s production is taken to be proportional to transcription factor activity (with the proportionality given by the sensitivity) and the mRNA is assumed to decay at a particular rate. The assumption that coregulated genes have similar profiles is equivalent to assuming both the decay and the sensitivity are high. In this talk we advocate model based target identification. We develop a simple probabilistic models of transcription (and translation) which encode mRNA (or Transcription Factor) production and decay. Our models are simple enough to allow genome wide target identification, but rich enough to encode dynamical behavior that, allowing us to identify putative targets even when decay rates are low.Mon, 19 Oct 2009 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-jhu09/model-based-target-identification-from-gene-expression-with-span-g-span-aussian-process.html
http://inverseprobability.com/talks/lawrence-jhu09/model-based-target-identification-from-gene-expression-with-span-g-span-aussian-process.htmlLawrence-jhu09Latent Force Modelling with <span>G</span>aussian ProcessesPhysics based approaches to data modeling involve constructing an accurate mechanistic model of data, often based on differential equations. Machine learning typically focuses on data driven approaches—perhaps through regularized function approximations.\
\
These two approaches to data modeling are often seen as polar opposites, but in reality they are two different ends to a spectrum of approaches we might take.\
\
In this talk we introduce latent force models. Latent force models are a new approach to data representation that model data through unknown forcing functions that drive differential equation models. By treating the unknown forcing functions with Gaussian process priors we can create probabilistic models that exhibit particular physical characteristics of interest, for example, in dynamical systems resonance and inertia. This allows us to perform a synthesis of the data driven and physical modeling paradigms. We will show applications of these models in systems biology and modelling of human motion capture data.Fri, 09 Oct 2009 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-newcastle09/latent-force-modelling-with-span-g-span-aussian-processes.html
http://inverseprobability.com/talks/lawrence-newcastle09/latent-force-modelling-with-span-g-span-aussian-processes.htmlLawrence-newcastle09Latent Force Models with <span>G</span>aussian ProcessesThu, 24 Sep 2009 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-inspire09/latent-force-models-with-span-g-span-aussian-processes.html
http://inverseprobability.com/talks/lawrence-inspire09/latent-force-models-with-span-g-span-aussian-processes.htmlLawrence-inspire09Efficient Multiple Output Convolution Processes for Multiple Task LearningWed, 16 Sep 2009 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-warwick09/efficient-multiple-output-convolution-processes-for-multiple-task-learning.html
http://inverseprobability.com/talks/lawrence-warwick09/efficient-multiple-output-convolution-processes-for-multiple-task-learning.htmlLawrence-warwick09Dealing with High Dimensional Data with Dimensionality ReductionSun, 06 Sep 2009 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-interspeech09/dealing-with-high-dimensional-data-with-dimensionality-reduction.html
http://inverseprobability.com/talks/lawrence-interspeech09/dealing-with-high-dimensional-data-with-dimensionality-reduction.htmlLawrence-interspeech09Latent Force Models and Multiple Output <span>G</span>aussian ProcessesWe are used to dealing with the situation where we have a latent variable. Often we assume this latent variable to be independently drawn from a distribution, e.g. probabilistic PCA or factor analysis. This simplification is often extended for temporal data where tractable Markovian independence assumptions are used (e.g. Kalman filters or hidden Markov models). In this talk we will consider the more general case where the latent variable is a forcing function in a differential equation model. We will show how for some simple ordinary differential equations the latent variable can be dealt with analytically for particular Gaussian process priors over the latent force. In this talk we will introduce the general framework and present results in systems biology and motion capture.Thu, 23 Jul 2009 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-lfm_slim09/latent-force-models-and-multiple-output-span-g-span-aussian-processes.html
http://inverseprobability.com/talks/lawrence-lfm_slim09/latent-force-models-and-multiple-output-span-g-span-aussian-processes.htmlLawrence-lfm_slim09Latent Force Models with <span>G</span>aussian ProcessesWe are used to dealing with the situation where we have a latent variable. Often we assume this latent variable to be independently drawn from a distribution, e.g. probabilistic PCA or factor analysis. This simplification is often extended for temporal data where tractable Markovian independence assumptions are used (e.g. Kalman filters or hidden Markov models). In this talk we will consider the more general case where the latent variable is a forcing function in a differential equation model. We will firstly give a brief introduction to Gaussian processes, then we will show how for some simple ordinary differential equations the latent variable can be dealt with analytically for particular Gaussian process priors over the latent force. In this talk we will introduce the general framework, present results in systems biology and motion capture.Mon, 13 Jul 2009 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-lfm_cagliary09/latent-force-models-with-span-g-span-aussian-processes.html
http://inverseprobability.com/talks/lawrence-lfm_cagliary09/latent-force-models-with-span-g-span-aussian-processes.htmlLawrence-lfm_cagliary09Non-linear Matrix Facorization with <span>G</span>aussian ProcesesA popular approach to collaborative filtering is matrix factorization. In this talk we consider the “probabilistic matrix factorization” and by taking a latent variable model perspective we show its equivalence to Bayesian PCA. This inspires us to consider probabilistic PCA and its non-linear extension, the Gaussian process latent variable model (GP-LVM) as an approach for probabilistic non-linear matrix factorization. We apply out approach to benchmark movie recommender data sets. The results show better than previous state-of-the-art performance.Fri, 03 Jul 2009 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-emmds09/non-linear-matrix-facorization-with-span-g-span-aussian-proceses.html
http://inverseprobability.com/talks/lawrence-emmds09/non-linear-matrix-facorization-with-span-g-span-aussian-proceses.htmlLawrence-emmds09An Introduction to Systems Biology from a Machine Learning Perspective <span>II</span>Tue, 23 Jun 2009 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-tutii09/an-introduction-to-systems-biology-from-a-machine-learning-perspective-span-ii-span.html
http://inverseprobability.com/talks/lawrence-tutii09/an-introduction-to-systems-biology-from-a-machine-learning-perspective-span-ii-span.htmlLawrence-tutII09An Introduction to Systems Biology from a Machine Learning PerspectiveMon, 22 Jun 2009 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-tut09/an-introduction-to-systems-biology-from-a-machine-learning-perspective.html
http://inverseprobability.com/talks/lawrence-tut09/an-introduction-to-systems-biology-from-a-machine-learning-perspective.htmlLawrence-tut09Non-linear Matrix Factorization with <span>G</span>aussian ProcessesTue, 14 Apr 2009 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-learning09/non-linear-matrix-factorization-with-span-g-span-aussian-processes.html
http://inverseprobability.com/talks/lawrence-learning09/non-linear-matrix-factorization-with-span-g-span-aussian-processes.htmlLawrence-learning09Estimation of Multiple Transcription Factor Activities using ODEs and <span>G</span>aussian ProcessesWed, 01 Apr 2009 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-licsb09/estimation-of-multiple-transcription-factor-activities-using-odes-and-span-g-span-aussi.html
http://inverseprobability.com/talks/lawrence-licsb09/estimation-of-multiple-transcription-factor-activities-using-odes-and-span-g-span-aussi.htmlLawrence-licsb09Python in Machine LearningIs Python a viable replacement for MATLAB?\
\
We are incredibly reliant on MATLAB, but should we be looking elsewhere for our ML programming needs? In this ML lunch I will try and share my recent experiences with Python and machine learning: good and bad. The main questions I think we should be considering are:\
\
Should we be trying to move to Python for our research?\
\
Should we be using Python in our teaching?\
\
I don’t know the answer, but I’ll try and use this MLO lunch to start the debate!Wed, 25 Mar 2009 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-python09/python-in-machine-learning.html
http://inverseprobability.com/talks/lawrence-python09/python-in-machine-learning.htmlLawrence-python09<span>GP-LVM</span> for Data ConsolidationSat, 20 Dec 2008 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-gpdc08/span-gp-lvm-span-for-data-consolidation.html
http://inverseprobability.com/talks/lawrence-gpdc08/span-gp-lvm-span-for-data-consolidation.htmlLawrence-gpdc08Latent Force Models with <span>G</span>aussian ProcessesWe are used to dealing with the situation where we have a latent variable. Often we assume this latent variable to be independently drawn from a distribution, e.g. probabilistic PCA or factor analysis. This simplification is often extended for temporal data where tractable Markovian independence assumptions are used (e.g. Kalman filters or hidden Markov models). In this talk we will consider the more general case where the latent variable is a forcing function in a differential equation model. We will firstly give a brief introduction to Gaussian processes, then we will show how for some simple ordinary differential equations the latent variable can be dealt with analytically for particular Gaussian process priors over the latent force. In this talk we will introduce the general framework, present results in systems biology.\
\
Joint work with Magnus Rattray, Mauricio Álvarez, Pei Gao, Antti Honkela, David Luengo, Guido Sanguinetti and Michalis K. Titsias.Thu, 16 Oct 2008 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-bristol08/latent-force-models-with-span-g-span-aussian-processes.html
http://inverseprobability.com/talks/lawrence-bristol08/latent-force-models-with-span-g-span-aussian-processes.htmlLawrence-bristol08Inference in Ordinary Differential Equations with Latent Functions through <span>G</span>aussian ProcessesIn biochemical interaction networks is a key problem in estimation of the structure and parameters of the genetic, metabolic and protein interaction networks that underpin all biological processes. We present a framework for Bayesian marginalisation of these latent chemical species through Gaussian process priors. We demonstrate our general approach on three different biological examples of single input motifs, including both activation and repression of transcription. We focus in particular on the problem of inferring transcription factor activity when the concentration of active protein cannot easily be measured. The uncertainty in the inferred transcription factor activity can be integrated out in order to derive a likelihood function that can be used for the estimation of regulatory model parameters.Wed, 08 Oct 2008 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-rss08/inference-in-ordinary-differential-equations-with-latent-functions-through-span-g-span.html
http://inverseprobability.com/talks/lawrence-rss08/inference-in-ordinary-differential-equations-with-latent-functions-through-span-g-span.htmlLawrence-rss08Dynamics with <span>G</span>aussian ProcessesWe are used to dealing with the situation where we have a latent variable. Often we assume this latent variable to be independently drawn from a distribution, *e.g.* probabilistic PCA or factor analysis. This simplification is often extended for temporal data where tractable Markovian independence assumptions are used (*e.g.* Kalman filters or hidden Markov models). In this talk we will consider the more general case where the latent variable is a forcing function in a differential equation model. We will firstly give a brief introduction to Gaussian processes, then we will show how for some simple ordinary differential equations the latent variable can be dealt with analytically for particular Gaussian process priors over the latent force. In this talk we will introduce the general framework, present results in systems biology.\
\
Joint work with Magnus Rattray, Mauricio Álvarez, Pei Gao, Antti Honkela, David Luengo, Guido Sanguinetti and Michalis K. Titsias.Wed, 10 Sep 2008 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-ncaf08/dynamics-with-span-g-span-aussian-processes.html
http://inverseprobability.com/talks/lawrence-ncaf08/dynamics-with-span-g-span-aussian-processes.htmlLawrence-ncaf08Ambiguity Modelling in Latent SpacesWe are interested in the situation where we have two or more representations of an underlying phenomenon. In particular we are interested in the scenario where the representation are complementary. This implies that a single individual representation is not sufficient to fully discriminate a specific instance of the underlying phenomenon, it also means that each representation is an ambiguous representation of the other complementary spaces. In this paper we present a latent variable model capable of consolidating multiple complementary representations. Our method extends canonical correlation analysis by introducing additional latent spaces that are specific to the different representations, thereby explaining the full variance of the observations. These additional spaces, explaining representation specific variance, separately model the variance in a representation ambiguous to the other. We develop a spectral algorithm for fast computation of the embeddings and a probabilistic model (based on Gaussian processes) for validation and inference. The proposed model has several potential application areas, we demonstrate its use for multi-modal regression on a benchmark human pose estimation data set.Mon, 08 Sep 2008 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-mlmi08/ambiguity-modelling-in-latent-spaces.html
http://inverseprobability.com/talks/lawrence-mlmi08/ambiguity-modelling-in-latent-spaces.htmlLawrence-mlmi08Latent Force Models with <span>G</span>aussian ProcessesWe are used to dealing with the situation where we have a latent variable. Often we assume this latent variable to be independently drawn from a distribution, *e.g.* probabilistic PCA or factor analysis. This simplification is often extended for temporal data where tractable Markovian independence assumptions are used (*e.g.* Kalman filters or hidden Markov models).\
\
In this talk we will consider the more general case where the latent variable is a forcing function in a differential equation model. We will show how for some simple ordinary differential equations the latent variable can be dealt with analytically for particular Gaussian process priors over the latent force. In this talk we will introduce the general framework, present results in systems biology and preview extensions.\
\
Joint work with Magnus Rattray, Mauricio Álvarez, Pei Gao, Antti Honkela, David Luengo, Guido Sanguinetti and Michalis K. Titsias.Sat, 06 Sep 2008 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-bark08/latent-force-models-with-span-g-span-aussian-processes.html
http://inverseprobability.com/talks/lawrence-bark08/latent-force-models-with-span-g-span-aussian-processes.htmlLawrence-bark08Statistical inference in systems biology through <span>G</span>aussian processes and ordinary differential equationsIn this talk we will summarise recent work from our group in Manchester on inferring ‘latent biochemical species’ in biological systems using Gaussian processes and differential equations. A key problem in biological data is when particular biochemical species of interest are not directly measurable. We will show how the framework of Gaussian processes can be brought to bear on the problem and values of latent chemical species can be inferred given data and a differential equation model.Tue, 17 Jun 2008 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-warwick08/statistical-inference-in-systems-biology-through-span-g-span-aussian-processes-and-ordi.html
http://inverseprobability.com/talks/lawrence-warwick08/statistical-inference-in-systems-biology-through-span-g-span-aussian-processes-and-ordi.htmlLawrence-warwick08Statistical Inference in Systems Biology through <span>G</span>aussian Processes and Ordinary Differential EquationsIn this talk we will summarise recent work from our group in Manchester on inferring ‘latent biochemical species’ in biological systems using Gaussian processes and differential equations. A key problem in biological data is when particular biochemical species of interest are not directly measurable. We will show how the framework of Gaussian processes can be brought to bear on the problem and values of latent chemical species can be inferred given data and a differential equation model.Wed, 07 May 2008 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-sysbiointrob08/statistical-inference-in-systems-biology-through-span-g-span-aussian-processes-and-ordi.html
http://inverseprobability.com/talks/lawrence-sysbiointrob08/statistical-inference-in-systems-biology-through-span-g-span-aussian-processes-and-ordi.htmlLawrence-sysbioIntroB08An Introduction to Systems Biology from a Machine Learning PerspectiveIn this talk we will introduce some of the challenges in systems biology and discuss the efforts being made to address them using statistical inference. General biological background will be interlaced with case studies that illustrate the salient issues in systems biology from the perspective of a machine learning researcher.Mon, 05 May 2008 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-sysbiointroa08/an-introduction-to-systems-biology-from-a-machine-learning-perspective.html
http://inverseprobability.com/talks/lawrence-sysbiointroa08/an-introduction-to-systems-biology-from-a-machine-learning-perspective.htmlLawrence-sysbioIntroA08Inferring Latent Functions with <span>G</span>aussian Processes in Differential EquationsIn this talk we will present recent work from Manchester in inference of latent functions in differential equations. Simple computational models for systems biology make use of ordinary differential equations that are driven from an often unobserved input function. We will describe how probabilistic inference over these latent functions may be performed through Gaussian process prior distributions. We will describe the algorithms and show results on toy problems and real biological systems.Wed, 30 Apr 2008 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-latentfunc08/inferring-latent-functions-with-span-g-span-aussian-processes-in-differential-equations.html
http://inverseprobability.com/talks/lawrence-latentfunc08/inferring-latent-functions-with-span-g-span-aussian-processes-in-differential-equations.htmlLawrence-latentFunc08Learning and Inference with <span>G</span>aussian Processes: An Overview of <span>B</span>ayesian Inference and <span>G</span>aussian ProcessesTue, 01 Apr 2008 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-gpbayes08/learning-and-inference-with-span-g-span-aussian-processes-an-overview-of-span-b-span-ay.html
http://inverseprobability.com/talks/lawrence-gpbayes08/learning-and-inference-with-span-g-span-aussian-processes-an-overview-of-span-b-span-ay.htmlLawrence-gpbayes08Human Motion Modelling with <span>G</span>aussian ProcessesThu, 07 Feb 2008 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-newton08/human-motion-modelling-with-span-g-span-aussian-processes.html
http://inverseprobability.com/talks/lawrence-newton08/human-motion-modelling-with-span-g-span-aussian-processes.htmlLawrence-newton08Human Motion Modelling through Dimensional Reduction with <span>G</span>aussian ProcessesTue, 29 Jan 2008 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-human08/human-motion-modelling-through-dimensional-reduction-with-span-g-span-aussian-processes.html
http://inverseprobability.com/talks/lawrence-human08/human-motion-modelling-through-dimensional-reduction-with-span-g-span-aussian-processes.htmlLawrence-human08<span>TP1</span>: Leveraging Complex Prior Knowledge in LearningMon, 28 Jan 2008 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-thematic08/span-tp1-span-leveraging-complex-prior-knowledge-in-learning.html
http://inverseprobability.com/talks/lawrence-thematic08/span-tp1-span-leveraging-complex-prior-knowledge-in-learning.htmlLawrence-thematic08Dimensionality ReductionWe approach dimensionality reduction from the perspective of multidimensional scaling. Starting from the basics, we draw the relationship between multidimensional scaling and principal component analysis. From this background we briefly review kernel PCA and Isomap. Finally, we consider the problem of model selection using Gaussian processes.Thu, 24 Jan 2008 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-data08/dimensionality-reduction.html
http://inverseprobability.com/talks/lawrence-data08/dimensionality-reduction.htmlLawrence-data08Exploiting Dimensional Dreduction In Modelling Of High Dimensional DistributionsSat, 08 Dec 2007 00:00:00 +0000
http://inverseprobability.com/talks/exploiting-dimensional-dreduction-in-modelling-of-high-dimensional-distributions.html
http://inverseprobability.com/talks/exploiting-dimensional-dreduction-in-modelling-of-high-dimensional-distributions.htmlLatent Variables, Differential Equations and <span>G</span>aussian ProcessesWe are used to dealing with the situation where we have a latent variable. Often we assume this latent variable to be independently drawn from a distribution, e.g. probabilistic PCA or factor analysis. This simplification is often extended for temporal data where tractable Markovian independence assumptions are used (e.g. Kalman filters or hidden Markov models). In this talk we will consider such models in the context of a biological problem: inferring transcription factor activities in simple transcription networks. We will extend the simpler formalisms described above to consider the case where the latent variable is a ’latent function’ and the relationship with the observed data is described by a linear differential equation. Through the use of a Gaussian process prior over the latent function we can perform inference tractably and learn parameters of interest in the system.Mon, 12 Nov 2007 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-msr07/latent-variables-differential-equations-and-span-g-span-aussian-processes.html
http://inverseprobability.com/talks/lawrence-msr07/latent-variables-differential-equations-and-span-g-span-aussian-processes.htmlLawrence-msr07Modelling Transcriptional Regulation with <span>G</span>aussian ProcessesWed, 07 Nov 2007 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-param07/modelling-transcriptional-regulation-with-span-g-span-aussian-processes.html
http://inverseprobability.com/talks/lawrence-param07/modelling-transcriptional-regulation-with-span-g-span-aussian-processes.htmlLawrence-param07Towards Computational Systems Biology with a Statistical Analysis Pipeline for Microarray DataSince the human genome project began mathematical models have become an integral part of biological data analysis. The growth in data availability has necessitated their use in summarization of the data (e.g. *statistical* approaches such as hierarchical clustering). Simultaneously, as more has become understood about the mechanisms underpinning particular pathways *mechanistic* models of interactions have become more widespread.\
\
The data-driven statistical approach and the mechanistic model approach each have their advantages. Data-driven models can be used in genome wide analyses to ’fish’ for genes that were not known to be relevant but provide a critical role in a pathway. Mechanistic models make real predictions about how systems will respond given particular interventions. The two approaches have interacted only loosely, often not through interaction between the ‘mathematicians’ but through indirect interaction via the biologists.\
\
In this talk we will follow describe a statistical analysis ‘pipeline’ for microarray data which handles the noise in the data. As we proceed down the pipeline we will come closer to mechanistic models of systems. We will finish with some general thoughts about the contribution that a combined statistical/mechanistic modelling approach can make.Wed, 31 Oct 2007 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-mbb07/towards-computational-systems-biology-with-a-statistical-analysis-pipeline-for-microarr.html
http://inverseprobability.com/talks/lawrence-mbb07/towards-computational-systems-biology-with-a-statistical-analysis-pipeline-for-microarr.htmlLawrence-mbb07Latent Variable Modelling with <span>G</span>aussian ProcessesIn this talk we will briefly describe the Gaussian process latent variable model, an approach to probabilistic modelling of data through non-linear dimensional reduction. The model takes a dual approach to statistical inference and can be shown to generalise PCA. We will briefly introduce the model and quickly show some example applications.Thu, 13 Sep 2007 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-inverse07/latent-variable-modelling-with-gaussian-processes.html
http://inverseprobability.com/talks/lawrence-inverse07/latent-variable-modelling-with-gaussian-processes.htmlLawrence-inverse07Probabilistic Inference for Modelling of Transcription Factor ActivityAccurate modelling of transcriptional processes in the cell requires the knowledge of a number of key biological quantities. In practice many of them are difficult to measure in vivo. For example, it is very hard to measure the active concentration levels of the transcription factor proteins that drive the process.\
\
In this talk we will show how, by making use of structural information about the interaction network (e.g. arising form ChIP-chip data), transcription factor activities can estimated using probabilistic inference. We propose two different probabilistic models: a simple linear model with Kalman filter based dynamics for genome/transcriptome wide studies and a differential equation based Gaussian process model with a more physically realistic parameterisation for smaller interaction networks.Thu, 05 Jul 2007 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-uc3tfa07/probabilistic-inference-for-modelling-of-transcription-factor-activity.html
http://inverseprobability.com/talks/lawrence-uc3tfa07/probabilistic-inference-for-modelling-of-transcription-factor-activity.htmlLawrence-uc3tfa07Probabilistic Dimensional Reduction with the <span>G</span>aussian Process Latent Variable ModelDensity modelling in high dimensions is a very difficult problem. Traditional approaches, such as mixtures of Gaussians, typically fail to capture the structure of data sets in high dimensional spaces. In this talk we will argue that for many data sets of interest, the data can be represented as a lower dimensional manifold immersed in the higher dimensional space. We will then present the Gaussian Process Latent Variable Model (GP-LVM), a non-linear probabilistic variant of principal component analysis (PCA) which implicitly assumes that the data lies on a lower dimensional space. Having introduced the GP-LVM we will review extensions to the algorithm, including dynamics, learning of large data sets and back constraints. We will demonstrate the application of the model and its extensions to a range of data sets, including human motion data, a vowel data set and a robot mapping problem.Wed, 04 Jul 2007 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-uc3mgplvm07/probabilistic-dimensional-reduction-with-the-span-g-span-aussian-process-latent-variabl.html
http://inverseprobability.com/talks/lawrence-uc3mgplvm07/probabilistic-dimensional-reduction-with-the-span-g-span-aussian-process-latent-variabl.htmlLawrence-uc3mgplvm07Fast Sparse Gaussian Process Methods: The Informative Vector MachineGaussian processes are a non parametric approach to learning regression models. In this talk we will given a brief review of the use of Gaussian processes for regression. We will then introduce the informative vector machine approach to learning Gaussian processes for Classification on large scale data sets. We will show extensions of the method including multi-task learning, semi-supervised learning and learning invariances.Tue, 03 Jul 2007 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-uc3mivm07/fast-sparse-gaussian-process-methods-the-informative-vector-machine.html
http://inverseprobability.com/talks/lawrence-uc3mivm07/fast-sparse-gaussian-process-methods-the-informative-vector-machine.htmlLawrence-uc3mivm07Hierarchical <span>G</span>aussian Process Latent Variable ModelsThe Gaussian process latent variable model (GP-LVM) is a powerful approach for probabilistic modelling of high dimensional data through dimensional reduction. In this paper we extend the GP-LVM through hierarchies. A hierarchical model (such as a tree) allows us to express conditional independencies in the data as well as the manifold structure. We first introduce Gaussian process hierarchies through a simple dynamical model, we then extend the approach to a more complex hierarchy which is applied to the visualisation of human motion data sets.Fri, 22 Jun 2007 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-icml07/hierarchical-span-g-span-aussian-process-latent-variable-models.html
http://inverseprobability.com/talks/lawrence-icml07/hierarchical-span-g-span-aussian-process-latent-variable-models.htmlLawrence-icml07Probabilistic Inference for Modelling of Transcription Factor ActivityAccurate modelling of transcriptional processes in the cell requires the knowledge of a number of key biological quantities. In practice many of them are difficult to measure in vivo. For example, it is very hard to measure the active concentration levels of the transcription factor proteins that drive the process.\
\
In this talk we will show how, by making use of structural information about the interaction network (e.g. arising form ChIP-chip data), transcription factor activities can estimated using probabilistic inference. We propose two different probabilistic models: a simple linear model with Kalman filter based dynamics for genome/transcriptome wide studies and a differential equation based Gaussian process model with a more physically realistic parameterisation for smaller interaction networks.Wed, 13 Jun 2007 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-gatsby07/probabilistic-inference-for-modelling-of-transcription-factor-activity.html
http://inverseprobability.com/talks/lawrence-gatsby07/probabilistic-inference-for-modelling-of-transcription-factor-activity.htmlLawrence-gatsby07<span>G</span>aussian Processes for Inference in Biological Interaction NetworksIn many biological applications key functions of interest, such as chemical species concentrations, are unobserved. In this talk we will briefly introduce Gaussian processes, which are probabilistic models of functions. We will show how they can be used, in combination with a simple differential equation model, to estimate the concentration of a transcription factor in a simple single input module network motif.Wed, 04 Apr 2007 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-mathbio07/span-g-span-aussian-processes-for-inference-in-biological-interaction-networks.html
http://inverseprobability.com/talks/lawrence-mathbio07/span-g-span-aussian-processes-for-inference-in-biological-interaction-networks.htmlLawrence-mathbio07Modelling Transcriptional Regulation with <span>G</span>aussian ProcessesModelling the dynamics of transcriptional processes in the cell requires the knowledge of a number of key biological quantities. While some of them are relatively easy to measure, such as mRNA decay rates and mRNA abundance levels, it is still very hard to measure the active concentration levels of the transcription factor proteins that drive the process and the sensitivity of target genes to these concentrations. In this paper we show how these quantities for a given transcription factor can be inferred from gene expression levels of a set of known target genes. We treat the protein concentration as a latent function with a Gaussian process prior, and include the sensitivities, mRNA decay rates and baseline expression levels as hyperparameters. We apply this procedure to a human leukemia dataset, focusing on the tumour repressor p53 and obtaining results in good accordance with recent biological studies.Wed, 28 Mar 2007 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-pesb07/modelling-transcriptional-regulation-with-span-g-span-aussian-processes.html
http://inverseprobability.com/talks/lawrence-pesb07/modelling-transcriptional-regulation-with-span-g-span-aussian-processes.htmlLawrence-pesb07Probabilistic Dimensional Reduction with the <span>G</span>aussian Process Latent Variable ModelDensity modelling in high dimensions is a very difficult problem. Traditional approaches, such as mixtures of Gaussians, typically fail to capture the structure of data sets in high dimensional spaces. In this talk we will argue that for many data sets of interest, the data can be represented as a lower dimensional manifold immersed in the higher dimensional space. We will then present the Gaussian Process Latent Variable Model (GP-LVM), a non-linear probabilistic variant of principal component analysis (PCA) which implicitly assumes that the data lies on a lower dimensional space. Having introduced the GP-LVM we will review extensions to the algorithm, including dynamics, learning of large data sets and back constraints. We will demonstrate the application of the model and its extensions to a range of data sets, including human motion data, a vowel data set and a robot mapping problem.Fri, 09 Mar 2007 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-ncrg07/probabilistic-dimensional-reduction-with-the-span-g-span-aussian-process-latent-variabl.html
http://inverseprobability.com/talks/lawrence-ncrg07/probabilistic-dimensional-reduction-with-the-span-g-span-aussian-process-latent-variabl.htmlLawrence-ncrg07Probabilistic Dimensional Reduction with the <span>G</span>aussian Process Latent Variable ModelDensity modelling in high dimensions is a very difficult problem. Traditional approaches, such as mixtures of Gaussians, typically fail to capture the structure of data sets in high dimensional spaces. In this talk we will argue that for many data sets of interest, the data can be represented as a lower dimensional manifold immersed in the higher dimensional space. We will then present the Gaussian Process Latent Variable Model (GP-LVM), a non-linear probabilistic variant of principal component analysis (PCA) which implicitly assumes that the data lies on a lower dimensional space. Having introduced the GP-LVM we will review extensions to the algorithm, including dynamics, learning of large data sets and back constraints. We will demonstrate the application of the model and its extensions to a range of data sets, including human motion data, a vowel data set and a robot mapping problem.Mon, 12 Feb 2007 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-google07/probabilistic-dimensional-reduction-with-the-span-g-span-aussian-process-latent-variabl.html
http://inverseprobability.com/talks/lawrence-google07/probabilistic-dimensional-reduction-with-the-span-g-span-aussian-process-latent-variabl.htmlLawrence-google07Probabilistic Dimensional Reduction with the <span>G</span>aussian Process Latent Variable ModelDensity modelling in high dimensions is a very difficult problem. Traditional approaches, such as mixtures of Gaussians, typically fail to capture the structure of data sets in high dimensional spaces. In this talk we will argue that for many data sets of interest, the data can be represented as a lower dimensional manifold immersed in the higher dimensional space. We will then present the Gaussian Process Latent Variable Model (GP-LVM), a non-linear probabilistic variant of principal component analysis (PCA) which implicitly assumes that the data lies on a lower dimensional space. Having introduced the GP-LVM we will review extensions to the algorithm, including dynamics, learning of large data sets and back constraints. We will demonstrate the application of the model and its extensions to a range of data sets, including human motion data, a vowel data set and a robot mapping problem.Fri, 09 Feb 2007 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-csail07/probabilistic-dimensional-reduction-with-the-span-g-span-aussian-process-latent-variabl.html
http://inverseprobability.com/talks/lawrence-csail07/probabilistic-dimensional-reduction-with-the-span-g-span-aussian-process-latent-variabl.htmlLawrence-csail07Learning and Inference with <span>G</span>aussian Processes: An Overview of <span>G</span>aussian Processes and the <span>GP-LVM</span>Fri, 03 Nov 2006 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-manchesterguest06/learning-and-inference-with-span-g-span-aussian-processes-an-overview-of-span-g-span-au.html
http://inverseprobability.com/talks/lawrence-manchesterguest06/learning-and-inference-with-span-g-span-aussian-processes-an-overview-of-span-g-span-au.htmlLawrence-manchesterGuest06Learning and Inference with <span>G</span>aussian ProcessesMany application domains of machine learning can be reduced to inference about the values of a function. Gaussian processes are powerful, flexible, probabilistic models that enable us to efficiently perform inference about functions in the presence of uncertainty. In this talk I will introduce Gaussian processes and review a few standard applications of these models. I will then show how Gaussian processes can be used to solve important and diverse real-world problems, including inference of the concentration of transcription factors which regulate gene expression and creating probabilistic models of human motion for animation and tracking.Mon, 21 Aug 2006 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-intel06/learning-and-inference-with-span-g-span-aussian-processes.html
http://inverseprobability.com/talks/lawrence-intel06/learning-and-inference-with-span-g-span-aussian-processes.htmlLawrence-intel06PUMA: Propagation of Uncertainty in Microarray AnalysisWed, 02 Aug 2006 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-tuebingen06/puma-propagation-of-uncertainty-in-microarray-analysis.html
http://inverseprobability.com/talks/lawrence-tuebingen06/puma-propagation-of-uncertainty-in-microarray-analysis.htmlLawrence-tuebingen06Probabilistic Dimensional Reduction with the <span>G</span>aussian Process Latent Variable ModelDensity modelling in high dimensions is a very difficult problem. Traditional approaches, such as mixtures of Gaussians, typically fail to capture the structure of data sets in high dimensional spaces. In this talk we will argue that for many data sets of interest, the data can be represented as a lower dimensional manifold immersed in the higher dimensional space. We will then present the Gaussian Process Latent Variable Model (GP-LVM), a non-linear probabilistic variant of principal component analysis (PCA) which implicitly assumes that the data lies on a lower dimensional space. Having introduced the GP-LVM we will review extensions to the algorithm, including dynamics, learning of large data sets and back constraints. We will demonstrate the application of the model and its extensions to a range of data sets, including human motion data, a vowel data set and a robot mapping problem.Tue, 11 Jul 2006 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-erice06/probabilistic-dimensional-reduction-with-the-span-g-span-aussian-process-latent-variabl.html
http://inverseprobability.com/talks/lawrence-erice06/probabilistic-dimensional-reduction-with-the-span-g-span-aussian-process-latent-variabl.htmlLawrence-erice06Local Distance Preservation in the GP-LVM through Back ConstraintsTue, 27 Jun 2006 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-icml06/local-distance-preservation-in-the-gp-lvm-through-back-constraints.html
http://inverseprobability.com/talks/lawrence-icml06/local-distance-preservation-in-the-gp-lvm-through-back-constraints.htmlLawrence-icml06Learning and Inference with <span>G</span>aussian ProcessesMany application domains of machine learning can be reduced to inference about the values of a function. Gaussian processes are powerful, flexible, probabilistic models that enable us to efficiently perform inference about functions in the presence of uncertainty. In this talk I will introduce Gaussian processes and review a few standard applications of these models. I will then show how Gaussian processes can be used to solve important and diverse real-world problems, including inference of the concentration of transcription factors which regulate gene expression and creating probabilistic models of human motion for animation and tracking.Thu, 22 Jun 2006 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-manchester06/learning-and-inference-with-span-g-span-aussian-processes.html
http://inverseprobability.com/talks/lawrence-manchester06/learning-and-inference-with-span-g-span-aussian-processes.htmlLawrence-manchester06A probabilistic dynamical model for quantitative inference of the regulatory mechanism of transcriptionMon, 10 Apr 2006 00:00:00 +0000
http://inverseprobability.com/talks/sanguinetti-masamb06/a-probabilistic-dynamical-model-for-quantitative-inference-of-the-regulatory-mechanism.html
http://inverseprobability.com/talks/sanguinetti-masamb06/a-probabilistic-dynamical-model-for-quantitative-inference-of-the-regulatory-mechanism.htmlSanguinetti-masamb06Probabilistic Dimensional Reduction with the <span>G</span>aussian Process Latent Variable ModelDensity modelling in high dimensions is a very difficult problem. Traditional approaches, such as mixtures of Gaussians, typically fail to capture the structure of data sets in high dimensional spaces. In this talk we will argue that for many data sets of interest, the data can be represented as a lower dimensional manifold immersed in the higher dimensional space. We will then present the Gaussian Process Latent Variable Model (GP-LVM), a non-linear probabilistic variant of principal component analysis (PCA) which implicitly assumes that the data lies on a lower dimensional space. Having introduced the GP-LVM we will review extensions to the algorithm, including dynamics, learning of large data sets and back constraints. We will demonstrate the application of the model and its extensions to a range of data sets, including human motion data, a vowel data set and a robot mapping problem.Tue, 07 Mar 2006 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-cued06/probabilistic-dimensional-reduction-with-the-span-g-span-aussian-process-latent-variabl.html
http://inverseprobability.com/talks/lawrence-cued06/probabilistic-dimensional-reduction-with-the-span-g-span-aussian-process-latent-variabl.htmlLawrence-cued06Computer Vision Reading Group: The <span>G</span>aussian Process Latent Variable ModelThe Gaussian process latent variable model (GP-LVM) is a recently proposed probabilistic approach to obtaining a reduced dimension representation of a data set. In this tutorial we motivate and describe the GP-LVM, giving a review of the model itself and some of the concepts behind it.Fri, 27 Jan 2006 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-oxford06/computer-vision-reading-group-the-span-g-span-aussian-process-latent-variable-model.html
http://inverseprobability.com/talks/lawrence-oxford06/computer-vision-reading-group-the-span-g-span-aussian-process-latent-variable-model.htmlLawrence-oxford06High Dimensional Probabilistic Modelling through ManifoldsDensity modelling in high dimensions is a very difficult problem. Traditional approaches, such as mixtures of Gaussians, typically fail to capture the structure of data sets in high dimensional spaces. In this talk we will argue that for many data sets of interest, the data can be represented as a lower dimensional manifold immersed in the higher dimensional space. We will then present the Gaussian Process Latent Variable Model (GP-LVM), a non-linear probabilistic variant of principal component analysis (PCA) which implicitly assumes that the data lies on a lower dimensional space. We will demonstrate the application of the model to a range of data sets, but with a particular focus on human motion data. We will show some preliminary work on facial animation and make use of a skeletal motion capture data set to illustrate differences between our model and traditional manifold techniques.Wed, 14 Dec 2005 00:00:00 +0000
http://inverseprobability.com/talks/lawrence--uw05/high-dimensional-probabilistic-modelling-through-manifolds.html
http://inverseprobability.com/talks/lawrence--uw05/high-dimensional-probabilistic-modelling-through-manifolds.htmlLawrence--uw05High Dimensional Probabilistic Modelling through ManifoldsDensity modelling in high dimensions is a very difficult problem. Traditional approaches, such as mixtures of Gaussians, typically fail to capture the structure of data sets in high dimensional spaces. In this talk we will argue that for many data sets of interest, the data can be represented as a lower dimensional manifold immersed in the higher dimensional space. We will then present the Gaussian Process Latent Variable Model (GP-LVM), a non-linear probabilistic variant of principal component analysis (PCA) which implicitly assumes that the data lies on a lower dimensional space. We will demonstrate the application of the model to a range of data sets, but with a particular focus on human motion data. We will show some preliminary work on facial animation and make use of a skeletal motion capture data set to illustrate differences between our model and traditional manifold techniques.Mon, 12 Dec 2005 00:00:00 +0000
http://inverseprobability.com/talks/lawrence--msrred05/high-dimensional-probabilistic-modelling-through-manifolds.html
http://inverseprobability.com/talks/lawrence--msrred05/high-dimensional-probabilistic-modelling-through-manifolds.htmlLawrence--msrred05High Dimensional Probabilistic Modelling through ManifoldsDensity modelling in high dimensions is a very difficult problem. Traditional approaches, such as mixtures of Gaussians, typically fail to capture the structure of data sets in high dimensional spaces. In this talk we will argue that for many data sets of interest, the data can be represented as a lower dimensional manifold immersed in the higher dimensional space. We will then present the Gaussian Process Latent Variable Model (GP-LVM), a non-linear probabilistic variant of principal component analysis (PCA) which implicitly assumes that the data lies on a lower dimensional space. We will demonstrate the application of the model to a range of data sets, but with a particular focus on human motion data. We will show some preliminary work on facial animation and make use of a skeletal motion capture data set to illustrate differences between our model and traditional manifold techniques.Fri, 02 Dec 2005 00:00:00 +0000
http://inverseprobability.com/talks/lawrence--ubc05/high-dimensional-probabilistic-modelling-through-manifolds.html
http://inverseprobability.com/talks/lawrence--ubc05/high-dimensional-probabilistic-modelling-through-manifolds.htmlLawrence--ubc05High Dimensional Probabilistic Modelling through ManifoldsDensity modelling in high dimensions is a very difficult problem. Traditional approaches, such as mixtures of Gaussians, typically fail to capture the structure of data sets in high dimensional spaces. In this talk we will argue that for many data sets of interest, the data can be represented as a lower dimensional manifold immersed in the higher dimensional space. We will then present the Gaussian Process Latent Variable Model (GP-LVM), a non-linear probabilistic variant of principal component analysis (PCA) which implicitly assumes that the data lies on a lower dimensional space. We will demonstrate the application of the model to a range of data sets, but with a particular focus on human motion data. We will show some preliminary work on facial animation and make use of a skeletal motion capture data set to illustrate differences between our model and traditional manifold techniques.Tue, 29 Nov 2005 00:00:00 +0000
http://inverseprobability.com/talks/lawrence--columbia05/high-dimensional-probabilistic-modelling-through-manifolds.html
http://inverseprobability.com/talks/lawrence--columbia05/high-dimensional-probabilistic-modelling-through-manifolds.htmlLawrence--columbia05High Dimensional Probabilistic Modelling through ManifoldsDensity modelling in high dimensions is a very difficult problem. Traditional approaches, such as mixtures of Gaussians, typically fail to capture the structure of data sets in high dimensional spaces. In this talk we will argue that for many data sets of interest, the data can be represented as a lower dimensional manifold immersed in the higher dimensional space. We will then present the Gaussian Process Latent Variable Model (GP-LVM), a non-linear probabilistic variant of principal component analysis (PCA) which implicitly assumes that the data lies on a lower dimensional space. We will demonstrate the application of the model to a range of data sets, but with a particular focus on human motion data. We will show some preliminary work on facial animation and make use of a skeletal motion capture data set to illustrate differences between our model and traditional manifold techniques.Mon, 28 Nov 2005 00:00:00 +0000
http://inverseprobability.com/talks/lawrence--ibm05/high-dimensional-probabilistic-modelling-through-manifolds.html
http://inverseprobability.com/talks/lawrence--ibm05/high-dimensional-probabilistic-modelling-through-manifolds.htmlLawrence--ibm05High Dimensional Probabilistic Modelling through ManifoldsDensity modelling in high dimensions is a very difficult problem. Traditional approaches, such as mixtures of Gaussians, typically fail to capture the structure of data sets in high dimensional spaces. In this talk we will argue that for many data sets of interest, the data can be represented as a lower dimensional manifold immersed in the higher dimensional space. We will then present the Gaussian Process Latent Variable Model (GP-LVM), a non-linear probabilistic variant of principal component analysis (PCA) which implicitly assumes that the data lies on a lower dimensional space. We will demonstrate the application of the model to a range of data sets, but with a particular focus on human motion data. We will show some preliminary work on facial animation and make use of a skeletal motion capture data set to illustrate differences between our model and traditional manifold techniques.Wed, 16 Nov 2005 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-gatsby05/high-dimensional-probabilistic-modelling-through-manifolds.html
http://inverseprobability.com/talks/lawrence-gatsby05/high-dimensional-probabilistic-modelling-through-manifolds.htmlLawrence-gatsby05High Dimensional Probabilistic Modelling through ManifoldsDensity modelling in high dimensions is a very difficult problem. Traditional approaches, such as mixtures of Gaussians, typically fail to capture the structure of data sets in high dimensional spaces. In this talk we will argue that for many data sets of interest, the data can be represented as a lower dimensional manifold immersed in the higher dimensional space. We will then present the Gaussian Process Latent Variable Model (GP-LVM), a non-linear probabilistic variant of principal component analysis (PCA) which implicitly assumes that the data lies on a lower dimensional space. We will demonstrate the application of the model to a range of data sets, but with a particular focus on human motion data. We will show some preliminary work on facial animation and make use of a skeletal motion capture data set to illustrate differences between our model and traditional manifold techniques.Wed, 02 Nov 2005 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-idiap05/high-dimensional-probabilistic-modelling-through-manifolds.html
http://inverseprobability.com/talks/lawrence-idiap05/high-dimensional-probabilistic-modelling-through-manifolds.htmlLawrence-idiap05High Dimensional Probabilistic Modelling through ManifoldsDensity modelling in high dimensions is a very difficult problem. Traditional approaches, such as mixtures of Gaussians, typically fail to capture the structure of data sets in high dimensional spaces. In this talk we will argue that for many data sets of interest, the data can be represented as a lower dimensional manifold immersed in the higher dimensional space. We will then present the Gaussian Process Latent Variable Model (GP-LVM), a non-linear probabilistic variant of principal component analysis (PCA) which implicitly assumes that the data lies on a lower dimensional space. We will demonstrate the application of the model to a range of data sets, but with a particular focus on human motion data. We will show some preliminary work on facial animation and make use of a skeletal motion capture data set to illustrate differences between our model and traditional manifold techniques.Mon, 31 Oct 2005 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-epfl05/high-dimensional-probabilistic-modelling-through-manifolds.html
http://inverseprobability.com/talks/lawrence-epfl05/high-dimensional-probabilistic-modelling-through-manifolds.htmlLawrence-epfl05Probabilistic Non-linear Component Analysis through <span>G</span>aussian Process Latent Variable ModelsMon, 15 Aug 2005 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-tuebingen05/probabilistic-non-linear-component-analysis-through-span-g-span-aussian-process-latent.html
http://inverseprobability.com/talks/lawrence-tuebingen05/probabilistic-non-linear-component-analysis-through-span-g-span-aussian-process-latent.htmlLawrence-tuebingen05Probabilistic Non-linear Component Analysis through <span>G</span>aussian Process Latent Variable ModelsIt is known that Principal Component Analysis has an underlying probabilistic representation based on a latent variable model. Principal component analysis (PCA) is recovered when the latent variables are integrated out and the parameters of the model are optimised by maximum likelihood. It is less well known that the dual approach of integrating out the parameters and optimising with respect to the latent variables also leads to PCA. The marginalised likelihood in this case takes the form of Gaussian process mappings, with linear Covariance functions, from a latent space to an observed space, which we refer to as a Gaussian Process Latent Variable Model (GPLVM). This dual probabilistic PCA is still a linear latent variable model, but by looking beyond the inner product kernel as a covariance function we can develop a non-linear probabilistic PCA. In the talk we will introduce the GPLVM and illustrate its application on a range of high dimensional data sets including motion capture data, hand written digits, a medical diagnosis data set and images.Wed, 11 May 2005 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-soton05/probabilistic-non-linear-component-analysis-through-span-g-span-aussian-process-latent.html
http://inverseprobability.com/talks/lawrence-soton05/probabilistic-non-linear-component-analysis-through-span-g-span-aussian-process-latent.htmlLawrence-soton05Probabilistic Non-linear Component Analysis through <span>G</span>aussian Process Latent Variable ModelsIt is known that Principal Component Analysis has an underlying probabilistic representation based on a latent variable model. Principal component analysis (PCA) is recovered when the latent variables are integrated out and the parameters of the model are optimised by maximum likelihood. It is less well known that the dual approach of integrating out the parameters and optimising with respect to the latent variables also leads to PCA. The marginalised likelihood in this case takes the form of Gaussian process mappings, with linear Covariance functions, from a latent space to an observed space, which we refer to as a Gaussian Process Latent Variable Model (GPLVM). This dual probabilistic PCA is still a linear latent variable model, but by looking beyond the inner product kernel as a covariance function we can develop a non-linear probabilistic PCA. In the talk we will introduce the GPLVM and illustrate its application on a range of high dimensional data sets including motion capture data, hand written digits, a medical diagnosis data set and images.Tue, 15 Mar 2005 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-msr05/probabilistic-non-linear-component-analysis-through-span-g-span-aussian-process-latent.html
http://inverseprobability.com/talks/lawrence-msr05/probabilistic-non-linear-component-analysis-through-span-g-span-aussian-process-latent.htmlLawrence-msr05Probabilistic Non-linear Component Analysis through <span>G</span>aussian Process Latent Variable ModelsIt is known that Principal Component Analysis has an underlying probabilistic representation based on a latent variable model. Principal component analysis (PCA) is recovered when the latent variables are integrated out and the parameters of the model are optimised by maximum likelihood. It is less well known that the dual approach of integrating out the parameters and optimising with respect to the latent variables also leads to PCA. The marginalised likelihood in this case takes the form of Gaussian process mappings, with linear Covariance functions, from a latent space to an observed space, which we refer to as a Gaussian Process Latent Variable Model (GPLVM). This dual probabilistic PCA is still a linear latent variable model, but by looking beyond the inner product kernel as a covariance function we can develop a non-linear probabilistic PCA. In the talk we will introduce the GPLVM and illustrate its application on a range of high dimensional data sets including motion capture data, hand written digits, a medical diagnosis data set and images.Wed, 09 Mar 2005 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-manchester05/probabilistic-non-linear-component-analysis-through-span-g-span-aussian-process-latent.html
http://inverseprobability.com/talks/lawrence-manchester05/probabilistic-non-linear-component-analysis-through-span-g-span-aussian-process-latent.htmlLawrence-manchester05Probabilistic Non-linear Component Analysis through <span>G</span>aussian Process Latent Variable ModelsIt is known that Principal Component Analysis has an underlying probabilistic representation based on a latent variable model. Principal component analysis (PCA) is recovered when the latent variables are integrated out and the parameters of the model are optimised by maximum likelihood. It is less well known that the dual approach of integrating out the parameters and optimising with respect to the latent variables also leads to PCA. The marginalised likelihood in this case takes the form of Gaussian process mappings, with linear Covariance functions, from a latent space to an observed space, which we refer to as a Gaussian Process Latent Variable Model (GPLVM). This dual probabilistic PCA is still a linear latent variable model, but by looking beyond the inner product kernel as a covariance function we can develop a non-linear probabilistic PCA. In the talk we will introduce the GPLVM and illustrate its application on a range of high dimensional data sets including motion capture data, hand written digits, a medical diagnosis data set and images.Tue, 01 Mar 2005 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-edinburgh05/probabilistic-non-linear-component-analysis-through-span-g-span-aussian-process-latent.html
http://inverseprobability.com/talks/lawrence-edinburgh05/probabilistic-non-linear-component-analysis-through-span-g-span-aussian-process-latent.htmlLawrence-edinburgh05Probabilistic Non-linear Component Analysis through <span>G</span>aussian Process Latent Variable ModelsIt is known that Principal Component Analysis has an underlying probabilistic representation based on a latent variable model. Principal component analysis (PCA) is recovered when the latent variables are integrated out and the parameters of the model are optimised by maximum likelihood. It is less well known that the dual approach of integrating out the parameters and optimising with respect to the latent variables also leads to PCA. The marginalised likelihood in this case takes the form of Gaussian process mappings, with linear Covariance functions, from a latent space to an observed space, which we refer to as a Gaussian Process Latent Variable Model (GPLVM). This dual probabilistic PCA is still a linear latent variable model, but by looking beyond the inner product kernel as a covariance function we can develop a non-linear probabilistic PCA. In the talk we will introduce the GPLVM and illustrate its application on a range of high dimensional data sets including motion capture data, hand written digits, a medical diagnosis data set and images.Mon, 21 Feb 2005 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-oxford05/probabilistic-non-linear-component-analysis-through-span-g-span-aussian-process-latent.html
http://inverseprobability.com/talks/lawrence-oxford05/probabilistic-non-linear-component-analysis-through-span-g-span-aussian-process-latent.htmlLawrence-oxford05Probabilistic Non-linear Component Analysis through <span>G</span>aussian Process Latent Variable ModelsIt is known that Principal Component Analysis has an underlying probabilistic representation based on a latent variable model. Principal component analysis (PCA) is recovered when the latent variables are integrated out and the parameters of the model are optimised by maximum likelihood. It is less well known that the dual approach of integrating out the parameters and optimising with respect to the latent variables also leads to PCA. The marginalised likelihood in this case takes the form of Gaussian process mappings, with linear Covariance functions, from a latent space to an observed space, which we refer to as a Gaussian Process Latent Variable Model (GPLVM). This dual probabilistic PCA is still a linear latent variable model, but by looking beyond the inner product kernel as a for a covariance function we can develop a non-linear probabilistic PCA.Thu, 09 Sep 2004 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-smlwgplvm03/probabilistic-non-linear-component-analysis-through-span-g-span-aussian-process-latent.html
http://inverseprobability.com/talks/lawrence-smlwgplvm03/probabilistic-non-linear-component-analysis-through-span-g-span-aussian-process-latent.htmlLawrence-smlwgplvm03Probabilistic Non-linear Component Analysis through Gaussian Process Latent Variable ModelsIt is known that Principal Component Analysis has an underlying
probabilistic representation based on a latent variable model. PCA
is recovered when the latent variables are integrated out and the
parameters of the model are optimised by maximum likelihood. It is
less well known that the dual approach of integrating out the
parameters and optimising with respect to the latent variables also
leads to PCA. The marginalised likelihood in this case takes the
form of Gaussian process mappings, with linear Covariance functions,
from a latent space to an observed space, which we refer to as a
Gaussian Process Latent Variable Model (GPLVM) [@Lawrence:gplvm03].
It is straightforward to *non-linearise* this model by
substituting the linear covariance function for a non-linear
one. The result is a non-linear probabilistic PCA model. In this
talk we will present a practical algorithm for optimising the latent
variables in a non-linear GPLVM and discuss some relations with
other models. Finally we will present results from a SIGGRAPH paper
which uses the GPLVM to learn styles in an inverse kinematics
problem [@Grochow:styleik04].
Thu, 06 May 2004 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-ucbgplvm03/probabilistic-non-linear-component-analysis-through-gaussian-process-latent.html
http://inverseprobability.com/talks/lawrence-ucbgplvm03/probabilistic-non-linear-component-analysis-through-gaussian-process-latent.htmlLawrence-ucbgplvm03Bayesian Processing of <span>cDNA</span> Microarray Images through the Variational Importance SamplerEach cell in the human body contains the same basic code in the form of the genome, however cells have differentiated roles which come about through different cells ‘expressing’ different genes. Key insights into gene interactions can be studied through measuring the level of expression of each gene at different times. Gene expression levels can be obtained from cDNA microarray experiments through the extraction of pixel intensities from a scanned image of a slide. In this talk we will start by briefly reviewing cDNA microarray technology. We will then focus on one problem that arises when processing these images: human error in locating the position of the spots can lead to variabilities in the extracted expression levels. We will present a Bayesian approach to the image processing which alleviates this problem. Our approach makes use of a novel combination of importance sampling and variational approximations. Finally if there is time we will briefly show some examples of the variational importance sampler applied to visual tracking problems.Thu, 04 Dec 2003 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-msrb03/bayesian-processing-of-span-cdna-span-microarray-images-through-the-variational-importa.html
http://inverseprobability.com/talks/lawrence-msrb03/bayesian-processing-of-span-cdna-span-microarray-images-through-the-variational-importa.htmlLawrence-msrb03Bayesian Processing of <span>cDNA</span> Microarray ImagesGene expression levels are obtained from microarray experiments through the extraction of pixel intensities from a scanned image of the slide. It is widely acknowledged that variabilities can occur in expression levels extracted from the same images by different users with the same software packages. These inconsistencies arise due to differences in the refinement of the placement of the microarray ’grids’. We introduce a novel automated approach to the refinement of grid placements that is based upon the use of Bayesian inference for determining the size, shape and positioning of the microarray ’spots’, capturing uncertainty that can be passed to downstream analysis. Our experiments demonstrate that variability between users can be significantly reduced using the approach. The automated nature of the approach also saves hours of researchers’ time normally spent in refining the grid placement.Fri, 20 Jun 2003 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-sussex03/bayesian-processing-of-span-cdna-span-microarray-images.html
http://inverseprobability.com/talks/lawrence-sussex03/bayesian-processing-of-span-cdna-span-microarray-images.htmlLawrence-sussex03Bayesian Processing of <span>cDNA</span> Microarray ImagesGene expression levels are obtained from microarray experiments through the extraction of pixel intensities from a scanned image of the slide. It is widely acknowledged that variabilities can occur in expression levels extracted from the same images by different users with the same software packages. These inconsistencies arise due to differences in the refinement of the placement of the microarray ‘grids’. We introduce a novel automated approach to the refinement of grid placements that is based upon the use of Bayesian inference for determining the size, shape and positioning of the microarray ‘spots’, capturing uncertainty that can be passed to downstream analysis. Our experiments demonstrate that variability between users can be significantly reduced using the approach. The automated nature of the approach also saves hours of researchers’ time normally spent in refining the grid placement. A MATLAB implementation of the algorithm is available from <http://inverseprobability.com/vis>.Wed, 21 May 2003 00:00:00 +0000
http://inverseprobability.com/talks/lawrence-manchester03/bayesian-processing-of-span-cdna-span-microarray-images.html
http://inverseprobability.com/talks/lawrence-manchester03/bayesian-processing-of-span-cdna-span-microarray-images.htmlLawrence-manchester03Particle Filters, Variational methods and Importance SamplingParticle filters allow tracking of systems with highly non-linear,
multi-modal posterior distributions, however they are prone to
failure when model likelihoods are sharply peaked or state spaces
are high dimensional. This failure is caused by a mismatch between
the proposal distribution and the true posterior. The number of
particles of samples then required to accurately represent the
posterior increases dramatically and with it the computational
demands of the algorithm. By formulating the problem within the
framework of variational inference we derive an algorithm in which
the proposal naturally adapts to more accurately reflect the true
posterior. This is achieved by replacing intractable moment
evaluations, arising from the highly non-linear nature of the
likelihood functions, with sample based approximations. In this
talk we shall first introduce the approach in a static setting:
Bayesian processing of cDNA microarray images. We will then add
dynamics to the model and demonstrate a marked improvement over
standard approaches on both synthetic and real-world tracking
examples.
Mon, 24 Mar 2003 00:00:00 +0000
http://inverseprobability.com/talks/particle-filters-variational-methods-and-importance-sampling.html
http://inverseprobability.com/talks/particle-filters-variational-methods-and-importance-sampling.html