AI Summer School 2017

AI and Healthcare

Abstract

Over the next decade, AI is set to transform the healthcare industry. Microsoft is working in a number of areas to empower healthcare experts with AI solutions to enable more effective treatment and diagnosis, and try to alleviate some of the current burdens on the healthcare system. Chris will cover some of the AI and healthcare projects underway, both within Microsoft Research and wider Microsoft

Biography

Chris Bishop is a Microsoft Technical Fellow and the Laboratory Director at Microsoft Research Cambridge. He is also Professor of Computer Science at the University of Edinburgh, and a Fellow of Darwin College, Cambridge. In 2004, he was elected Fellow of the Royal Academy of Engineering, in 2007 he was elected Fellow of the Royal Society of Edinburgh, and in 2017 he was elected Fellow of the Royal Society.

Chris obtained a BA in Physics from Oxford, and a PhD in Theoretical Physics from the University of Edinburgh, with a thesis on quantum field theory. He then joined Culham Laboratory where he worked on the theory of magnetically confined plasmas as part of the European controlled fusion programme.

From there, he developed an interest in pattern recognition, and became Head of the Applied Neurocomputing Centre at AEA Technology. He was subsequently elected to a Chair in the Department of Computer Science and Applied Mathematics at Aston University, where he led the Neural Computing Research Group. Chris then took a sabbatical during which time he was principal organiser of the six month international research programme on Neural Networks and Machine Learning at the Isaac Newton Institute for Mathematical Sciences in Cambridge, which ran in 1997.

After completion of the Newton Institute programme Chris joined the Microsoft Research Laboratory in Cambridge.

Timing and Coordination in Physically Situated Language Interaction

View the slides

Abstract

Over the recent years, we have seen significant advances in speech recognition, language and dialog technologies. Voice-enabled personal assistants are now deployed in the hands of millions of people. At the same time, numerous challenges remain largely unaddressed in the realm of physically situated language interaction (e.g., in-car systems, robots in public spaces, ambient assistance). In this talk, I will highlight some of these challenges. I will outline a set of basic skills required for supporting dialog in physically situated settings, and I will highlight the importance of fine-grained timing and coordination in regulating language interaction processes like engagement and turn-taking.

Biography

Dan Bohus is a Senior Researcher in the Adaptive Systems and Interaction Group at Microsoft Research Redmond. His research agenda is focused on physically situated, open-world spoken language interaction. Before joining Microsoft Research, Dan has received his Ph.D. degree (2007) in Computer Science from Carnegie Mellon University.

Assistive AI for Cancer Treatment

Abstract

I will talk about how a decade of research in machine learning and computer vision is culminating in a tool that helps treat cancer in a more efficient, targeted way. Surprisingly, the technology behind it all was inspired by the work we have done on Kinect. A collaboration with the cancer center at Addenbrookes is helping turn our research into a practical tool with enormous potential impact for the oncologists, and ultimately for the patients too.

Biography

In June 2000, Antonio Criminisi joined Microsoft Research in Cambridge (Machine Learning and Perception group) as a visiting researcher. In February 2001, he moved to the Interactive Visual Media Group in Redmond (WA, United States) as a post-doctorate researcher. In October 2002, he moved back to the Machine Learning and Perception Group in Cambridge as a researcher. In September 2011, he became senior researcher and is now leading the Medical Image Analysis team.

Antonio’s current research interests are in the area of medical image analysis, object category recognition, image and video analysis and editing, one-to-one teleconferencing, 3D reconstruction from single and multiple images with application to virtual reality, forensic science, and history of art.

Antonio Criminisi was born in 1972 in Italy. In October 1990, he was appointed “Alfiere del Lavoro” by the Italian President F. Cossiga for his successful studies. In July 1996, he received a degree in Electronics Engineering at the University of Palermo and in December 1999, he obtained a PhD in Computer Vision at the University of Oxford. His thesis. Accurate Visual Metrology from Single and Multiple Uncalibrated Images, won the British Computer Society Distinguished Dissertation Award for the year 2000 and was published by Springer-Verlag London Ltd. in August 2001. Antonio was a Research Fellow at Clare Hall College, Cambridge from 2002 to 2005. Antonio has won a number of best paper prizes in top computer vision conferences.

Session “Design” – On the Human Side

Abstract

This session addresses what is human about artificial intelligence. The session will inspect this question from three very different perspectives.  Alex Taylor will discuss human perspectives on AI; Anab Jain will comment on how design is shaping and commenting on AI; and Rebecca Fiebrink will speak about how people can learn to train AI tools for their own use. A commentary across fields, this session is intended to open up the discussion about where AI fits in the lives of people and the research challenges that that presents.

Biography

Dr. Rebecca Fiebrink is a Lecturer at Goldsmiths, University of London. Her research focuses on designing new ways for humans to interact with computers in creative practice, including the use of machine learning as a creative tool. Fiebrink is the developer of the Wekinator system for real-time interactive machine learning, and the creator of a MOOC titled “Machine Learning for Artists and Musicians,” which launched in 2016 on the Kadenze platform. She was previously an Assistant Professor at Princeton University, where she co-directed the Princeton Laptop Orchestra. She has worked with companies including Microsoft Research, Sun Microsystems Research Labs, Imagine Research, and Smule, where she helped to build the #1 iTunes app “I am T-Pain.” She holds a PhD in Computer Science from Princeton University.

Cancer AI

Abstract

Cancers are pathologies driven by genetic mutations that disrupt a multitude of signalling pathways operating across different cell types interacting in highly complex ways. No two cancers, even of the same type, are the same. The holy grail of cancer treatment is to analyse the patient’s genome and predict a sequence and combination of therapies that will destroy that patient’s cancer with no adverse side effects. By developing executable models that can simulate cancer tumours at different levels of abstraction, we are on the threshold of being able to deliver on this vision. The state-of-the-art in executable biology is the use of formal methods and software verification to describe biological systems and explore their properties. Using program synthesis methods we can directly build such models from patients’ data. This approach has already been used to find new more efficient therapies for Leukaemia in partnership with AstraZenenca. The next big question, as we collect more and more patient genomic data and history of cancer treatments, is how can we use AI methods to drive therapeutic regimes directly from patients’ data? In the talk, I will showcase recent results and share some of the ambitions in this space.

Biography

Jasmin Fisher is a Senior Researcher at Microsoft Research Cambridge and an Associate Professor (Reader) of Systems Biology in the Department of Biochemistry at the University of Cambridge. She is a member of the Cambridge Cancer Centre, Cambridge Systems Biology Centre and the Cambridge Stem Cell Institute, and in 2016 she was elected Fellow of Trinity Hall, Cambridge. Jasmin received her PhD in Neuroimmunology from the Weizmann Institute in 2003. She then started her work on the application of formal methods to biology as a postdoctoral fellow in the Department of Computer Science at the Weizmann Institute, and then continued to work on the development of novel formalisms and tools that are specifically-tailored for modelling biological processes in the School of Computer Science at the EPFL in Switzerland. In 2007, Jasmin joined the Microsoft Research Lab in Cambridge. In 2009, she was also appointed a Research Group Leader in the University of Cambridge. Jasmin has devoted her career to develop methods for Executable Biology. She is a pioneer in using formal verification methods to analyse mechanistic models of cellular processes and disease. Her research group focuses on cutting-edge technologies to understand the molecular basis of cancer and the development of novel drug therapies, currently also in use by the pharmaceutical industry.

Probabilistic Machine Learning and AI

View the slides

Abstract

How can a machine learn from experience? Probabilistic modelling provides a mathematical framework for understanding what learning is, and has therefore emerged as one of the principal approaches for designing computer algorithms that learn from data acquired through experience.  The field of machine learning underpins recent advances in artificial intelligence, and data science, and has the potential to play an important role in scientific data analysis. I will highlight some current areas of research at the frontiers of machine learning, including our project on developing an Automatic Statistician.

Biography

Zoubin Ghahramani is a Professor at the University of Cambridge and Chief Scientist at Uber.  He is also Deputy Director of the Leverhulme Centre for the Future of Intelligence, Fellow of the Royal Society and of St John’s College. He was a founding Director of the Alan Turing Institute and co-founder of Geometric Intelligence. His research focuses on probabilistic approaches to machine learning and AI.

The Malmo Collaborative AI Challenge

View the slides

Abstract

A long-term goal of artificial intelligence research is to develop artificial assistants (AI agents) that can collaborate with and empower their users. This goal raises important questions, such as how AI agents may understand a person’s goals, plans and beliefs in order to facilitate effective collaboration.

In this talk I give an overview of the Malmo Collaborative AI Challenge – a challenge our team recently conducted to increase awareness of these research challenges and enable steps towards solutions. I will give an overview of the challenge task and initial findings. Finally, the winning teams of the challenge will present their approach and insights gained towards developing collaborative AI.

Biography

Katja Hofmann is a researcher at Microsoft Research Cambridge. As part of the Machine Intelligence and Perception group, she is research lead of Project Malmo. Before joining Microsoft Research, Katja received her PhD in Computer Science from the University of Amsterdam, her MSc in Computer Science from California State University, and her BSc in Computer Science from the University of Applied Sciences in Dresden, Germany. Katja’s main research goal is to develop interactive learning systems. Her dream is to develop AIs that learn to collaborate with human players in Minecraft.

A Brief Tour Through Optimisation Methods for Machine Learning

View the slides

Abstract

We’ll try to embark on a small guided tour through the current zoo of optimisation algorithms for many machine learning applications. Motivated from convex optimisation, we will cover a selection of main algorithmic building blocks together with their theoretical and practical efficiency properties. Finally, we’ll touch on distributed optimisation for current heterogenous and large systems, and some open challenges.

Biography

Martin Jaggi is a Tenure Track Assistant Professor at EPFL, heading the Machine Learning and Optimization Laboratory. Before that, he was a post-doctoral researcher at ETH Zurich, at the Simons Institute in Berkeley, US, and at École Polytechnique in Paris, France. He has earned his PhD in Machine Learning and Optimization from ETH Zurich in 2011, and a MSc in Mathematics also from ETH Zurich, interrupted with several shorter stints in industry.

Session “Design” – On the Human Side

Abstract

This session addresses what is human about artificial intelligence. The session will inspect this question from three very different perspectives.  Alex Taylor will discuss human perspectives on AI; Anab Jain will comment on how design is shaping and commenting on AI; and Rebecca Fiebrink will speak about how people can learn to train AI tools for their own use. A commentary across fields, this session is intended to open up the discussion about where AI fits in the lives of people and the research challenges that that presents.

Biography

Anab is the Director and Co-founder of critically acclaimed design studio Superflux. The Studio consistently produces inventive work in the realm of emerging technologies for business, cultural, and social purposes. Recent clients include Govt. of UAE, UNDP, Suncorp, Cabinet Office, Microsoft Research and the V&A Museum.

Anab is also Professor of Industrial Design at the University of Applied Arts in Vienna, Austria.

She is the recipient of the Award of Excellence ICSID, UNESCO Digital Arts Award, the Grand Prix at Geneva Human Rights Festival, as well as awards from Apple and the UK Government’s Innovation Department. Her work has been exhibited at MoMA New York, V&A Museum, Science Gallery Dublin, National Museum of China, Vitra Design Museum, and Tate Modern.

Session “Malmo – Interactive Learning”

Biography

Sean Kuno is a Research Program Manager of Microsoft Research Outreach. He is based in Redmond U.S.S. and he is a member of Artificial Intelligence Outreach team. Kuno leads the ideation, design and launch of community programs for AI projects such as Project Malmo, working in partnership with universities and government agencies worldwide.

Kuno joined Microsoft Research Asia in 2009 as a University Relations Manager in Japan. Before he joined Microsoft, he worked for the Japan Science and Technology Agency (JST), the second largest funding agency in Japan, where he had more than four years’ experience of project funding, program management and program evaluation and promotion of basic science research projects and academic exchange events. Before JST, he worked as a manager of marketing and product & business development in the cable and satellite industry in Japan. He received a bachelor degree (1996) and a master’s degree (1998) in Quantum Engineering and Systems Science from the Graduate School of Engineering, the University of Tokyo.

Session “Design” – On the Human Side

Abstract

This session addresses what is human about artificial intelligence. The session will inspect this question from three very different perspectives.  Alex Taylor will discuss how AI can be applied to understanding people; Anab Jain will comment on how design is shaping and commenting on AI; and Rebecca Fiebrink will speak about how people can learn to train AI tools for their own use. A commentary across fields, this session is intended to open up the discussion about where AI fits in the lives of people and the research challenges that that presents.

Biography

Cecily Morrison is a researcher in the Human Experience & Design (HXD) group at Microsoft Research Cambridge. Her research lies at the intersection of Human-Computer Interaction and Artificial Intelligence, understanding how we design and build systems that extend people’s capability through machine and person working in symbiosis.

Monte Carlo and Machine Learning

View the slides

Abstract

“Monte Carlo” methods use random sampling to understand a system, estimate averages, or compute integrals. Monte Carlo methods were amongst the earliest applications run on electronic computers in the 1940s, and continue to see widespread use and research as our models and computational power grow. In machine learning, random sampling is widely used within optimisation methods, and as a way to perform inference in probabilistic models. Here “inference” simply means obtaining multiple plausible settings of model parameters that could have led to some observed data. Obtaining a range of explanations tells us both what we can and cannot know from our data, and prevents us from making overconfident (wrong) predictions.

This talk will describe some current research areas in Monte Carlo methods, and their interaction with machine learning.

Biography

Iain Murray is a reader in Machine Learning at the University of Edinburgh. His research interests include probabilistic reasoning using Machine Learning, Density Estimation, and Markov chain Monte Carlo.

Session “Optimisation”

Biography

Sebastian Nowozin is a principal researcher at Microsoft Research managing the Machine Intelligence and Perception group. He works on models and algorithms for artificial intelligence. Previous to joining Microsoft Research he received his PhD in 2009 from the Max Planck Institute for Biological Cybernetics and the Technical University of Berlin. Sebastian is associate editor of IEEE TPAMI and the Journal of Machine Learning Research and an area chair at machine learning and computer vision conferences (NIPS, ICML, CVPR, ICCV).

Confidential Machine Learning in the Cloud

Abstract

In this talk, I will focus on techniques that allow one to use cloud computing infrastructure for running machine learning algorithms on sensitive and confidential datasets in a secure and privacy-preserving manner. Our goal is to provide users of cloud computing with complete control over their data: no one, including malicious employees and hackers, should be able to access the data without the user’s permission. We start off by isolating machine learning (ML) code and data from the rest of the cloud environment using encryption and secure hardware (for example, Intel SGX). Though this essential step reduces the attack surface significantly, the interaction of the code with untrusted environment via data-dependent accesses to memory can leak confidential information. To this end, we propose data-oblivious counterparts of several machine learning algorithms, including decision trees and neural networks. These algorithms are designed to access memory without revealing secret information about their input. We use algorithmic techniques as well as platform specific hardware features to ensure that only public information, such as dataset size, is revealed. We will see that an efficient implementation of our framework on Intel Skylake processors scales up to realistic datasets, with overheads several orders of magnitude lower than with cryptographic-based approaches.

Biography

Olya Ohrimenko is a researcher in Constructive Security Group at Microsoft Research, Cambridge, and a research fellow at Darwin College, Cambridge University. Her research interests include privacy, integrity and security issues that emerge in the cloud computing environment. Olya received her Ph.D. degree from Brown University in 2013 and a B.CS. (Hons) degree from The University of Melbourne in 2007. She was an intern at Google, Microsoft Research Redmond and IBM Research Zurich.

How to Write a Great Research Paper

View the slides

How to Give a Great Research Talk

View the slides

Abstract

Writing papers and giving talks are key skills for any researcher, but they aren’t easy. In this pair of presentations, I’ll describe simple guidelines that I follow for writing papers and giving talks, which I think may be useful to you too. I don’t have all the answers-far from it-so I hope that the presentation will evolve into a discussion in which you share your own insights, rather than a lecture.

Biography

Simon Peyton Jones, MA, MBCS, CEng, graduated from Trinity College Cambridge in 1980. After two years in industry, he spent seven years as a lecturer at University College London, and nine years as a professor at Glasgow University, before moving to Microsoft Research in 1998. His main research interest is in functional programming languages, their implementation, and their application. He has led a succession of research projects focused around the design and implementation of production-quality functional-language systems for both uniprocessors and parallel machines. He was a key contributor to the design of the now-standard functional language Haskell, and is the lead designer of the widely used Glasgow Haskell Compiler (GHC). He has written two textbooks about the implementation of functional languages. More generally, he is interested in language design, rich type systems, software component architectures, compiler technology, code generation, runtime systems, virtual machines, garbage collection, and more. He is particularly motivated by direct use of principled theory to practical language design and implementation-that’s one reason he loves functional programming so much. He is also keen to apply ideas from advanced programming languages to mainstream settings.

Representation Construction for Reinforcement Learning

View the slides

Confidentiality and Privacy Threats in Machine Learning

View the slides

Abstract

Machine learning models are increasingly deployed widely, necessitating an understanding of the often nuanced interplay between ML and security. In this talk, I’ll provide a brief introduction to some recent research trends in ML security, and then focus on those that deal with confidentiality and privacy.

First, I will discuss recent work showing how to steal machine learning models from prediction APIs. We show simple approaches for extracting near-exact replicas of target linear models, neural networks, and decision trees from production machine learning cloud services such as Amazon Prediction API and BigML.  I’ll then explore one specific privacy threat that arises from access to models, what we call model inversion. In the model inversion attacks we develop, an adversary can exploit access to a model and some partial information about a person to improve their ability to guess sensitive information about that person, such as a recognizable image of their face, their genotype, or private lifestyle behaviours. Finally, I will discuss the ability of malicious training algorithms to exfiltrate training data using the spare capacity of typical machine learning models.

Biography

Thomas Ristenpart is an Associate Professor at Cornell Tech and a member of the Computer Science department at Cornell University. Before joining Cornell Tech in May, 2015, he spent four and a half years as an Assistant Professor at the University of Wisconsin-Madison. He completed his PhD at UC San Diego in 2010.  His research spans a wide range of computer security topics, with recent focuses including new threats to, and improved opportunities for, cloud computing security, confidentiality and privacy in machine learning, and topics in applied and theoretical cryptography.  His work has been featured in the New York Times, the MIT Technology Review, ABC News, U.S. News and World Report, and elsewhere.  His work has been recognized by the UC San Diego Computer Science and Engineering Department Dissertation Award, an NSF CAREER Award, Best Paper Award at USENIX Security 2014, Distinguished Student Paper Award at Oakland 2016, and a Sloan Research Fellowship.

Policy Gradient Methods: Tutorial and New Frontiers

View the slides

Abstract

In this tutorial we discuss several recent advances in deep reinforcement learning involving policy gradient methods. These methods have shown significant success in a wide range of domains, including continuous-action domains such as manipulation, locomotion, and flight. They have also achieved the state of the art in discrete action domains such as Atari. We will provide a unifying overview of a variety of different policy gradient methods, and we will also discuss the formalism of stochastic computation graphs for computing gradients of expectations.

Biography

John Schulman is a research scientist at OpenAI. He received his PhD in computer science from UC Berkeley, after receiving a Bachelor of Science in physics from Caltech. He has written various publications on reinforcement learning and robotics, and has twice co-taught a course on deep reinforcement learning at UC Berkeley.

Welcome to AI Summer School 2017

Biography

Scarlet Schwiderski-Grosche is a Principal Research Program Manager at Microsoft Research Cambridge, working for Microsoft Research Outreach in the Europe, Middle East, and Africa (EMEA) region. She is responsible for academic research partnerships in the region, especially for the Joint Research Centres with Inria in France and EPFL and ETH Zurich in Switzerland. Scarlet has a PhD in Computer Science from University of Cambridge. She was in academia for almost 10 years before joining Microsoft in March 2009. In academia, she worked as Lecturer in Information Security at Royal Holloway, University of London.

HoloLens: Computer Vision meets Mixed Reality

Abstract

Microsoft HoloLens is the world’s first fully-untethered, self-contained Holographic computer, and has been made possible by recent advances in computer vision and AI technology. In this talk, we will explore the incredible combination of hardware and software that goes into building a device that can track your head motion without any external sensors, recognize your hand gestures, and reconstruct the room in 3D, all in real-time on a low-power embedded processor. Finally, we will see some of the exciting opportunities for HoloLens, for both research and real-world applications.

Biography

Jamie Shotton is a Partner Scientist and leads the HoloLens Science team at Microsoft in Cambridge, UK, where his team focuses on the visual understanding of people to improve interaction and communication in mixed reality.  He studied Computer Science at the University of Cambridge, where he remained for his PhD in computer vision and machine learning. He joined Microsoft Research in 2008 where he was a research scientist and head of the Machine Intelligence & Perception group, before founding the HoloLens Science Cambridge team in 2016. His research focuses at the intersection of computer vision, AI, machine learning, and graphics, with particular emphasis on systems that allow people to interact naturally with computers. He has received multiple Best Paper and Best Demo awards at top-tier academic conferences. His work on machine learning for body part recognition for Kinect was awarded the Royal Academy of Engineering’s MacRobert Award 2011, and he shares Microsoft’s Outstanding Technical Achievement Award for 2012 with the Kinect engineering team. In 2014 he received the PAMI Young Researcher Award, and in 2015 the MIT Technology Review Innovator Under 35 Award (“TR35”).

AI for Health

Abstract

The pace of technology continues to promise revolutionary improvements in health, and AI has the potential to positively improve lives for patients and citizens. The reality is that improving healthcare is complex due to the myriad of people, processes, and systems that must work together. We are at a pivotal moment where the convergence of seamless communication, artificial intelligence, personalised medicine, and secure, trusted, hyper-scale cloud computing is giving us hope for a brighter future. We will discuss how AI is driving innovation by designing healthcare experiences that are centred around empowering patients, empowering clinicians, and empowering healthcare providers.

Biography

Dr Kenji Takeda is Director for Azure for Research at Microsoft Research. He is helping researchers take best advantage of cloud computing, including through data science, high-performance computing, and the internet of things. He is working on the application of AI in health, including in medical imaging and using the cloud. He has extensive experience in cloud computing, high performance and high productivity computing, data-intensive science, scientific workflows, scholarly communication, engineering and educational outreach. He has a passion for open science, and developing novel computational approaches to tackle fundamental and applied problems in science and engineering. He is a visiting industry fellow at the Alan Turing Institute, and visiting Associate Professor at the University of Southampton, UK.

Session “Design” – On the Human Side

Abstract

This session addresses what is human about artificial intelligence. The session will inspect this question from three very different perspectives.  Alex Taylor will discuss human perspectives on AI; Anab Jain will comment on how design is shaping and commenting on AI; and Rebecca Fiebrink will speak about how people can learn to train AI tools for their own use. A commentary across fields, this session is intended to open up the discussion about where AI fits in the lives of people and the research challenges that that presents.

Biography

Alex Taylor is a sociologist in the Human Experiences & Design group at Microsoft Research, Cambridge. He has been contributing to both academic and industrial research in Science and Technology Studies and Human-Computer Interaction for almost fifteen years. Showing a broad fascination for the entanglements between social life and machines, his research ranges from empirical studies of technology in everyday life to explorations into method, and the ‘inventive methods’ that are afforded through speculative and material interventions.

Counterfactual Multi-Agent Policy Gradients

View the slides

Abstract

Many real-world problems, such as network packet routing and the coordination of autonomous vehicles, are naturally modelled as cooperative multi-agent systems. In this talk, I overview some of the key challenges in developing reinforcement learning methods that can efficiently learn decentralised policies for such systems. I also propose a new multi-agent actor-critic method called counterfactual multi-agent (COMA) policy gradients. COMA uses a centralised critic to estimate the Q-function and decentralised actors to optimise the agents’ policies.  In addition, to address the challenges of multi-agent credit assignment, it uses a counterfactual baseline that marginalises out a single agent’s action, while keeping the other agents’ actions fixed. COMA also uses a critic representation that allows the counterfactual baseline to be computed efficiently in a single forward pass. Finally, I present results evaluating COMA in the testbed of StarCraft unit micromanagement.

Biography

Shimon Whiteson is an associate professor in the Department of Computer Science at the University of Oxford, and a tutorial fellow at St. Catherine’s College. His research focuses on artificial intelligence, with a particular focus on machine learning and decision-theoretic planning. In addition to theoretical work on these topics, he has in recent years also focused on applying them to practical problems in robotics and search engine optimisation. He studied English and Computer Science at Rice University before completing a doctorate in Computer Science at the University of Texas at Austin in 2007.  He then spent eight years as an Assistant and then an Associate Professor at the University of Amsterdam before joining Oxford as an Associate Professor in 2015.  He was awarded an ERC Starting Grant from the European Research Council in 2014 and a Google Faculty Research Award in 2017.

Distributed Bayesian Learning with Stochastic Natural-gradient Expectation Propagation and the Posterior Server

Abstract

View the slides

We make two contributions to Bayesian machine learning algorithms. Firstly, we propose stochastic natural gradient expectation propagation (SNEP), a novel alternative to expectation propagation (EP), a popular variational inference algorithm. SNEP is a black box variational algorithm, in that it does not require any simplifying assumptions on the distribution of interest, beyond the existence of some Monte Carlo sampler for estimating the moments of the EP tilted distributions. Further, as opposed to EP which has no guarantee of convergence, SNEP can be shown to be convergent, even when using Monte Carlo moment estimates.

Secondly, we propose a novel architecture for distributed Bayesian learning which we call the posterior server. The posterior server allows scalable and robust Bayesian learning in cases where a dataset is stored in a distributed manner across a cluster, with each compute node containing a disjoint subset of data. An independent Markov chain Monte Carlo (MCMC) sampler is run on each compute node, with direct  access only to the local data subset, but which targets an approximation to the global posterior distribution given all data across the whole cluster. This is achieved by using a distributed asynchronous implementation of SNEP to pass messages across the cluster. We demonstrate SNEP and the posterior server on distributed Bayesian learning of logistic regression and neural networks.

Biography

Yee Whye Teh is a Professor of Statistical Machine Learning at the University of Oxford, working as a University Lecturer at the Department of Statistics, and as a Research Scientist at DeepMind. He is a European Research Council Consolidator Fellow and an Alan Turing Institute Faculty Fellow.

From 2013 to 2016 he was also a Tutorial Fellow at University College, Oxford. From 2007 to 2012 Yee Whye was Lecturer then Reader of Computational Statistics and Machine Learning at the Gatsby Computational Neuroscience Unit, UCL. Prior to UCL he was Lee Kuan Yew Postdoctoral Fellow at the National University of Singapore and a postdoctoral fellow at University of California at Berkeley working with Michael I. Jordan.

He obtained his Ph.D. in Computer Science at the University of Toronto in 2003 with Geoffrey Hinton, and his B.Math. in Computer Science and Pure Mathematics at the University of Waterloo in 1997.

Session “Language”

Biography

Dr John Winn is a principal researcher in machine learning at Microsoft Research Cambridge.  He obtained his Ph.D. in 2003 from the University of Cambridge, having previously studied at the Massachusetts Institute of Technology AI Lab.  John’s research focuses on tools and languages that make it easier to do machine learning – with applications including information retrieval, computational biology, healthcare and machine vision.  His work has been used in the Xbox Kinect body tracker and for detecting email clutter in Microsoft Office.  In addition, he has created a deep learning paint program that could only paint horses and an object recognizer that caused Bill Gates to take off his glasses.  His current focus is on probabilistic programming for very large scale text understanding.

Applications of Deep Learning in Spoken Dialogue Systems

View the slides

Abstract

Spoken dialogue systems (SDS) provide the core enabling technology for building intelligent personal assistants.  The function of an SDS is to understand each user input, decode the users intention or goal and then respond accordingly. Whereas historically much of this functionality has been provided by hand-crafted rule systems, modern systems increasingly rely on statistical models and machine learning. This talk will briefly review the structure and components of a modern SDS and show how they can be implemented using deep neural networks. The provision of adequate quantities of annotated training data is a major limitation on progress and the talk will conclude with some recent work on end-to-end training which has the potential to greatly reduce or even eliminate the need for labelled data.

Biography

Steve Young is Professor of Information Engineering at Cambridge University and a Senior Member of Technical Staff in the Apple Siri Development team based in Cambridge, UK.  His main research interests lie in the area of statistical spoken language systems including speech recognition, speech synthesis and dialogue management.  He is the recipient of a number of awards including an IEEE Signal Processing Society Technical Achievement Award, an ISCA Medal for Scientific Achievement and an IEEE James L Flanagan Speech and Audio Processing Award. He is a Fellow of the Royal Academy of Engineering and the Institute of Electrical and Electronics Engineers (IEEE).