Research Centre for Machine Learning Seminar

Prof Alan Bundy, CBE, FRS, University of Edinburgh

Date & Time: Wed, 26 Oct 2016, 12:00noon

Venue: AG21 (College Building)

TITLE: Reformation: a generic algorithm for repairing faulty representations.

ABSTRACT: Autonomous agents require representations of their environment to interpret sensory data, make plans to achieve their goals and solve other problems. Good representations are key to effective problem solving. Historically, they have been manually constructed and tuned to suit a particular task and environment. As we aspire to build persistent, autonomous agents in changing environments interacting with changing populations of other agents and changing tasks, then these agents’ representations must also evolve automatically. Such evolution will not just be to change beliefs within a fixed language, but the language itself must also sometimes change, i.e., old concepts will be replaced with new. Faults in representations are often detected by failures of inference, e.g., inconsistency and incompleteness. The reformation algorithm generalises our previous work on domain-specific representational repair to provide a generic mechanism with a potentially wide application. I will describe reformation and illustrate it on a variety of examples.

Speaker Bio: Alan Bundy is Professor of Automated Reasoning in the School of Informatics at the University of Edinburgh.  His research interests include: the automation of mathematical reasoning, with applications to reasoning about the correctness of computer software and hardware; and the automatic construction, analysis and evolution of representations of knowledge. His research combines artificial intelligence with theoretical computer science and applies this to practical problems in the development and maintenance of computing systems.  He is the author of over 290 publications and has held over 60 research grants. He is a fellow of several academic societies, including the Royal Society, the Royal Society of Edinburgh, the Royal Academy of Engineers and the Association for Computing Machinery. His awards include the IJCAI Research Excellence Award (2007), the CADE Herbrand Award (2007) and a CBE (2012). He was: Edinburgh’s founding Head of Informatics (1998-2001); founding Convener of UKCRC (2000-05); and a Vice President and Trustee of the British Computer Society with special responsibility for the Academy of Computing (2010-12). He was also a member of: the Hewlett-Packard Research Board (1989-91); the ITEC Foresight Panel (1994-96); both the 2001 and 2008 Computer Science RAE panels (1999-2001, 2005-8); and the Scottish Science Advisory Council (2008-12).

The Research Centre for Machine Learning (RCML) at City is pleased to announce the seminar by Dr. Luis Lamb – Professor and Dean (Director) of the Institute of Informatics (2011-2015 & 2015-2019), ex officio (2011-2015 & 2015-2019) and Elected (2010-2012) Member of the University Council at the Federal University of Rio Grande do Sul, Porto Alegre, Brazil..

Please find below the details of the talk:

Venue: A108 (College Building)

Date & Time: Jun 24, 2016 (12:00-13:00)

Title: Learning and Reasoning in AI and Cognitive Computation

Abstract: Cognitive Computation and Artificial Intelligence have received increasing interest from both academics and relevant industries. The recent impact of AI and machine learning applications have led many to question the future influence that these research fields may have upon human life. Moreover, several leading scientists and intellectuals have voiced concerns about possible consequences of Artificial Intelligence, Cognitive Computing and Machine Learning. In this talk, we highlight recent developments on the research towards the integration of robust reasoning and learning mechanisms in cognitive computing and AI and the impact that these fields may have in the near future.

Speaker Bio: Luis Lamb is Professor and Dean (Director) of the Institute of Informatics (2011-2015 & 2015-2019), ex officio (2011-2015 & 2015-2019) and Elected (2010-2012) Member of the University Council at the Federal University of Rio Grande do Sul, Porto Alegre, Brazil. He was Deputy Dean of the Institute of Informatics at UFRGS from August 2006 to October 2011.

He holds both the Ph.D. in Computing Science from the Imperial College London (2000) and the Diploma of the Imperial College (D.I.C.) (2000), MSc by research (1995) and BSc in Computer Science (1992) from the Federal University of Rio Grande do Sul, Brazil. In 2010 he received the MIT Executive Certificate in Strategy and Innovation and in 2014 he received the Executive Certificate in Management and Leadership (Massachusetts Institute of Technology – Sloan School of Management). He is Honorary Visiting Fellow at the Department of Computing, City University London and Visiting Research Fellow, Abductive Systems Group, Department of Philosophy, University of British Columbia, Canada (group led by John Woods).

His research interests include: Logic in Computer Science and Artificial Intelligence, Neural Computation; Social Computing and Computing in the Physical and Social Sciences. Lamb has co-authored two research monographs: Neural-Symbolic Cognitive Reasoning, with d’Avila Garcez and Gabbay (Springer 2009) and Compiled Labelled Deductive Systems, with Broda, Gabbay and Russo (IoP 2004). He is co-editor-in-chief of the Revista de Informática Teórica e Aplicada and he is on the editorial board of the Logic Journal of the IGPL (Oxford) and of the Journal of the Brazilian Computer Society (Springer). Lamb’s research has led to publications in ACM Transactions on Autonomous and Adaptive Systems, Behavioral and Brain Sciences, Theoretical Computer Science, Neural Computation, Journal of Logic and Computation, IEEE Transactions on Neural Networks and Learning Systems, Physica A, Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, The Journal of Theoretical Biology, and at the flagship Artificial Intelligence and Neural Computation conferences AAAI, IJCAI, NIPS, HCOMP. His research on embedded systems software modelling and adaptation has led to publications at ICSE-11, ASE2010, and DATE2008. He was co-organizer of the Dagstuhl Seminar 14381: Neural-Symbolic Learning and Reasoning in September 2014. He is, or has been, member of the Programme or Organizing Committee of a large number of international conferences and workshops on Artificial Intelligence, Neural and Cognitive Computation, Logic in Computer Science, Embedded Systems and Formal Methods. Lamb holds an Advanced Research Fellowship (2010-2017) from the Brazilian National Research Council CNPq. He is a professional member of the ACM, ACM SIGACT, AAAI, AMS, ASL, IEEE, C&GCA, and the Brazilian Computer Society.

The Research Centre for Machine Learning (RCML) at City is pleased to announce the seminar by Dr. Lucian Busoniu – associate professor with the Department of Automation at the Technical University of Cluj-Napoca.

Please find below the details of the talk:

Venue: AG21 (College Building)

Date & Time: Jun 30, 2016 (12:00-13:00)

Title: Planning Methods for Near-Optimal Nonlinear Control

Abstract: We propose an optimistic planning method to search for near-optimal sequences of actions in discrete-time, infinite-horizon optimal control problems with discounted rewards. The dynamics are nonlinear, while the action is continuous and lies in a bounded interval. The method works like model-based predictive control, but exploits insights from bandit theory in reinforcement learning to run the search. Specifically, it iteratively refines adaptive-dimensionality hyperboxes, always choosing an optimistic box with the largest upper bound on the rewards. Under certain Lipschitz conditions on the dynamics and rewards, a guaranteed near-optimality bound is obtained as a function of the computation invested. We also give an empirical extension that does not require knowing the Lipschitz constants, and works much better in experiments, strongly indicating that such a method is the best next step.

Speaker Bio: Lucian Busoniu received the Ph.D. degree (cum laude) from the Delft University of Technology, the Netherlands, in 2009. He is an associate professor with the Department of Automation at the Technical University of Cluj-Napoca, and has previously held research positions in the Netherlands and France. His research interests include planning for nonlinear optimal control, reinforcement learning and approximate dynamic programming, multiagent systems, and robotics. He received the 2009 Andrew P. Sage Award for the best paper in the IEEE Transactions on Systems, Man, and Cybernetics.

In the next seminar of the City Research Centre for Machine Learning, we will have a talk by Dr. Greg Wayne who is a researcher at Google DeepMind.

Venue: B104 (University Building)

Date & Time: Mar 18, 2016 (12:00-13:00)

Title: Differentiable Neural Computers for Memory-Based Control

Abstract: I will describe a neural network control circuit that interfaces to a large, external memory buffer, which it can learn to read from and write to. I will show that this system, called a Differentiable Neural Computer, excels at learning to represent and compute transformations of data structures. It can also learn strictly from task-related reinforcement signals to compute beneficial courses of action.

Speaker Bio: Greg Wayne received his B.S. in Symbolic Systems from Stanford University, his M.S. in Applied Mathematics from City University of New York, and his Ph.D. in Neuroscience from Columbia University, working in the theoretical neuroscience laboratory of Larry Abbott. Since 2014 he has been at Google DeepMind in London, pursuing research primarily on artificial neural network memory systems and motor control.

Google DeepMind is a research centre whose mission is “to solve intelligence and use it to make the world a better place”.

In the next Machine Learning Group seminar, we will have a talk by Dr. Carlos Eduardo Thomaz who is a Professor of Statistical Pattern Recognition at the Department of Electrical Engineering, Centro Universitario da FEI (FEI-SP), Sao Paulo, Brazil.

Venue: AG10 (College Building)

Date & Time: Sep 30, 2015 (12:00-13:00)

Title: A Photo-Realistic Generator of Most Expressive and Discriminant Changes in 2D Face Images

Abstract: In this talk, I will describe a photo-realistic generator that creates semi-automatically face images of unseen subjects. Unlike previously described methods for generating face imagery, this approach incorporates texture and shape in a single computational framework based on high dimensional encoding of variance and discriminant information from sample groups. The method produces realistic, frontal pose, images with minimum manual intervention, and might be a useful tool for face perception applications where privacy-preserving analysis is an issue and the goal is not the recognition of the face itself, but rather its characteristics like gender, age or race, commonly explored in social and forensic contexts.

Speaker Bio: I am Professor of Statistical Pattern Recognition at the Department of Electrical Engineering, Centro Universitario da FEI (FEI-SP), Sao Paulo, Brazil.  I am also a CNPq Research Fellow and head of the Image Processing Lab funded by FAPESP at FEI.  In 1993, I received my B.Sc. degree in Electronic Engineering from Pontifical Catholic University of Rio de Janeiro (PUC-RJ), Rio de Janeiro, Brazil.  After working for six years in industry, I obtained the M.Sc. degree in Electrical Engineering from PUC-RJ in 1999.  In October 2000, I joined the Department of Computing at Imperial College London where I obtained the Ph.D. degree in Statistical Pattern Recognition in 2004.  I was a Research Associate at the Department of Computing, Imperial College London, from December 2003 to January 2005 working in the UK EPSRC e-science project called Information eXtraction from Images (IXI).  In 2012, I was awarded a University of Nottingham Brazil Visiting Fellowship to work in the Sir Peter Mansfield Magnetic Resonance Centre from middle April to the first week of July. My general interests are in Statistical Pattern Recognition, Computer Vision, Medical Image Computing, and Machine Learning, whereas my specific research interests are in limited-sample-size problems in pattern recognition (Lattes cv, in portuguese).

In the next Machine Learning Group seminar, we will have a talk by Dr. Luke Dickens who is a Lecturer in the Department of Information Studies at UCL.

Venue: A226 (College Building)

Date & Time: Apr 10, 2015 (12:00-13:00)

Title: Part 1: Efficient Knowledge Aquisition in Crowdsourcing ; Part 2: The Human Gamma Project

Abstract:

This talk will be in two parts.

In the first part, I will talk about my work in crowdsourcing and crowdsensing. Crowdsourcing, and its younger sibling crowdsensing, provide ways to harness the time, expertise, intellectual capacity, organisation skill, moral judgement, and distributed nature of large groups of people, from interested laypeople to focused experts. There has been a wealth of work recently, investigating how to establish high-quality ground-truth predictions using multiple semi-trusted sources. The underlying idea behind much of this work uses correspondence between sources for mutual validation. In simple terms, two sources are more likely to agree on a label, if there is a shared cause, such as the sources being reliable. I will discuss when these approaches work, and what can cause them to fail, as well as potential mitigation strategies. I will then go on to talk about our methods that use these models to efficiently acquire new labels, and our techniques for fast ground truth prediction across multiple contexts.

In the second part of the talk, I will briefly outline my work in behavioural modelling of humans undertaking reinforcement tasks, and the implicit discounting we use to choose between short term small gains versus longer term larger rewards. Reinforcement learning models offer a biologically plausible framework in which to study human behaviour in sequential learning tasks. In particular, reward prediction errors found in the brain, have a close analog to ‘temporal differences’ in the widely used temporal difference (TD) machine learning algorithm. I will discuss our psychophysics experiments, designed to elicit human behaviours, and investigate reward discounting characteristics. I will also present some preliminary findings that suggest that humans adapt their reward discounting to certain features of task complexity. This work may help us to develop new reinforcement learning algorithms with adaptive reward discounting.

These works have been undertaken with a number of researchers at Imperial College, and were supported by EPSRC and EIT ICT Labs funding.

Speaker Bio: I am a Machine Learning specialist with a particular interest in reinforcement learning, probabilistic modelling and systems neuroscience. I completed my PhD at Imperial College London under the supervision of Dr Alessandra Russo and Dr Krysia Broda, investigating the use of Reinforcement Learning for non-cooperative multi-agent environments with hidden state.

Since then, I have held a number of post-doctoral posts at Imperial College, engineering and developing machine learning techniques for various application areas, including: security & privacy, behavioural modelling, systems neuroscience, and crowdsourcing. My current research focuses on applying probabilistic modelling and information theory to these domains. I now work as a Lecturer in the Department of Information Studies at UCL.