HLC

Human-Like Computing Third Wave of AI Workshop (3AI-HLC 2019)


Invited Speakers


Stephen Muggleton
Professor Stephen Muggleton, Department of Computing, Imperial College London

Title: What do we need from Third Wave Artificial Intelligence

Abstract:
Third Wave Artificial Intelligence is a term introduced by DARPA to describe new forms intelligent computer systems which transcend black box machine learning. This presentation analyses key drivers which motivate such developments. In  doing so we revisit discussions, from the late 1980s, between Geoff Hinton and Donald Michie concerning the expected contributions of Neural Networks versus Symbolic Machine Learning. At the time Hinton argued the Scientific case for Neural Networks as testable models for Neurobiology. By contrast, Michie advocated the Engineering value of learned models which were interpretable to the user and could even be used for coaching skills in games such as Chess. In this talk we describe recent research at Imperial College London in which prototype Third Wave AI systems have been developed which use modern Symbolic Machine Learning techniques. These systems support natural language explanation of symbolic Machine Learned solutions for areas ranging from game-playing to high-level computer vision. In each case models are learned from fewer examples than comparable Second-Wave AI systems, and often require minimal supervision. The data efficiency and explanaibility of such systems is compatible with interactive learning involving users. We close the talk by summarising challenges to be addressed in future research.

Short biography:
Professor Stephen Muggleton FREng FAAAI is Professor of Machine Learning in the Department of Computing at Imperial College London, Director of the UK's Human-Like Network and is internationally recognised as the founder of the field of Inductive Logic Programming. SM’s career has concentrated on the development of theory, implementations and applications of Machine Learning, particularly in the field of Inductive Logic Programming (ILP) and Probabilistic ILP (PILP). Over the last decade he has collaborated with biological colleagues, such as Prof Mike Sternberg, on applications of Machine Learning to Biological prediction tasks. SM’s group is situated within the Department of Computing and specialises in the development of novel general-purpose machine learning algorithms, and their application to biological prediction tasks. Widely applied software developed by the group includes the ILP system Progol (publication has over 1700 citations on Google Scholar) as well as a family of related systems including ASE-Progol (used in the Robot Scientist project), Metagol and Golem.

Francesca Toni
Professor Francesca Toni, Department of Computing, Imperial College London

Title: Extracting Dialogical Explanations for Review Aggregations with Argumentative Dialogical Agents

Abstract:
The aggregation of online reviews is fast becoming the chosen method of quality control for users in various domains, from retail to entertainment. Consequently, fair, thorough and explainable aggregation of reviews is increasingly sought-after. We consider the movie review domain, and in particular Rotten Tomatoes’ ubiquitous (and arguably over-simplified) aggregation method, the Tomatometer Score (TS). For a movie, this amounts to the percentage of critics giving the movie a positive review. We define a novel form of argumentative dialogical agent (ADA) for explaining the reasoning within the reviews. ADA integrates: 1.) NLP with reviews to extract a Quantitative Bipolar Argumentation Framework (QBAF) for any chosen movie to provide the underlying structure of explanations, and 2.) gradual semantics for QBAFs for deriving a dialectical strength measure for movies, as an alternative to the TS, satisfying desirable properties for obtaining explanations. We evaluate ADA using some prominent NLP methods and gradual semantics for QBAFs. We show that they provide a dialectical strength which is comparable with the TS, while at the same time being able to provide dialogical explanations of why a movie obtained its strength via interactions between the user and ADA.

Short biography:
Francesca Toni is Professor in Computational Logic in the Department of Computing, Imperial College London, UK, and the funder and  leader of the CLArg (Computational Logic and Argumentation) research group. Her research interests  lie within the broad area of Knowledge Representation and Reasoning in Artificial Intelligence, and in particular include Argumentation, Logic-Based Multi-Agent Systems, Logic Programming for Knowledge Representation and  Reasoning, Non-monotonic and Default Reasoning. She graduated, summa cum laude, in Computing at the University of Pisa, Italy, in 1990, and received her PhD in Computing in 1995 from Imperial College London. She has coordinated two EU projects, received funding from EPSRC and the EU, and awarded a Senior Research Fellowship from The Royal Academy of Engineering and the Leverhulme Trust. She is currently Technical Director of  the ROAD2H EPSRC-funded project. She has co-chaired ICLP2015 (the 31st International Conference on Logic Programming) and KR 2018 (the 16th Conference on Principles of Knowledge Representation and Reasoning). She is a member of the steering committe of AT (Agreement Technologies) and KR Inc (Principles of Knowledge Representation and Reasoning, Incorporated), corner editor on Argumentation for the Journal of Logic and Computation , and in the editorial board of the Argument and Computation journal and the AI journal.

Ute Schmid
Professor Ute Schmid, Cognitive System Group, University of Bamberg

Title: Cooperative Learning with Mutual Explanations

Abstract :
Explainable AI most often refers to visual highlighting of information which is relevant for the classification decision for a given instance. In contrast, interpretable machine learning means that the learned models are represented in a human readable form. While a transparent and comprehensible model communicates how a class can be characterized (e.g., what is a cat in contrast to non-cat instances), an explanation gives reasons why a specific instance is classified in a specific way (e.g., why is this image classified as a cat). I will argue that presenting either visualisations or rules to a user will often not suffice as a helpful explanation. Visualisations can only inform the user of the relevance of certain features but not on the relevance of a specific feature value. For example, when classifying a facial expression, it is not enough to explain that the region of the eye contributes to the classification. It is of importance whether the eye is wide open or whether the lid is tightened to classify an expression as surprise or pain. Futhermore, visual highlighting can only inform about conjunction of features (e.g., there is a red block and a green block) while rules can convey relations (e.g., the red block is on the green block) and negation. Often, a combination of visual and textual explanations might be most helpful for a user. Additionally, near-miss examples can help to increase the understanding of what aspects of an instance are crucial for class membership. It can be assumed that explanations are not "one size fits all" but that it depends on the user, the problem, and the current situation which type of explanation is most helpful. Finally, I will present a new method which allows a machine learning system to exploit not only class corrections but also explanations from the user to correct and adapt learned models in interactive, cooperative learning scenarios.

Short biography:
Ute Schmid holds a diploma in psychology and a diploma in computer science, both from Technical University Berlin (TUB), Germany. She received her doctoral degree (Dr. rer.nat.) in computer science from TUB in 1994 and her habilitation in computer science in 2002. From 1994 to 2001 she was assistant professor (wissenschaftliche Assistentin) at the AI/Machine Learning group, Department of Computer Science, TUB. Afterwards she worked as lecturer (akademische Rätin) for Intelligent Systems at the Department of Mathematics and Computer Science at University Osnabrück. Since 2004 she holds a professorship of Applied Computer Science/Cognitive Systems at the University of Bamberg. Research interests of Ute Schmid are mainly in the domain of comprehensible machine learning, explainable AI, and high-level learning on relational data, especially inductive programming, knowledge level learning from planning, learning structural prototypes, analogical problem solving and learning. Further research is on various applications of machine learning (e.g., classifier learning from medical data and for facial expressions) and empirical and experimental work on high-level cognitive processes. Ute Schmid dedicates a significant amount of her time to measures supporting women in computer science and to promote computer science as a topic in elementary, primary, and secondary education.
Nick Chater
Professor Nick Chater, Warwick Business School, University of Warwick

Title: Virtual bargaining - A microfoundation for the theory of social interaction

Abstract:
How can people coordinate their actions or make joint decisions? One possibility is that each person attempts to predict the actions of the other(s), and best-responds accordingly. But this can lead to bad outcomes, and sometimes even vicious circularity. An alternative view is that each person attempts to work out what the two or more players would agree to do, if they were to bargain explicitly. If the result of such a "virtual" bargain is "obvious," then the players can simply play their respective roles in that bargain. I suggest that virtual bargaining is essential to genuinely social interaction (rather than viewing other people as instruments), and may even be uniquely human. This approach aims to respect methodological individualism, a key principle in many areas of social science, while explaining how human groups can, in a very real sense, be "greater" than the sum of their individual members.

Short biography:
Nick Chater is Professor of Behavioural Science at Warwick Business School. He works on the cognitive and social foundations of rationality and language. He has published more than 250 papers, co-authored or edited more than a dozen books, has won four national awards for psychological research, and has served as Associate Editor for the journals Cognitive Science, Psychological Review, and Psychological Science. He was elected a Fellow of the Cognitive Science Society in 2010 and a Fellow of the British Academy in 2012. Nick is co-founder of the research consultancy Decision Technology and is a member on the UK’s Committee on Climate Change. He is the author of The Mind is Flat (2018).

Murray Shanahan
Professor Murray Shanahan, Google DeepMind & Imperial College London

Title : Reconciling Deep Learning with Symbolic AI

Abstract :
In spite of its undeniable effectiveness, conventional deep learning architectures have a number of limitations, such as data inefficiency, brittleness, and lack of interpretatbility. One way to address these limitations is to import a central idea from symbolic AI, namely the use of compositional representations based on objects and relations. In this talk I will discuss recent work on neural network architectures that learn to acquire and exploit relational information, which are a step in this direction.

Short biography:
Murray Shanahan is Professor of Cognitive Robotics in the Dept. of Computing at Imperial College London, and a senior research scientist at DeepMind. Educated at Imperial College and Cambridge University (King’s College), he became a full professor at Imperial in 2006, and joined DeepMind in 2017. His publications span artificial intelligence, robotics, machine learning, logic, dynamical systems, computational neuroscience, and philosophy of mind. He has written several books, including “Embodiment and the Inner Life” (2010) and “The Technological Singularity” (2015). His main current research interests are neurodynamics, deep reinforcement learning, and the future of AI.
Kristian Kersting
Professor Kristian Kersting, Technische Universität Darmstadt

Title: Deep Machines That Know When They Do not Know

Abstract:
Our minds make inferences that appear to go far beyond standard machine learning. Whereas people can learn richer representations and use them for a wider range of learning tasks, machine learning algorithms have been mainly employed in a stand-alone context, constructing a single function from a table of training examples. In this talk, I shall touch upon a view on machine learning, called probabilistic programming, that can help capturing these human learning aspects by combining high-level programming languages and probabilistic machine learning — the high-level language helps reducing the cost of modelling and probabilities help quantifying when a machine does not know something. Since probabilistic inference remains intractable, existing approaches leverage deep learning for inference. Instead of “going down the full neural road,” I shall argue to use sum-product networks, a deep but tractable architecture for probability distributions. This can speed up inference in probabilistic programs, as I shall illustrate for unsupervised science understanding, and even pave the way towards automating density estimation, making machine learning accessible to a broader audience of non-experts. This talk is based on joint works with many people such as Carsten Binnig, Zoubin Ghahramani, Andreas Koch, Alejandro Molina, Sriraam Natarajan, Robert Peharz, Constantin Rothkopf, Thomas Schneider, Patrick Schramwoski, Xiaoting Shao, Karl Stelzner, Martin Trapp, Isabel Valera, Antonio Vergari, and Fabrizio Ventola.

Short biography:
Kristian Kersting is a Professor (W3) for Machine Learning at the TU Darmstadt University, Germany. After receiving his Ph.D. from the University of Freiburg in 2006, he was with the MIT, Fraunhofer IAIS, the University of Bonn, and the TU Dortmund University. His main research interests are statistical relational artificial intelligence (AI) and probabilistic deep learning. Kristian has published over 160 peer-reviewed technical papers and co-authored a book on statistical relational AI. He regularly serves on the PC (often at senior level) for several top conference (NeurIPS, AAAI, IJCAI, KDD, ICML, UAI, ECML PKDD etc.), co-chaired the PC of ECML PKDD 2013, 2020 and UAI 2017, and is the elected PC co-chair of ECML PKDD 2020. He is the Speciality Editor-in-Chief for Machine Learning and AI of Frontiers in Big Data, and is/was an action editor of TPAMI, JAIR, AIJ, DAMI, and MLJ.