Keynotes and Tutorials

AIxIA 2022 comes with a great program of invited talks and tutorials. Refer to the Program Overview page to look for the time slots. All the keynotes and tutorials will take place in the Plenary room.

Keynotes

Despite the surprising power of many modern AI systems that often learn their own representations, there is significant discontent about their inscrutability and the attendant problems in their ability to interact with humans. While alternatives such as neuro-symbolic approaches have been proposed, there is a lack of consensus on what they are about. There are often two independent motivations (i) symbols as a lingua franca for human-AI interaction and (ii) symbols as system-produced abstractions used by the AI system in its internal reasoning. The jury is still out on whether AI systems will need to use symbols in their internal reasoning to achieve general intelligence capabilities. Whatever the answer there is, the need for (human-understandable) symbols in human-AI interaction seems quite compelling. In particular, humans would be interested in providing explicit (symbolic) knowledge and advice -- and expect machine explanations in kind. This alone requires AI systems to maintain a symbolic interface for interaction with humans. In this talk, I will motivate this point of view, and describe recent efforts in our research group along this direction.

Short bio: Subbarao Kambhampati is a professor of computer science at Arizona State University. Kambhampati studies fundamental problems in planning and decision making, motivated in particular by the challenges of human-aware AI systems. He is a fellow of Association for the Advancement of Artificial Intelligence, American Association for the Advancement of Science,  and Association for Computing machinery, and was an NSF Young Investigator. He served as the president of the Association for the Advancement of Artificial Intelligence, a trustee of the International Joint Conference on Artificial Intelligence,  the chair of AAAS Section T (Information, Communication and Computation), and a founding board member of Partnership on AI.

Kambhampati’s research as well as his views on the progress and societal impacts of AI have been featured in multiple national and international media outlets.

He can be followed on Twitter (@rao2z).

I have worked on various subjects, but ever since my years as PostDoc, I have been fascinated by Datalog, a logical programming language for reasoning and database querying. I wil start this talk with a short introduction to Datalog, followed by an overview of some theoretical results on a number of Datalog variants. This will be interleaved with a tale of four Datalog-related companies I co-founded: DLVSystem, Lixto, Wrapidity, and DeepReason.ai. I will also comment on the difficulties an academic has to face when spinning out a start-up, and on the satisfaction you may experience when your results are used in practice.

Short Bio: Georg Gottlob is a Royal Society Research Professor and a Professor of Informatics at Oxford University, and an Adjunct Professor at TU  Wien.  At Oxford he is a Fellow of St John's College.  His interests include knowledge representation, logic and complexity, and database and Web querying. He has received various awards, among which the Wittgenstein Award (Austria) and the Ada Lovelace Medal (UK), and, jointly with co-authors the Alonzo Church Award. He is a Fellow of the Royal Society, of the Austrian and German National Academies of Science, and of the Academia Europaea. He was a founder of Lixto, a company specialised in semi-automatic web data extraction which was acquired by McKinsey in 2013. Gottlob was awarded an ERC Advanced Investigator's Grant for the project "DIADEM: Domain-centric Intelligent Automated Data Extraction Methodology". Based on the results of this project, he co-founded Wrapidity Ltd, a company that specialised in fully automated web data extraction, which was acquired in 2016 by Meltwater. He was also a co-founder of the currently  active company  DLVSystem and of DeepReason.AI Ltd, which is a second recently acquired by Meltwater.

Tutorials

This tutorial reviews the recent advancements made in using constrained optimization to incorporate structural information and domain knowledge into machine learning models. We start by reviewing the foundation of differentiable optimization and discuss how to convert optimization into differentiable layers to use in machine learning models. Such integration enables to enforce structural information and domain knowledge into machine learning models. Next, we focus on extending this setting by integrating constrained optimization to enforce structure in the outputs of learned embeddings, leading to end-to-end decision-focused learning, that trains models to directly optimize the performance in targeted applications. Finally. we review techniques to learn constrained optimization surrogates by leveraging a distribution of optimization problems and their solutions leading to enhanced optimization modeling technology for operations research decision tasks. The tutorial concludes with a discussion of challenges and open questions.

Speaker bio: Ferdinando Fioretto is an assistant professor at Syracuse University, NY. He works on machine learning, optimization, differential privacy, and fairness. His recent work focuses on (1) the integration of constrained optimization and machine learning to enhance the expressive ability of machine learning models and (2) understanding the interplay among privacy, equity, robustness, and performance in machine learning models and decision tasks. He is a recipient of the 2023 Amazon Research Award, 2022 NSF CAREER award, the 2022 Google Research Scholar Award, the 2022 Caspar Bowden PET award, the 2021 ISSNAF Mario Gerla Young Investigator Award, the 2021 ACP Early Career Researcher Award, the 2017 AIxIA Best AI dissertation award, and several best paper awards.

In the last decades, the research in Artificial Intelligence (AI) has reached remarkable results in a variety of fields ranging from computer vision to language technologies. Despite these results, however, the realization of artificial systems endowed with human-level competencies in real-life scenarios is still far from being achieved and, in general, the gap between natural and artificial intelligence is still enormous. Given this state of affairs, in recent years, the area of cognitively inspired artificial systems has attracted a renewed attention, both from academia and industry, and the awareness about the need for additional research in this interdisciplinary intersection is gaining widespread acceptance. In this tutorial, I will show how the insights coming from the cognitive research can be used to shape the design of intelligent artificial systems with the twofold goal of i) building better (i.e. more intelligent) machines and ii) detecting and discovering, via simulation, novel aspects and mechanisms of the biological system taken as a source of inspiration.

Short bio: Antonio Lieto is a researcher in Computer Science at the Department of Computer Science of the University of Turin (Italy) and at the ICAR-CNR Institute in Palermo (Italy). His main research topics include commonsense reasoning, language and knowledge technologies, cognitive architectures for intelligent interactive agents (embodied and not). He is the current Vice-President of the Italian Association of Cognitive Sciences (AISC, 2017-2022), the recipient of the “Outstanding BICA Research Award” from the Biologically Inspired Cognitive Architecture Society (USA), and is an ACM Distinguished Speaker on the topics of cognitively inspired AI. He has authored the book “Cognitive Design for Artificial Minds” (Routledge/Taylor & Francis, 2021).