Speaker Yoshua Bengio
Wednesday October 21st, 2020 at 9.00 am Montreal Time
Priors for deep learning of semantic representations
Deep learning has been very successful at capturing what mostly corresponds to unverbalizable knowledge which humans possess, in specific domains of applications. The research described here aims at extending deep learning towards representing and reasoning with high-level semantic variables which form the basis for natural language communication and expressing algorithmic knowledge in software. These aspects of the world around us which are captured in natural language and refer to semantic high-level variables often have a causal role (referring to agents, objects, and actions or intentions). These high-level variables also seem to satisfy very peculiar characteristics which low-level data (like images or sounds) do not share, and it would be good to clarify these characteristics in the form of priors which can guide the design of machine learning systems benefitting from these assumptions. Since these priors are not just about the joint distribution between the semantic variables (e.g. it has a sparse factor graph corresponding to a modular decomposition of knowledge) but also about how the distribution changes (typically by causal interventions), this analysis may also help to build machine learning systems which can generalize better out-of-distribution. Introducing such assumptions is necessary to even start having a theory about generalizing out-of-distribution. There are also fascinating connections between these priors and what is hypothesized about conscious processing in the brain, with conscious processing allowing us to reason (i.e., perform chains of inferences about the past and the future, as well as credit assignment) at the level of these high-level variables. This involves attention mechanisms and short-term memory to form a bottleneck of information being broadcast around the brain between different parts of it, as we focus on different high-level variables and some of their interactions. The presentation summarizes a few recent results using some of these ideas for discovering causal structure and modularizing recurrent neural networks with attention mechanisms in order to obtain better out-of-distribution generalization and move deep learning towards capturing some of the functions associated with conscious processing over high-level semantic variables.
Yoshua Bengio is Professor in the Computer Science and Operations Research department at Université de Montréal. He is founder and scientific director of Mila (Quebec Artificial Intelligence Institute) and of IVADO (Institute for Data Valorization). He is a Fellow of the Royal Society of London and of the Royal Society of Canada, and an officer of the Order of Canada. He has received a Canada Research Chair and a Canada CIFAR AI Chair and is a recipient of the 2018 Turing Award for pioneering deep learning. He is a member of the NeurIPS advisory board, co-founder and member of the board of the ICLR conference, and program director of the CIFAR program on Learning in Machines and Brains. His goal is to contribute to uncovering the principles giving rise to intelligence through learning, as well as favour the development of AI for the benefit of all.
Speaker Betty Cheng
Thursday October 22nd, 2020 at 9.00 am Montreal Time
Model-Driven Engineering for Data-Centric Autonomous Systems
Model-Driven Engineering for Data-Centric Autonomous Systems
Increasingly complex autonomous systems involve different types of models and large volumes of data–both coming from disparate communities. Scientific, engineering, and machine learning models are traditionally quite different in format, granularity, time-scale, use, and so forth. Furthermore, these disparate model types make use of heterogeneous data–for instance, time-series data and cross-sectional environmental data, which may be incompatible. As system complexity and data volume both increase, integrating these heterogeneous models and data in a consistent and systematic fashion becomes challenging. Furthermore, disparate disciplines and fields of study perceive and use models differently. One community attempting to use models from another community may not only miss out on their benefits, but may even misuse the models (e.g., by overlooking their limitations or misunderstanding their assumptions). This talk will explore the roles that the various model and data types play in a system’s development. It will also discuss strategies for bridging the gap between model-driven engineering and data-centric autonomous systems development. We will overview MODA, a recently proposed unified conceptual reference framework intended to serve as a scientific foundation for data-centric and model-driven systems engineering. We will illustrate the MODA framework in terms of autonomous systems that we have developed with our collaborators. These instantiations and others have helped us to identify research challenges when using an engineering-based approach to develop autonomous systems that integrate heterogeneous models and data.
Betty H.C. Cheng is a professor in the Department of Computer Science and Engineering at Michigan State University. She is also the Industrial Relations Manager and senior researcher for BEACON, the National Science Foundation Science and Technology Center in the area of Evolution in Action. Her research interests include self-adaptive autonomous systems, requirements engineering, model-driven engineering, automated software engineering, and harnessing evolutionary computation and search-based techniques to address software engineering problems. These research areas are used to support the development of high-assurance adaptive systems that must continuously deliver acceptable behavior, even in the face of environmental and system uncertainty. Example applications include intelligent transportation and vehicle systems. She collaborates extensively with industrial partners in her research projects in order to ensure real-world relevance of her research and to facilitate technology exchange between academia and industry. Her collaborators include Ford, General Motors, ZF, BAE, Motorola, and Siemens. Previously, she was awarded a NASA/JPL Faculty Fellowship to investigate the use of new software engineering techniques for a portion of the NASA space shuttle software. She has recently launched new projects in the areas of model-driven approaches to sustainability, cyber security for automotive systems, and feature interaction detection and mitigation for autonomic systems, all in the context of operating under uncertainty while maintaining assurance objectives. Her research has been funded by several federal funding agencies, including NSF, AFRL, ONR, DARPA, NASA, ARO, and numerous industrial organizations. She serves on the journal editorial boards for Requirements Engineering and Software and Systems Modeling; she is Co-Associate Editor-in-Chief for IEEE Transactions for Software Engineering, where she previously served twice as an Associate Editor. She was the Technical Program Co-Chair for IEEE International Conference on Software Engineering (ICSE-2013), the premier and flagship conference for software engineering. She received her Bachelor of Science degree from Northwestern University, and her MS and PhD from the University of Illinois-Urbana Champaign, all in computer science.
Speaker Krzysztof Czarnecki
Friday October 23rd, 2020 at 9.00 am Montreal Time.
Models in AI-enabled systems: the case of automated driving
Systems and software models and variability are established disciplines to help engineers manage complexity and design flexible systems. These models are expert-built abstractions, often parameterized and composable, to answer engineering questions and guide system development or its runtime execution. In contrast, AI-enabled systems rely mainly on machine-learned models, which are derived purely from data, to handle complex prediction tasks, such as scene recognition or decision making. Machine-learned models have enabled amazing progress in AI, but suffer from the lack of specifications and human interpretability, which is a serious problem in safety-critical applications, such as automated driving. Illustrated with examples from automated driving, this talk will contrast expert- and data-driven models in terms of their strengths and limitations, discuss the role of expert-driven models in AI-enabled systems, explore the integration of both types of models, and identify future directions to achieve best of both worlds.
Krzysztof Czarnecki is a Professor of Electrical and Computer Engineering at the University of Waterloo, where he heads the Waterloo Intelligent Systems Engineering (WISE) Laboratory. He is a leading expert in the safety of automated driving systems (ADS), with focus on assuring the safety of driving behavior and machine-learned functions. As part of his research, he has co-lead the development of UW Moose (started in 2016), Canada’s first self-driving research vehicle, which has been tested on public roads since 2018 (autonomoose.net). His recent research contributions related to ADS safety assurance include an uncertainty-centric framework for assuring the safety of perceptual components based on machine learning, a framework for specifying driving behavior requirements, and methods for modeling and sampling road user behavior. He serves on SAE task forces on driving automation definitions, reference architecture, verification and validation, and maneuvers and behaviors, the Canadian Mirror Committee of ISO TC 22/SC 32 (contribution to ISO/PAS 21448 Safety of the Intended Functionality), and the UL STP 4600 Evaluation of Autonomous Products standards committee. Before working on automated driving, he advised Pratt and Whitney Canada on creating reusable software designs and components for aircraft engine control systems (2011-2015). Before coming to Waterloo, he was a researcher at DaimlerChrysler Research (1995-2002), Germany, focusing on improving software development practices and technologies in enterprise, automotive, and aerospace sectors. He co-authored the book on “Generative Programming” (Addison- Wesley, 2000), which pioneered automated software engineering based on feature modeling, domain-specific languages, and program generation. While at Waterloo, he held the NSERC/Bank of Nova Scotia Industrial Research Chair in Requirements Engineering of Service-oriented Software Systems (2008-2013) and worked on methods and tools for engineering complex software-intensive systems. He received the Premier’s Research Excellence Award in 2004 and the British Computing Society in Upper Canada Award for Outstanding Contributions to IT Industry in 2008. He has also received seven Best Paper Awards, two ACM Distinguished Paper Awards, and two Most Influential Paper Awards. His publications have been widely cited (over 25,000 citations on Google Scholar).