FUN2MODEL is a five-year project at the Department of Computer Science, University of Oxford, funded by a £2.4m European Research Council Advanced Investigators Grant and led by Marta Kwiatkowska with Dave Parker, University of Birmingham, as external collaborator. The project starts on 1st October 2019.
Abstract: Machine learning is revolutionising computer science and AI. Much of its success is due to deep neural networks, which have demonstrated outstanding performance in perception tasks such as image classification. Solutions based on deep learning are now being deployed in real-world systems, from virtual personal assistants to self-driving cars. Unfortunately, the black-box nature and instability of deep neural networks is raising concerns about the readiness of this technology. Efforts to address robustness of deep learning are emerging, but are limited to simple properties and function-based perception tasks that learn data associations. While perception is an essential feature of an artificial agent, achieving beneficial collaboration between human and artificial agents requires models of autonomy, inference, decision making, control and coordination that significantly go beyond perception. To address this challenge, this project will capitalise on recent breakthroughs by the PI and develop a model-based, probabilistic reasoning framework for autonomous agents with cognitive aspects, which supports reasoning about their decisions, agent interactions and inferences that capture cognitive information, in presence of uncertainty and partial observability. The objectives are to develop novel probabilistic verification and synthesis techniques to guarantee safety, robustness and fairness for complex decisions based on machine learning, formulate a comprehensive, compositional game-based modelling framework for reasoning about systems of autonomous agents and their interactions, and evaluate the techniques on a variety of case studies. Addressing these challenges will require a fundamental shift towards Bayesian methods, and development of new, scalable, techniques, which differ from conventional probabilistic verification. If successful, the project will result in major advances in the quest towards provably robust and beneficial AI.