Talk, DeepMind, London, UK
In this talk I discuss the problem of AI alignment from a generalization perspective. I cover the main issues involved in creating safe AGI for practical and scalable deployment along with several methods, analysis and applications.
Talk, Waymo, Oxford, UK
In this talk I discuss how the problem of generalization in AI can be approached from three different directions towards creating agents that are robust to changes in environment, goals and coplayers. I present several algorithmic approaches along with analysis and applications.
Talk, DeepMind, Oxford, UK
In this talk I discuss how open ended learning in the XLand domain helps create generally capable AI agents, and how intent prediction for co-players in the environment can help improve generalization performance.
Talk, ICML 2021, Oxford, UK
Talk Slides: In this talk I discuss how tensor decompositions can be used for sample efficient Reinforcement Learning for large factored action spaces.
Talk, GoodAI's Meta-Learning & Multi-Agent Learning Workshop, Oxford, UK
Talk link: In this talk I motivate why multi-agent learning would be an important component of AI and elucidate some frameworks where it can be used in designing an AI system.
Talk, NeurIPS 2019, Oxford, UK
Talk Slides: In this talk I discuss the sub-optimality in value based methods arising from representational constraints and propose a new method MAVEN to overcome it.
Talk, NeurIPS 2019, Oxford, UK
Talk Slides: In this talk I discuss the exact framework for encoding RL as probabilistic inference and the algorithms that arise from it.