Greedy Robots in a Changing World – Forgetting and Transfer in Continual Reinforcement Learning
Although the world we’re living in is changing rapidly, deep learning models are designed primarily for static data and stationary environments. In my talk, I will advocate for continual reinforcement learning, a challenging yet promising paradigm of training RL agents from a stream of evolving tasks and scenarios. A continual learner should be able to adapt to new tasks by leveraging the experience it already possesses and retain the knowledge of the previously acquired skills. In our NeurIPS 2021 and NeurIPS 2022 papers, we proposed a robotics benchmark dedicated to this setting and used it to analyze several existing solutions. Based on these investigations, I will talk about the promises of this paradigm, the major problems still to solve, and the question at the center of this problem – what does it mean to adapt?
Maciej is a fourth-year PhD student at GMUM at the Jagiellonian University. He is working on the efficiency and adaptability of machine learning systems in areas such as continual learning, adaptive computation, and generative models. He regularly publishes and reviews at several top ML venues (ICML, NeurIPS, ICLR, AAAI). During his PhD he did internships at ETH Zurich and Lyft Level-5. He co-organized the EEML 2020 summer school and MLSS^N 2022 in Kraków.