Continuous Self-supervision and Curiosity - Abhinav Gupta - CoLLAs 2022

Описание: Abstract:
In this talk, I will describe my efforts to build robots that learn like babies: continuously by observing and interacting with the world. In 2013, we released NEIL (one of the world’s first visual never-ending system) which used semantics to build a commonsense knowledge base. After two continuous years of NEIL, we learned several lessons. First, scalability requires self-supervision. Second, commonsense requires not passive internet data but physical interaction and experimentation driven by curiosity. Third, multimodality is a core requirement of lifelong learning. While in last few years, we have made significant advances in each of these areas, I argue we have lost the sense of purpose for continual/lifelong learning. My talk will focus on reestablishing how we can move these research areas back to being the core of lifelong learning. More specifically, I will introduce our recent efforts in developing continuous versions of self-supervision and curiosity/exploration. Finally, I will highlight how democratization is core to good science, yet continual learning suffers from democratization issues. With that in mind, I will introduce the CORA framework, which seeks to be a one stop shop for Continual RL providing benchmarks, new metrics and baseline implementations for several continual learning algorithms.