AASS Seminar - A few observations on combining deep learning and reinforcement learning

The research centre AASS arranges a seminar with Razvan Pascanu, Deepmind. The seminar will be via Zoom, follow the link below to join.

Link to Zoom

Abstract

In this presentation I will start with a brief introduction to deep learning. I will focus on a few fundamental principles and components that power the recent advancement of deep learning, rather than focusing on any architecture in particular. In particular I will talk about the role of function composition, weight sharing, initialization and learning dynamics. Based on this notion, in the second part of the talk I will look at what role neural networks can play in reinforcement learning. In particular my aim is to highlight that treating deep reinforcement learning as just reinforcement learning with neural networks as function approximators, or just deep learning with an RL specific objective can be misleading. I will highlight that updates coming from RL objectives do not form a gradient vector field, an observation known but sometimes overlooked by practitioners and discuss some of its implications. I will discuss how due to the function approximator, DRL systems might be forced to learn tasks sequentially.