Time and place
TBA
Presenters
Alex Lamb (Microsoft Research NYC), John Langford (Microsoft Research NYC)
Contact email: lambalex@microsoft.com
Expected Background
MDPs, some basic deep learning, some self-supervised learning may be useful
Expected Gained Skills
Should be able to write code to learn agent-centric latent state in a simple setting (such as a gridworld with some distractors). Should understand the most important few results and counterexamples from theory.
Brief Description
A latent state can enable vastly better planning, exploration, and credit assignment by keeping task-relevant information while discarding distractions and irrelevant details. For example, in video games, there is a game-engine state which has all relevant information for the underlying dynamics. This tutorial will discuss how we can discover such a latent state in the real world directly from observations. The tutorial will discuss theoretical developments at a high-level, to explain the key pieces of understanding as well as their limitations. The tutorial will discuss where the state-of-the-art is experimentally, and what is currently ready for usage in real-world applications.