Anthill Inside 2017

Reinforcement Learning (RL) is a natural computational paradigm for agents learning from interaction to achieve a goal. Deep learning (DL) provides a powerful general-purpose representation learning framework. A combination of these two has recently emerged as a strong contender for artificial general intelligence. This tutorial will povide a gentle exposition of RL concepts and DL based RL with a focus on policy gradients. We will briefly touch upon some applications and end with some recent work on multi-modal conversational systems.

Outline

• Introduction to Reinforcement Learning (RL) -- 15 minutes
  (states, actions, rewards, returns)
  (concepts of value function,model,policy,MDPS)
  (highlight key difference from supervised learning)
  (Abstracting problems in a RL paradigm)
  (walk through some concrete examples)

• Deep RL -- 20 minutes
  (value/policy/model based deep RL)
  (PG,DQN,A3C with a focus on policy gradients)
  (Current successes Atari,Go,OpenAI Gym etc)

• Application to multi-modal conversational systems -- 5 minutes
  (end with some current work we are doing)

Speaker bio

Vikas C. Raykar works as a researcher at IBM Research, Bangalore, India. An expert in machine learning he is currently focused on building machines that can understand natural language and images in par with humans. He finished his doctoral studies in the computer science department at the University of Maryland, College Park.