Explore cutting-edge techniques in data augmentation for image-based reinforcement learning in this 52-minute seminar by Rob Fergus at MIT. Delve into a model-free reinforcement learning algorithm for visual continuous control that achieves state-of-the-art results on the DeepMind Control Suite, including complex humanoid locomotion. Learn about a self-supervised framework that combines representation learning with exploration through prototypical representations. Discover how pre-trained task-agnostic representations and prototypes enable superior downstream policy learning on challenging continuous control tasks. Gain insights into the latest advancements in computer vision, reinforcement learning, and artificial intelligence from a leading expert in the field.
Data Augmentation for Image-Based Reinforcement Learning
Massachusetts Institute of Technology via YouTube
Overview
Syllabus
Introduction
Outline
Problem
Image Augmentation
Other Augmentation Strategies
Hyper Parameters
Models and Auxiliary Tasks
Results
Atari Benchmark
Image Augmentations
Summary
Dr Q
Dr Qv2
Dreamer
Conclusion
Reinforcement with prototypical representations
Limitations
Task Exploration
Selfsupervised Learning
ProtoRL Approach
Example
Importance of Exploration
Benchmarking
Wrapup
Taught by
MIT Embodied Intelligence
