Explore the challenges and implications of task specification in artificial intelligence through this MIT EI seminar featuring Dr. Pulkit Agrawal. Delve into the fundamental mismatch between human communication of tasks and machine understanding, examining issues such as narrow transfer and non-robust feature learning. Investigate potential solutions, including reward design, data augmentation, domain randomization, and contrastive learning. Gain insights into the importance of task specification in achieving true transfer in AI systems, and learn about cutting-edge research in robotics, deep learning, computer vision, and reinforcement learning from a leading expert in the field.
Overview
Syllabus
Embodied Intelligence
reward design
Model Predictions
Task objective
Deep networks likes to CHEAT Learn unwanted solutions
Adversarial Examples are not Bugs, they are Features
How to align Human and DNN features?
Data Augmentation Increases Robustness
Domain Randomization for Transfer
Procgen: Procedurally generated games
Critical Analysis of out-of-distribution Generalization
Consider Objects
What is a Model?
Co-optimization of control and environment
Contrastive Learning
The Task Specification Problem
Taught by
MIT Embodied Intelligence
