Explore a lecture on interpretable machine learning through program synthesis presented by Osbert Bastani at IPAM's Explainable AI for the Sciences workshop. Delve into novel approaches for creating custom model families using domain-specific programming languages, moving beyond traditional fixed model families. Discover applications in learning interpretable control policies and RNA splice prediction. Examine topics such as parallel parking control, video trajectory queries, deep reinforcement learning, and multi-agent reinforcement learning. Investigate the Viper algorithm, state machine policies, and neurosymbolic transformers. Gain insights into programmatic attention rules, sparse communication structures, and modular networks for RNA splicing in this comprehensive exploration of cutting-edge interpretable machine learning techniques.
Interpretable Machine Learning via Program Synthesis - IPAM at UCLA
Institute for Pure & Applied Mathematics (IPAM) via YouTube
Overview
Syllabus
Intro
What is Interpretability?
RNA Splicing Mechanism
RNA Splice Prediction
Control: Parallel Parking
Learning Interpretable Models
Program Synthesis for Interpretable ML
Video Trajectory Queries
Control & Reinforcement Learning
Deep Reinforcement Learning
Imitation Learning
Dataset Aggregation (DAgger)
Our Approach: Leverage the Q-Function
Viper Algorithm
Verifying Correctness of a Toy Pong Controller
Learning State Machine Policies
Teacher Policy
Interpretability of State Machine Policies
Example: Single Group
Multi-Agent Reinforcement Learning
Transformer Communication Graph
Neurosymbolic Transformers
Learning Algorithm
Programmatic Attention Rules
Sparse Communication Structure
Modular Networks for RNA Splicing
Conclusion
Taught by
Institute for Pure & Applied Mathematics (IPAM)
