ML for the Real World: Distributionally Robust Extrapolation
summary
Introduces a distributionally robust minimax framework that gives formal extrapolation guarantees under covariate shift, constraining an adversary by source features and evaluating loss on an unlabeled target; integrates with deep nets and a feedback-linearization controller in simulation. Implements a safe exploration loop that picks trajectories only if worst-case error bounds permit data collection, yielding faster landings after ~10–15 epochs, dramatically fewer collisions and ~30% higher success rate versus standard planning. conservative extrapolation and adversary-constrained uncertainty.
AI-written summary · about
chapters (19)
- 0:00Introduction
- 0:38Welcome
- 0:59Realworld machining tasks
- 1:26Control experts and social scientists
- 2:05How we usually train machinery models
- 4:23Safety Critical
- 5:08Classification Problem
- 7:38Running Example
- 8:56Roadmap
- 9:15Background
- 11:17Financial robustness
- 11:46Distribution shift
- 14:47Takeaways
- 15:21Demonstration
- 17:20Safe Exploration
- 20:31Spaceship Docking
- 23:40Summary
- 24:46Outlook
- 26:16Questions
related talks
42mToward Understanding the Extrapolation of Nonlinear Models to Unseen Domains or the Whole Domain
34mDistributional Robustness, Learning, and Empirical Likelihood
01h 28mMulti-Distribution Learning, for Robustness, Fairness, and Collaboration
54mGeneralization and Robustness in Offline Reinforcement Learning
01h 01mML for discrete optimization: Theoretical foundations
01h 01mRobustness, Stochastics, Uncertainty 2
41mTransfer learning via local convergence rates of the nonparametric least squares estimator
36mModel-Based Reinforcement Learning with Value-Targeted Regression