Invariance, Causality and Novel Robustness

on YouTube ↗44m

" summary

Claim: causal regression coefficients are the minimizers of a worst‑case L2 risk over perturbations that do not act on Y, and a new estimator called anchor regression trades residual fit against correlation with observed anchor variables via a tunable gamma regularizer. It solves a convex objective (add L1 for high‑D), handles anchors that affect hidden confounders or Y, shows ~15–25% predictive gains on heterogeneous data, and extends heuristically to random‑forest‑style nonlinear adjustments, though identifiability can remain limited.

AI-written summary · about

" chapters (15)

1. 0:00Intro
2. 1:43Robust Statistics
3. 3:26Causality
4. 5:19Modern Applications
5. 10:56The Meditators
6. 12:48Heterogeneous Data
7. 14:25Prediction Problem
8. 15:59Robustness
9. 17:02The Causality Solution
10. 24:02Causality Definition
11. 36:57Shifty
12. 39:04Shift perturbation
13. 40:56In practice
14. 41:36Random forests
15. 41:59Heterogeneity

" related talks

• 53mAdversarial Machine Learning and Instrumental Variables for Flexible Causal ModelingVasilis Syrgkanis · Adversarial Approaches in Machine Learning 2022
• 22mPanel DiscussionChristopher Harshaw, Liam Solus, Arnab Bhattacharyya · Learning from Interventions 2022
• 12mDistribution-Free RobustnessSitan Chen · Meet the Fellows Welcome Event
• 01h 15mAn overview of classical robust statistics and generalizations to the futurePo-Ling Loh · Modern Paradigms in Generalization Boot Camp 2024
• 01h 23mImplicit Regularization IINati Srebro · Deep Learning Boot Camp 2019
• 45mCausally motivated robustness to shortcut learningMaggie Makar · Domain Adaptation and Related Areas 2024
• 34mDistributional Robustness, Learning, and Empirical LikelihoodJohn Duchi · Optimization, Statistics and Uncertainty
• 45mMulticalibration, Universal Adaptability and CausalityOmer Reingold · Algorithmic Aspects of Causal Inference 2022