ICML: Where causality meets machine learning

causalitymachine-learningicmlcausal-inferenceexplainable-ai

Abstraction: Growing role of causality in modern machine learning research

Key points:

Dominik Janzing (AWS) co-authored 4 ICML 2022 papers with "causal" in the title; the 2012 paper "On causal and anticausal learning" with Bernhard Schölkopf received a Test of Time Honorable Mention at ICML 2022
Causality interest in ML grew from experimental design, economics, and graphical-models communities; causal discovery — inferring a causal graph from passive observations — is an ambitious open problem
Key ML topics intersecting causality include explainable AI, fairness, and distribution-shift-robust representations
Fundamental formal definitions (e.g., "root cause of an extreme event") still do not exist, indicating an early-stage field
Mathematical tools used span statistics, kernel methods, linear algebra, Shannon and algorithmic information theory, Fourier analysis, group theory, and game theory — no dominant formalism yet
Janzing predicts causality will play an even bigger role in ML over the next decade

Connections: Amazon · Dominik Janzing · Bernhard Scholkopf · Causal Inference · Causal Discovery · Graphical Models · Explainable AI