Why machine learning struggles with causality - TechTalks

machine-learningcausalitydeep-learninggeneralizationcausal-inference

Abstraction: ML systems lack causal reasoning needed for robust generalization

Key points:

ML models rely on i.i.d. (independent and identically distributed) data and learn statistical correlations rather than causal relationships, limiting generalization outside training distributions
A 2021 paper "Towards Causal Representation Learning" from Max Planck Institute, Mila, and Google Research outlines structural causal models and independent causal mechanisms as paths forward
Causal models allow counterfactual reasoning, domain transfer, and robustness to adversarial attacks — capabilities current deep learning lacks
Self-driving car failures and pandemic-era model degradation are concrete examples of i.i.d. assumption breakdown in the real world
Judea Pearl (Turing Award, causal inference pioneer) and Yoshua Bengio (deep learning pioneer) represent converging perspectives on the need for causal ML
Proposed solution is embedding structural causal models (SCMs) into modular neural architectures for reusable, fine-tunable knowledge modules

Connections: Judea Pearl · Yoshua Bengio · Causal Representation Learning · Reinforcement Learning · Transfer Learning