[24] Martin Arjovsky - Out of Distribution Generalization in Machine Learning
The Thesis Review
00:00
Is There a Key Idea Behind Invariant Risk Minimization?
The key idea about iron is that in causality problems being the causal parents of a target variable is equivalent to the conditional distribution from these variables to the target. So it's an equivalence between causality and certain statistical properties that are testable being stable across different environments. Then irem is just an algorithm to find the features that have those properties so that's the key idea.
Transcript
Play full episode
