#invariant-risk-minimization News & Analysis

Unsupervised Representation Learning -- an Invariant Risk Minimization Perspective

Researchers propose a new unsupervised framework for Invariant Risk Minimization (IRM) that learns robust representations without labeled data. The approach introduces two methods - Principal Invariant Component Analysis (PICA) and Variational Invariant Autoencoder (VIAE) - that can capture invariant structures across different environments using only unlabeled data.