Reproduce “Mitigating Spurious Correlations with Causal Logit Perturbation” on OpenResearch

About this paper

Xiaoling Zhou, Wei Ye, Rui Xie, Shikun Zhang

Causal Logit Perturbation (CLP) is a framework that integrates causal reasoning into deep learning to mitigate model reliance on spurious correlations by simulating causal interventions via logit modifications. It achieves state-of-the-art or improved performance across various biased learning scenarios, including long-tail learning, noisy label learning, and subpopulation shift.

Key insights

• Spurious correlations in deep learning models can be understood and addressed within a causal framework, specifically by identifying and intervening on 'backdoor paths' in causal graphs.
• Simulating causal interventions through learnable logit perturbations can effectively break spurious correlations without direct manipulation of input data.
• A meta-learning approach can train a perturbation network to generate sample-specific logit modifications, guiding the main classifier to focus on causal features.

Summary by alphaXiv.