CDC-FSL: : A Causal De-Confounding Framework for Few-Shot Learning

In Few-Shot Learning (FSL), Out-of-Distribution (OOD) samples often introduce confounders, limiting models’ generalization ability. Existing causal-based FSL approaches adjust for these confounders but typically rely on fixed assumptions derived from prior knowledge. In this work, we argue that confounders should be adaptively learned to suit specific scenarios. To address this, we propose a Structural Causal Model (SCM) that explains how confounders lead to misclassification in FSL, particularly when causal features and confounder features are entangled. Building on this SCM, we introduce the Causal De-Confounding framework for FSL (CDC-FSL), which dynamically learns disentangled representations of causal and confounder features. Using a novel co-learning strategy and back-door adjustment, CDC-FSL mitigates confounding effects during FSL recognition. Extensive experiments on standard benchmarks demonstrate that CDC-FSL achieves state-of-the-art performance in both 1-shot and 5-shot settings. Additionally, it exhibits superior cross-domain transferability, outperforming existing methods across diverse target domains.