RELAX: Representation Learning Explainability

Abstract Despite the significant improvements that self-supervised representation learning has led to when from unlabeled data, no methods have been developed explain what influences learned representation. We address this need through our proposed approach, RELAX, which is first approach for attribution-based explanations of representations. Our can also model uncertainty in its explanations, essential produce trustworthy explanations. RELAX explains representations by measuring similarities space between an input and masked out versions itself, providing intuitive significantly outperform gradient-based baselines. provide theoretical interpretations conduct a novel analysis feature extractors trained using supervised unsupervised learning, insights into different strategies. Moreover, we user study assess how well aligns with human intuition show method outperforms baselines both quantitative evaluation studies. Finally, illustrate usability several use cases highlight incorporating be faithful taking crucial step towards explaining