Contrastive Learning Improves Feature Attribution in Neural Networks

A new study on arXiv (2604.22540) empirically demonstrates that neural networks trained with Supervised Contrastive Learning (SCL) produce higher-quality feature attribution explanations compared to those trained with cross-entropy. SCL creates an embedding space where similar data points cluster together, offering advantages in adversarial robustness and out-of-distribution detection, making it preferable for safety-critical applications. The research focuses on image classification tasks.