Balancing Robustness and Sensitivity using Feature Contrastive Learning
This addresses a key challenge in deploying robust models for real-world applications, though it is incremental in nature.
The paper tackles the trade-off between robustness and sensitivity in large neural networks by introducing Feature Contrastive Learning (FCL), which improves generalization on noisy vision and NLP datasets.
It is generally believed that robust training of extremely large networks is critical to their success in real-world applications. However, when taken to the extreme, methods that promote robustness can hurt the model's sensitivity to rare or underrepresented patterns. In this paper, we discuss this trade-off between sensitivity and robustness to natural (non-adversarial) perturbations by introducing two notions: contextual feature utility and contextual feature sensitivity. We propose Feature Contrastive Learning (FCL) that encourages a model to be more sensitive to the features that have higher contextual utility. Empirical results demonstrate that models trained with FCL achieve a better balance of robustness and sensitivity, leading to improved generalization in the presence of noise on both vision and NLP datasets.