On the Properties of Feature Attribution for Supervised Contrastive Learning

The paper's core contribution is not just that SCL is more robust, but that standard feature attribution methods (like gradient-based saliency or SHAP-style approaches) behave differently when the model was never trained to optimize a classification boundary directly. That mismatch between how attribution is computed and how the model actually learned its representations is the buried problem here.

This connects directly to the interpretability thread running through recent arXiv coverage. The ORCA paper from mid-April tackled a structurally similar problem for SVMs: how do you extract meaningful feature contributions from a model whose internal geometry doesn't map cleanly onto standard attribution assumptions? SCL poses the same challenge one level up, in deep neural networks. Both papers are essentially asking whether our interpretability tools are actually measuring what we think they are, or whether they're artifacts of the training objective rather than the model's learned knowledge.

The practical test is whether any safety-critical deployment team (medical imaging or autonomous systems are the obvious candidates) adopts SCL with attribution-aware tooling in the next 12 months. If attribution methods are adapted specifically for contrastive objectives and validated on a public benchmark, that confirms the problem is real and tractable. If the field keeps applying cross-entropy-derived attribution tools to SCL models without adjustment, this paper's warnings will go unheeded.