[2604.08573] Silhouette Loss: Differentiable Global Structure Learning for Deep Representations

Computer Science > Machine Learning arXiv:2604.08573 (cs) [Submitted on 27 Mar 2026] Title:Silhouette Loss: Differentiable Global Structure Learning for Deep Representations Authors:Matheus Vinícius Todescato, Joel Luís Carbonera View a PDF of the paper titled Silhouette Loss: Differentiable Global Structure Learning for Deep Representations, by Matheus Vin\'icius Todescato and Joel Lu\'is Carbonera View PDF HTML (experimental) Abstract:Learning discriminative representations is a central goal of supervised deep learning. While cross-entropy (CE) remains the dominant objective for classification, it does not explicitly enforce desirable geometric properties in the embedding space, such as intra-class compactness and inter-class separation. Existing metric learning approaches, including supervised contrastive learning (SupCon) and proxy-based methods, address this limitation by operating on pairwise or proxy-based relationships, but often increase computational cost and complexity. In this work, we introduce Soft Silhouette Loss, a novel differentiable objective inspired by the classical silhouette coefficient from clustering analysis. Unlike pairwise objectives, our formulation evaluates each sample against all classes in the batch, providing a batch-level notion of global structure. The proposed loss directly encourages samples to be closer to their own class than to competing classes, while remaining lightweight. Soft Silhouette Loss can be seamlessly combined with cross...