Silhouette Loss: Differentiable Global Structure Learning for Deep Representations

Researchers Matheus Todescato and Joel Carbonera have introduced Soft Silhouette Loss, a novel training objective for deep learning models that directly optimizes the geometric structure of learned representations. Unlike standard cross-entropy loss, which focuses solely on correct classification, this new loss function explicitly encourages two desirable properties: intra-class compactness (samples from the same class cluster together) and inter-class separation (different classes are well-distanced). The technique is inspired by the classical silhouette coefficient from clustering analysis but reformulated as a differentiable function suitable for gradient-based optimization.

In practical terms, Soft Silhouette Loss evaluates each sample against all classes within a training batch, providing a global view of structure rather than just pairwise relationships. This makes it computationally lighter than alternatives like supervised contrastive learning (SupCon), which requires comparing many sample pairs. The researchers propose using it alongside cross-entropy in a hybrid objective, jointly optimizing for both accurate classification and well-structured embeddings. Extensive testing on seven diverse datasets showed consistent improvements, with the hybrid method achieving the best performance.

The results are significant: augmenting standard cross-entropy training with Soft Silhouette Loss improved average top-1 accuracy from 36.71% to 39.08%, a 6.5% relative gain. It also outperformed supervised contrastive learning alone (37.85%) while incurring substantially lower computational overhead. This demonstrates that classical clustering principles can be effectively translated into modern deep learning objectives, offering a more efficient path to better model performance without dramatic increases in training cost.