Pairwise Gaussian Loss for Convolutional Neural Networks

Convolutional neural networks (CNNs) have demonstrated great competence in feature representation, and then, achieved a good performance to many classification tasks. Cross-entropy loss, together with softmax, is arguably one of the most commonly used loss functions in CNNs (that is generally called softmax loss). However, the softmax loss can result in a weakly discriminative feature representation since it focuses on the interclass separability rather than the intraclass compactness. This article proposes a pairwise Gaussian loss (PGL) for CNNs that can well address the intraclass compactness through significantly penalizing those similar sample pairs with a relatively large distance. At the same time, PGL can still ensure a good interclass separability. Experiments show that PGL can guarantee that CNNs obtain a better classification performance compared to not only the softmax loss but also others often used in CNNs. Our experiments also show that PGL has a stable convergence for the stochastic gradient descent optimization method in CNNs and a good generalization ability for different structures of CNNs.