GO Loss: A Gaussian Distribution-Based Orthogonal Decomposition Loss for Classification

We present a novel loss function, namely, GO loss, for classification. Most of the existing methods, such as center loss and contrastive loss, dynamically determine the convergence direction of the sample features during the training process. By contrast, GO loss decomposes the convergence direction into two mutually orthogonal components, namely, tangential and radial directions, and conducts optimization on them separately. The two components theoretically affect the interclass separation and the intraclass compactness of the distribution of the sample features, respectively. Thus, separately minimizing losses on them can avoid the effects of their optimization. Accordingly, a stable convergence center can be obtained for each of them. Moreover, we assume that the two components follow Gaussian distribution, which is proved as an effective way to accurately model training features for improving the classification effects. Experiments on multiple classification benchmarks, such as MNIST, CIFAR, and ImageNet, demonstrate the effectiveness of GO loss.