Most of the discriminative feature learning methods are specifically developed for metric learning, however, the effectiveness may be not obvious for other tasks. In this letter, we propose a novel discrimination regularization method for image classification, which enhances the intra-class compactness and inter-class discrepancy simultaneously, through decreasing the minimal angle (DMA) between the feature vector and any one of the weight vectors in classification layer. This method can robustly improve the discriminability and generalizability of neural networks and easily exert its effect by plugging the DMA regularization term into the loss function with negligible computational overhead. The DMA regularization is simple, efficient, and effective. Therefore, it can be used as a basic regularization method for models based on neural networks. We evaluate DMA by applying it to various modern models on CIFAR10, CIFAR100, and TinyImageNet datasets, decreasing the test error rate by 0.2–0.4%, 0.2–1.5%, and 0.3-0.4% respectively. Code is available at: https://github.com/wznpub/DMA_Regularization.

中文翻译：

---

DMA 正则化：通过减小最小角度来增强神经网络的可辨别性

大多数判别特征学习方法是专门为度量学习而开发的，然而，对于其他任务，其有效性可能并不明显。在这封信中，我们提出了一种新的图像分类判别正则化方法，通过减小特征向量与分类中任何一个权重向量之间的最小角度（DMA），同时增强类内紧凑性和类间差异层。该方法可以稳健地提高神经网络的判别性和泛化性，并通过将DMA正则化项插入损失函数中，计算开销可忽略不计，轻松发挥其效果。DMA 正则化简单、高效、有效。因此，它可以作为基于神经网络的模型的基本正则化方法。我们通过将 DMA 应用于 CIFAR10、CIFAR100 和 TinyImageNet 数据集上的各种现代模型来评估 DMA，将测试错误率分别降低 0.2-0.4%、0.2-1.5% 和 0.3-0.4%。代码位于：https://github.com/wznpub/DMA_Regularization。