A Softmax-Free Loss Function Based on Predefined Optimal-Distribution of Latent Features for Deep Learning Classifier

In the field of pattern classification, training deep learning classifiers is mostly end-to-end learning, and loss function constraint on final output (posterior probability) network, so existence Softmax essential. case there usually no effective that completely relies features middle layer to restrict resulting in distribution sample latent not optimal, still room for improvement classification accuracy. Based concept Predefined Evenly-Distributed Class Centroids (PEDCC), this article proposes a Softmax-free based predefined optimal-distribution features—POD Loss. The only restricts samples, including norm-adaptive Cosine distance between feature vector center evenly-distributed class, correlation samples. Finally, used classification. Compared with commonly Loss, some typical related functions PEDCC-Loss, experiments several datasets networks show performance POD Loss always significant better easier converge. Code available https://github.com/TianYuZu/POD-Loss .