Hierarchical Prototypes Polynomial Softmax Loss Function for Visual Classification

A well-designed loss function can effectively improve the characterization ability of network features without increasing amount calculation in model inference stage, and has become focus attention recent research. Given that existing lightweight adds a to last layer, which severely attenuates gradient during backpropagation process, we propose hierarchical polynomial kernel prototype this study. In function, addition multiple stages deep neural enhances efficiency return, only multi-layer functions training stage stage. addition, good non-linear expression improves characteristic performance network. Verification on public datasets shows trained with proposed higher accuracy than other functions.