Multi proxy anchor family loss for several types of gradients

Shozo Saeki, Minoru Kawahara, Hirohisa Aman

Abstract: Highlights•MPA-family losses can learn the real-world dataset with multi-local centers.•MPA-family losses improve the training capacity of a neural network owing to solving the gradient issues.•MPA-family losses have data-wise or class-wise properties with respect to gradient generation.