Constrainted Loss Function for Classification Problems

Abstract

Producing good representative features is the key to perform high accuracy in conventional methods of computer vision in many tasks. Our work is to show that whether good features are still critical in deep learning models. We bring two loss functions so that one works well in the classification problems and the other achieve good performance in verification problems, together to see whether they will improve the performance on classification problems. In literature, loss functions for classification problems and verification problems are working independently within the deep learning domain, losses of classification emphasis the discrimininative power in distinguishing data of different classes while the verification losses focus on establishing invariant mapping within an embedding via metric learning. In this paper, the major work is to look for a better balancing for loss functions of different kinds. The verification loss together with the classification loss requires a novel scheme in achieving better tradeoff while there are conflicts exited between losses to be involved. Experiments are designed based on the well-known models of ResNet and VGG-16 evaluating on CIFAR-10 and CIFAR-100.