Hyperspherical Mixed Prototypes Networks

This paper introduces hyperspherical mixed prototype networks. The key difference compared to hyperspherical prototype networks is mixing class prototypes and refined optimization objectives. This work proposes mixing data samples and the corresponding class prototypes in their respective spaces. In this case, the objective is to maximize the cosine similarity between a mixed sample and its corresponding mixed prototype. The other proposed objective is to minimize the cross-entropy between the dot product of a mixed sample with the original prototypes and the corresponding vector representing the proportion of each prototype in the mixture. The experiments show a performance improvement in the visual classification task compared to the baseline hyperspherical prototype networks for both optimization objectives. In addition, the results presented in this work outperform those given in the original hyperspherical prototype networks paper. Another key finding is that, when mixing is used, maximizing similarities as an optimization objective results in better accuracy and much better intraclass clustering of embeddings around their respective prototypes than cross-entropy.