Multi-granular Representation Learning Using Multiple Clustered Features

Abstract

In this paper, we focus on the granularity that indicates the level of fineness of features obtained from data. We define the granularity of features as the number of clusters formed by a batch of those feature vectors, and analyze the effect of granularity distribution of features on machine learning tasks. For this purpose, we propose a new neural network architecture and the regularization method that can control the granularity of features explicitly. The proposed network has a branched structure inside and learns multiple feature subvectors regularized to be evenly divided into the specified number of clusters. In experiments with classification task, we achieved higher accuracy than the usual one when we learn feature vectors in an unsupervised manner. We also confirmed that the existence ratio of coarse-grained features leads to better representation for classification rather than the multi-granularity property.