Acquisition of Highly Independent Latent Space for Distribution Control Based on Features

Abstract

In machine learning, the quality of the acquired features significantly influences the accuracy of the target task. Recently, use of large-scale data has expanded in various fields. Large-scale data generally contains much information. Furthermore, machine learning makes it challenging to extract independent features from them. Furthermore, if the extracted features have a lot of redundancy, the learning results cannot obtain sufficient performance. This study reduces the redundancy of the acquired features and realizes intuitive mapping to latent space according to the information of the data. The proposed method has two learning steps. Initially, we create a feature extractor with sufficient performance for a particular data type. Then, the second step creates another feature extractor. Here, we learn that the features acquired by the extractor are independent of the features acquired in the first step. The experiment extracted features from the data based on the collaboration of two feature extractors. The first extractor extracts many common features, and the second extractor realizes feature extraction according to the individuality of the data. Then, we succeeded in mapping data groups with different characteristics to different regions of the latent space.