Extraction of easily interpretable representation using five-layered autoencoder

In Deep Learning, which has become a topic of intense study in recent years, an auto encoder that learns an identity map by a neural network plays an important role to extract features of data. One disadvantage of this method is that it is not always possible to extract appropriate features in the middle layer. In this study, a five-layer auto-encoder based on a sensory integration model proposed by Fukumura et.al is examined. In order to gain easily interpretable and symbolic expressions for pattern recognition, the outputs of neurons in the middle layer are forced to take a binary value and parts of these outputs are forced to take the same value when the input data belongs to the same category. As a result of computer simulation focused on simple pattern recognition, it was confirmed that it is possible to obtain the desired expression in the middle layer.