Modular representation of autoencoder networks
Chihiro Watanabe, Kaoru Hiramatsu, Kunio Kashino · 2017
An autoencoder (AE) is a nonlinear extension of principal component analysis (PCA). It can extract abstract information about input data with low dimensions by combining multiple dimensions of input data through a layered neural network. A trained AE network can be used in various applications like parameter initialization for another inference; however, there is still no method for interpreting the feature values provided by an AE. Here we propose applying the method for extracting modular representations of layered neural networks to AE networks. We show in experiments using both synthetic and practical datasets that this approach can properly decompose a trained AE network and provide clues for discovering knowledge of feature values extracted in a hidden layer of an AE network.