Deep Embedded Clustering with Asymmetric Residual Autoencoder

Abstract

Deep clustering methods have obtained excellent performance on clustering tasks with the benefit of feature representations learned with deep neural networks. Even though promising performance of deep clustering has been shown in different applications, the efficiency of the features achieved is limited by the symmetric structure of the autoencoders employed. Deeper autoencoder will lead to less reliable features extracted from the encoder due to the strong decoding capability of the symmetric deep decoder. To address this issue, a novel Asymmetric Deep Residual Embedded Clustering algorithm is proposed in this paper. Specifically, an asymmetric residual deep autoencoder is constructed to learn the features embedded in high dimensional data. The asymmetric residual autoencoder uses residual connection to enhance the feature extraction ability of the encoder with deeper network, while a shallow CNN is adopted as the decoder. This arrangement could make the feature representation ability of the encoder stronger than decoder's reconstruction ability, which ensures the reliability of the extracted features. In addition, a clustering layer has been incorporated to form an end to end solution. Experiments on benchmark datasets have shown the effectiveness of the proposed method.