DILC-ESAE: Data-Info envelope stacked autoencoder on correlation among samples rather than themselves.

Abstract

Stacked autoencoders (SAEs) are a powerful and popular tool for structured classification. However, the existing AEs have a common shortcoming: they only focus on the original sample individual without considering the correlation information between similar samples. In fact, the information can significantly enhance the separability, thereby further improving the classification performance. To solve this problem, a new SAE that can effectively mine the correlation between samples in the modeling process is explored in this paper, inspired by knowledge pyramid and human recognition mechanism. The proposed Data-Info envelope stacked autoencoder (DILC-ESAE) models on the correlation among them, thereby improving the classification accuracy. It mainly consists of two parts. The first part is the Data-Info layer construction model (DILC) which is to extract local and global correlation information among samples, thereby constructing envelope samples. The second part consists of embedded stacked autoencoder (ESAE), which fuses the original samples during the training phase and within the network architecture, ultimately enhancing complementarity between deep features and original features in a superior fashion. The experimental results demonstrate that the proposed DILC-ESAE exhibits notably superior performance compared to existing, representative SAEs. Unlike existing SAEs, the major originality of this proposed DILC-ESAE is in that it can conduct deep feature extraction on the correlation among samples rather than themselves, thereby realizing hierarchical deep feature extraction. Moreover, the proposed DILC can be applied to other kinds of autoencoders even other deep neural networks, so has wide reference value. The data and codes can be found in https://github.com/acceptthisjj/DILC-ESAE.