Bilateral discriminative autoencoder model orienting co-representation learning

Abstract

Autoencoder is an important representation learning model which has attracted extensive research attention. However, an autoencoder learns latent representation by reducing reconstruction error without emphasis on discrimination, which is vital to downstream machine learning tasks like classification and clustering. Many existing works have improved the discrimination of autoencoders. But as far as we know, there is no work focusing on bilateral discriminative representation learning(i.e. co-representation learning). Our work unlocks the potential of autoencoder on co-representation learning and proposes a bilateral discriminative autoencoder model for co-representation learning(CRBDAE). By utilizing a fuzzy set, the topological relationship between samples and features is represented as fuzzy information. In the bilateral discriminative autoencoder, by means of regularization, fuzzy information is employed to enhance the self-supervised co-representation learning ability. Thus, the corresponding loss function is illustrated. We also inferred the parameters updating method and proposed the model training algorithm. Finally, the availability of the CRBDAE model was demonstrated on 12 datasets and the results proved that the performance of the proposed model meets our expectations.