FVAE: a regularized variational autoencoder using the Fisher criterion

Jie Lai,Xiaodan Wang,Qian Xiang,Rui Li,Yafei Song
DOI: https://doi.org/10.1007/s10489-022-03422-6
IF: 5.3
2022-03-26
Applied Intelligence
Abstract:As a deep generative model, the variational autoencoder (VAE) is widely applied to solve problems of insufficient samples and imbalanced labels. In the VAE, the distribution of latent variables affects the quality of the generated samples. To obtain discriminative latent variables and generated samples, a Fisher variational autoencoder (FVAE) based on the Fisher criterion is proposed in this study. The FVAE introduces the Fisher criterion into the VAE by adding the Fisher regularization term to the loss function, aiming to maximize the between-class distance and minimize the within-class distance of latent variables. Different from the unsupervised learning of the VAE, the FVAE requires class labels to calculate the Fisher regularization loss so that the learned latent variables and generated samples have sufficient category information to complete classification tasks. Experiments on benchmark datasets show that the learned latent variables of the FVAE are more discriminative and that the generated samples are more efficient in improving the performance of various classifiers than the VAE, β-variational autoencoder (β-VAE), conditional variational autoencoder (CVAE), Denoising variational autoencoder (DVAE) and information maximizing variational autoencoder (IM-VAE).
computer science, artificial intelligence
What problem does this paper attempt to address?