mcVAE: disentangling by mean constraint

Ming-fei Hu

doi:10.1007/s00371-023-02843-9

2023-04-06

PRE

DOI：10.1007/s00371-023-02843-9

原文链接

原文求助

摘要

En 中文

Disentanglement tends to automatically learn and separate the interpretable factors of variation hidden in the data. Disentangled representations are more transferable and robust for the chosen model, and they are commonly used in image attack detection and anti-fraud, as well as classification and recommendation systems in special situations. As a popular method for learning unsupervised disentanglement, beta-VAE re-weights the KL divergence by an adjustable hyperparameter. However, good disentangled representations always lead to blurry reconstructions and mode collapse on complex datasets. We find that the variance vector of the variational posterior is related to the nature of the dataset and representation space, limiting its value to 1 is not reasonable enough. More importantly, constraining mean variable alone can achieve better disentanglement and reconstruction performance. Therefore, we introduce mean constraint VAE, a simple and effective replacement of the beta-VAE for improving the poor reconstruction and learning a higher degree of disentanglement. In addition, a classifier-free measure of disentanglement called variance proportion metric is proposed. Experiments show that our framework outperforms beta-VAE on several benchmark datasets.

Keyword:

Variational autoencoder

Disentanglement

Representation learning