VAE Based Disentanglement Learning by Dimension-Wise Constraints to Latent Variables

Disentanglement representation learning usually refers to the separation of factors that vary within the data, which are often physically meaningful and independent features that are important in real-world scenarios. β-VAE, as the basic model for unsupervised disentanglement, employs a mechanism to re-weight the KL divergence. However, it is often faced with the problem of how to adjust β to trade off reconstruction fidelity against disentanglement. We find that disentanglement appears to be embodied in individual dimensions of latent features that contain more information than others. Therefore, We propose Dimensionally Constrained VAE (DCVAE), which is a simple alternative to β-VAE for improving reconstruction quality and learning to achieve higher disentanglement scores. The effectiveness of this model is verified on the mainstream dSprites and CelebA datasets.