RTC-VAE: HARNESSING THE PECULIARITY OF TOTAL CORRELATION IN LEARNING DISENTANGLED REPRESENTATIONS
In the problem of unsupervised learning of disentangled representations, one of the promising methods is to penalize the total correlation of sampled latent vari-ables. Unfortunately, this well-motivated strategy often fail to achieve disentanglement due to a problematic difference between the sampled latent representation and its corresponding mean representation. We provide a theoretical explanation that low total correlation of sample distribution cannot guarantee low total correlation of the mean representation. We prove that for the mean representation of arbitrarily high total correlation, there exist distributions of latent variables of abounded total correlation. However, we still believe that total correlation could be a key to the disentanglement of unsupervised representative learning, and we propose a remedy, RTC-VAE, which rectifies the total correlation penalty. Experiments show that our model has a more reasonable distribution of the mean representation compared with baseline models, e.g.,β-TCVAE and FactorVAE.
PDF Abstract
dSprites
