Learning Disentangled Joint Continuous and Discrete Representations

goal: disentanglement on continuous and discrete variables

• revisit beta-VAE: larger beta leads to efficient and disentangled representations
  • KL as a upper bound of mutual information
  • penalizing KL improves disentanglement at the cost of reconstruction error
  • controlled information capacity -> force KL to be C, which increases during training
• approx. of discrete variables
  • Gumbel: differentiable relaxation
• diff. C for continuous and discrete variables
  • avoid model assigning all capacity to continuous var.s