TL;DR
- task : image generation
- problem : posterior collapse in generation model
- idea : discrete latent variable (idea from vector quantization)
- architecture : #45 with codebook(find nearest embedding vector) -> need copying gradient!
- objective : reconstruction error + embedding loss w.r.t. reconstruction error + commitment loss(to train embedding + encoder in similar pace)
- baseline : VAE, VIMCO
- data : CIFAR10
- result : qualitatively good!
- contribution : VAE with discrete latent vector