DisUnknown Distilling Unknown Factors for Disentanglement Learning

not all factors are labeled

• assume all other factors labeled and this single factor and all other task-irrlevent factors encapsulate into one factor
• two stage training
  • extract unknown factor disentangledly
    • gt labels + predicted unknown -> reconstruct image
    • random labels + predicted unknown -> cannot reconstruct gt labels
      • s.t. no informations are leaked!!!
  • train a conditional GAN: embed labeled factor into the latent space
    • s.t. random factor is preserved, images are photo-realistic, gt labels are preserved, good!