ALIGNFLOW LEARNING FROM MULTIPLE DOMAINS VIA NORMALIZING FLOWS

objective: learn a joint distribution over two domains

• CycleGAN
  • indirect constraint of joint distribution
  • careful design of loss functions
• a single latent variable shared by two domains
  • forward mapping implemented as a normalizing flow: cycle consistant in nature
  • also can be trained by MLE objective
    • only trained by MLE fails to identify optimal mapping: any volume-preserving transformation in latent space will achieve the same L_{MLE}