Cycle-Consistent Adversarial Learning as Approximate Bayesian Inference
This provides a theoretical foundation for unpaired domain translation methods, connecting them to Bayesian inference, which is incremental but clarifies existing approaches.
The paper tackles the problem of learning interdomain correspondences without paired data by framing it as Bayesian inference in a latent variable model, and shows that cycle-consistent adversarial learning (CYCLEGAN) can be derived as a special case of their variational inference framework.
We formalize the problem of learning interdomain correspondences in the absence of paired data as Bayesian inference in a latent variable model (LVM), where one seeks the underlying hidden representations of entities from one domain as entities from the other domain. First, we introduce implicit latent variable models, where the prior over hidden representations can be specified flexibly as an implicit distribution. Next, we develop a new variational inference (VI) algorithm for this model based on minimization of the symmetric Kullback-Leibler (KL) divergence between a variational joint and the exact joint distribution. Lastly, we demonstrate that the state-of-the-art cycle-consistent adversarial learning (CYCLEGAN) models can be derived as a special case within our proposed VI framework, thus establishing its connection to approximate Bayesian inference methods.