LOGAN-B: Bidirectional Latent Optimized Generative Adversarial Networks

The inductive bias of generative adversarial networks is a powerful tool for learning distributions of data. However, these models are difficult to train and very susceptible to mode collapse, capturing only a few modes of the data distribution. In this work, we introduce bidirectional latent optimized generative adversarial networks as a framework for training generative adversarial networks to mitigate this problem by encouraging the generator to maintain a mapping to the data distribution.

[Report]