A Search for Visual Features in Adversarial Networks

Abstract:
Scarcity and imbalance of the training data usually hinder adversarial networks to have proper training and thus make it difficult for the framework to learn representations with adequate quality. This is especially true when rare instances look kind of exotic compared to the rest of the instances in the dataset. In this work, we investigate approaches for leveraging relevant knowledge from a different dataset to improve the model performance and demonstrate that auxiliary data enables the networks to learn underrepresented features.