GAN Dissection: Visualizing and Understanding Generative Adversarial Networks

Generative Adversarial Networks (GANs) have recently achieved impressive for many real-world applications, and many GAN variants have emerged improvements in sample quality and training stability. However, they have been well visualized or understood. How does a GAN represent our visual internally? What causes the artifacts in GAN results? How do choices affect GAN learning? Answering such questions could us to develop new insights and better models. In this work, we present an analytic framework to visualize and understand at the unit-, object-, and scene-level. We first identify a group of units that are closely related to object concepts using a-based network dissection method. Then, we quantify the causal of interpretable units by measuring the ability of interventions to objects in the output. We examine the contextual relationship between units and their surroundings by inserting the discovered object concepts new images. We show several practical applications enabled by our, from comparing internal representations across different layers, , and datasets, to improving GANs by locating and removing-causing units, to interactively manipulating objects in a scene. We open source interpretation tools to help researchers and practitioners understand their GAN models.