Interpreting Galaxy Deblender GAN from the Discriminator's Perspective

Abstract: Generative adversarial networks (GANs) are well known for their unsupervised learning capabilities. A recent success in the field of astronomy is deblending two overlapping galaxy images via a branched GAN model. However, it remains a significant challenge to comprehend how the network works, which is particularly difficult for non-expert users. This research focuses on behaviors of one of the network's major components, the Discriminator, which plays a vital role but is often overlooked, Specifically, we enhance the Layer-wise Relevance Propagation (LRP) scheme to generate a heatmap-based visualization. We call this technique Polarized-LRP and it consists of two parts i.e. positive contribution heatmaps for ground truth images and negative contribution heatmaps for generated images. Using the Galaxy Zoo dataset we demonstrate that our method clearly reveals attention areas of the Discriminator when differentiating generated galaxy images from ground truth images. To connect the Discriminator's impact on the Generator, we visualize the gradual changes of the Generator across the training process. An interesting result we have achieved there is the detection of a problematic data augmentation procedure that would else have remained hidden. We find that our proposed method serves as a useful visual analytical tool for a deeper understanding of GAN models.