Understanding Invariance via Feedforward Inversion of Discriminatively Trained Classifiers
Deep image classifiers’ logits retain far more visual information than just class labels, enabling realistic image reconstruction via a GAN and revealing model invariances, robustness, and sensitiv...
