Article Source
Exploring And Attacking Neural Networks With Activation Atlases
Abstract
By using feature inversion to visualize millions of activations from an image classification network, we create an explorable activation atlas of features the network has learned which can reveal how the network typically represents some concepts.