The Activation Atlas is a method for representing processes in an artificial neural network during image recognition.
The visualization shows a collection of visual data that the network uses to recognize things. This means the operating principle of individual algorithms can be observed. It not only shows abstract forms, colors, and patterns that are recognized by the program, but also how these elements are combined to identify certain objects, animals, and scenes.
Based on the study „Exploring Neural Networks with Activation Atlases“ by Shan Carter, Zan Armstrong, Ludwig Schubert, Ian Johnson, Chris Olah