Automatic Discovery of Visual Circuits

Achyuta Rajaram, Neil Chowdhury, Antonio Torralba +2

To date, most discoveries of network subcomponents that implement human-interpretable computations in deep vision models have involved close study of single units and large amounts…