Leaky ReLUs That Differ in Forward and Backward Pass Facilitate Activation Maximization in Deep Neural Networks.

Christoph Linse, Erhardt Barth, Thomas Martinetz

Browse the full IJCNN paper archive.