Interpretability of deep models

Excitation Backprop

Visualization method that propagates activations from the output layer to the input using a modified version of backpropagation, where weights are replaced by conditional probabilities to better capture excitation paths in the network.

← Geri