DeepLIFT: Learning Important Features Through Propagating Activation Differences
DeepLIFT is a method for assigning feature importance that compares a neuron's activation to its'reference', where the reference is the activation that the neuron has when the network is provided a'reference input'.
BibTex:
