Deep Learning Important Features (DeepLIFT)

DeepLIFT (Deep Learning Important FeaTures) is a method for attributing feature importance to the input features of a deep neural network. DeepLIFT works by comparing the activations of each neuron in the network under the actual input with the activations that would have been produced by a reference input, and then using the difference to attribute importance scores to the input features.

​

The reference input is typically chosen to be a neutral input that represents a baseline or background state. For example, in image classification tasks, the reference input might be an image with all pixels set to a neutral gray color. The idea behind DeepLIFT is to measure how much each input feature deviates from the reference input, and then attribute importance scores to the input features based on the extent of the deviation.

​

To compute the importance scores, DeepLIFT first computes the difference between the activations of each neuron under the actual input and the reference input. This difference is then multiplied by a set of weights that reflect the contribution of each input feature to the neuron's activation, yielding a set of importance scores that indicate the contribution of each input feature to the overall output of the network.

​

DeepLIFT has been used in a variety of applications, including image classification, natural language processing, and genomics, and has been shown to be effective in identifying important input features and detecting potential sources of bias or error in deep neural networks.