Fig. 1
From: Visualizing complex feature interactions and feature sharing in genomic deep neural networks
Illustration of DeepResolve’s working flow. a Feature Importance Vectors calculation. After a network is trained and a intermediate layer is selected, DeepResolve first computes the feature importance maps (FIM) of each of the channels using gradient ascent. Then for each channel, the Feature Importance Vector (FIV) score is calculated as the spatial average of its FIM scores. b Overall Feature Importance Vector calculation. For each class, DeepResolve repeats the FIV calculation T times with different random initializations. The weighted variance over the T times is then calculated as an indicator of inconsistency level (IL) of each channel. A Gaussian Mixture Model is trained on IL scores to determine the non-additiveness of a channel. For each channel, the T FIVs are combined with the reference to the inconsistency level to generate an Overall Feature Importance Vector (OFIV) which summarizes all ‘favored’ and ‘unfavored’ patterns of a class. Finally, we use the non-negative OFIVs of each class to analyze class similarity and the OFIVs to analyze class differences