Fig. 2
From: Visualizing complex feature interactions and feature sharing in genomic deep neural networks
Illustration of the generation of OFIV from FIVs generated by all 10 runs of gradient ascent in synthetic data set I. Red circles on the X-axis represent positive channels and blue circles represent negative channels. Circle size is proportional to the absolute FIV value. The weighted variance (IL score) of each channel is plotted below the FIVs, where the darkness and circle size is proportional to the variance. The OFIV is visualized below, where the circle size reflect the overall importance score of a channel. The channels that are predicted as non-additive by the Gaussian Mixture Model fitted on the IL scores are labeled by a star. A seqlogo visualization of the filter weight is plotted next to the corresponding channel. Filter {a,f} and {c,d} which capture sequences that involve in XOR logic are correctly predicted as non-additive. Among the remaining filters, the top-OFIV ones {b,c,g} which capture the sequence that involve in AND logic are correctly predicted as additive