Fig. 15: Illustration of the interpretable CNN designed by Zhang et al.118. | npj Computational Materials

Fig. 15: Illustration of the interpretable CNN designed by Zhang et al.118.

From: Explainable machine learning in materials science

Fig. 15

The top four rows show example filer feature maps from the interpretable CNN and the bottom two rows show example filter feature maps from an ordinary CNN. Note that feature maps from the interpretable CNN filters tend to be meaningful (animal heads in this example) while feature maps from ordinary CNN filters are usually meaningless. These filter feature maps are computed from filter receptive fields following a technique proposed by Zhou et al.161. Figure reprinted from ref. 118 with permission. Copyright 2018 IEEE.

Back to article page