Fig. 3: Example heat map explanations for a CNN with image inputs.
The task is to predict the ionic conductivity of a ceramic material from its image quality maps. a CNN model architecture. The last convolution layer is connected with the first fully connected layer via global average pooling (GAP), which allows the tracking of implicit attentive response weights from the top fully connected layers to pixel locations on the original image. b Heat map masked input images. The blue and red regions hide image features that were ignored by the CNN when predicting low and high ionic conductivities. Figure reprinted from ref. 21 with permission. Copyright 2017 Elsevier.
