Grad-CAM as a function of training time. A single batch means that the network has trained on 240 spectra from the artificial dataset. Initially the ConvNet focuses randomly on the left wing of the k-line core, before pooling all of its attention onto the artifact, resulting in a classification accuracy of 100%. Grad-CAM allows us to deduce that the most important feature for the network’s decision making was exclusively the artifact, which by design was the only way to distinguish the two classes of spectra.