Visualization using Grad-CAM
The original images of the native and hybrid species, and the activation
maps visualized by Grad-CAM, are shown in Figure 3 and Figure 4. Other
individuals used in the analysis are shown in Supplementary Figures 4
and 5. The red indicates the regions the model considers when outputting
the prediction results. The results show that the head spots of
salamanders are a key area for the classification of native and hybrid
species. For native species, the visualized figures indicated that the
network learns to recognize relatively distinctive and large black
spots. On the other hand, for hybrids, the visualization indicated that
the network learned to focus on pale and ambiguous wide regions rather
than distinctive and large black spots. However, because the spot
pattern differs from individual to individual, the results of Grad-CAM
visualization also differ from individual to individual (Supplementary
Figures 4 and 5).