After training, click "Compute saliency" to see which pixels the network focuses on. We grey out 4×4 patches one at a time and measure how much each patch changes the top-class score — bright patches matter most.
Brighter = removing this patch hurts confidence more.