The network's first layer has 32 learned filters, each 3×3 pixels × RGB. Nobody designed them; they come out of training. Click a filter to see where it fires in the photo: some respond to edges in one direction, others to a colour. At only 3×3 pixels they look less like the famous Gabor-like filters of AlexNet (11×11), but their responses tell the same story.
The same photo as the network sees it at increasing depth. Each tile is one channel; stronger colour = stronger response. Early maps are sharp and look like edge images; deeper maps get coarser and respond to whole regions (fur, wheels, a face) instead of pixels.
A class activation map (CAM) shows which regions pushed the score of a class up. Bright = evidence for that class. This is how you catch shortcut learning: does "dog" light up the dog, or the grass behind it? Pick one of the top classes: