In AlexNet, the convolutional layers learn features in a hierarchy, becoming more abstract the deeper the data travels. The earliest layers react to basic visual cues like edges or color contrasts. Deeper layers, however, begin to fire for complex shapes and class-specific patterns, such as parts of faces or objects. We can actually see this process by passing an image through the network and extracting activation maps from selected filters. Visualizing these maps shows which regions of the imag