If you look at the 'pictures' inside a CNN layer, and some parts are very bright, what does that tell you about what the network found in the image?

The 'pictures' inside a CNN layer are technically called feature maps, also known as activation maps. Each feature map is the output produced by a specific convolutional filter, or kernel, after it has processed the input image or a feature map from a preceding layer. A very bright part within one of these feature maps indicates a high activation value at that specific location. This high activation signifies that the particular visual pattern or feature which that unique filter was trained to detect was strongly identified and present in the corresponding region of the input image. For instance, filters in earlier layers often learn to detect basic, low-level features like edges (e.g., horizontal, vertical, diagonal lines) or specific textures. If a bright spot appears in a feature map generated by an edge-detecting filter, it means a prominent edge of that specific orientation was found at that exact location in the image. As the network progresses to deeper layers, filters combine these simpler features to recognize more complex, high-level patterns, such as corners, curves, or even parts of objects like an 'eye' or a 'wheel'. Therefore, a bright spot in a deeper layer's feature map reveals that the network strongly detected a more intricate and abstract feature at that particular position within the image.