feature map

A feature map is the picture a filter produces after it has scanned an image — a kind of heat map showing where one particular pattern was found. If a filter is tuned to detect slanting edges, its feature map is bright along every slanting edge in the photo and dark everywhere else. So while the input is a picture of pixels, a feature map is a picture of presence: it answers, at each spot, "how strongly does my pattern show up here?"

Inside a convolutional network this happens in parallel. A layer doesn't apply one filter but many — dozens or hundreds — and each produces its own feature map: one for vertical edges, one for a reddish blob, one for a fur-like texture, and so on. Stacked together, these maps form the layer's output, and the next layer treats that stack as its new input, building maps of maps. Early layers hold feature maps of simple things (edges, colors); deeper layers hold maps of complex things (eyes, wheels). The word "map" is literal: each one preserves the layout of the image, so a bright spot in a feature map points to roughly where in the original photo that feature lives.

Feature maps are how a deep network's thinking becomes visible. You can render them as grayscale images and actually see what a layer is reacting to — a rare window into a system that is otherwise a wall of numbers. Be careful with the romance, though: deeper feature maps respond to abstract mixtures that rarely line up with a single human word, and reading them is interpretation, not proof. They show what the network attends to, not why, and not whether it is right.