Feature maps and filters
One filter asks one question. A layer asks dozens.
Last lesson you slid one filter and produced one feature map: a spatial answer to a single question ("vertical edge here?"). But one question is a pathetic description of an image. A real convolutional layer runs MANY filters over the same input, 32, 64, 256, each with its own learned weights, each asking a different question, each producing its own map.
The output of a conv layer is therefore a STACK of feature maps: same width and height, one map per filter. Think of it as the image re-described, no longer "brightness at each pixel" but "vertical-edge-ness here, horizontal-edge-ness here, green-blob-ness here…", a translation from pixels into evidence.
And here's the move that creates the hierarchy from the why-deep lesson: the NEXT conv layer convolves over that stack. Its filters see all the maps at once, so they detect combinations of answers, 'vertical edge here AND horizontal edge just below' is a corner. Corners combine into shapes, shapes into parts. Each layer asks questions about the previous layer's answers, and depth compounds.
Same image, three different questions
One 8×8 image: a solid square and a horizontal bar. Switch between filters and watch the feature map transform, bright cells mean "my pattern is here". Same pixels, completely different maps.
Fires on the left and right sides of the square, where brightness changes horizontally. The long bar barely interests it.
Nobody designs the filters: and what they become is famous
For decades, computer vision experts hand-crafted filters exactly like the ones you just used (the edge detectors are the classic Sobel filters). Convolutional networks ended that era: initialize the filters randomly, let backpropagation tune them against the loss, done. The network invents its own questions.
And what it invents is one of the most repeated results in the field: train a CNN on almost any large visual task and its first-layer filters converge to edge detectors at various angles and color-contrast blobs, remarkably similar to the receptive fields of neurons in V1, the first visual cortex area. The biological inspiration lesson from Fundamentals comes full circle: we borrowed a sketch of the brain, trained it with math, and it grew the same front-end vision hardware.
Deeper layers get harder to interpret, their filters operate on stacks of abstract maps, not pixels, but visualization tools confirm the hierarchy: motifs, textures, parts, objects. You already sorted these in the why-deep lesson; now you know the machinery that builds each floor.
Check your understanding
A convolutional layer has 32 filters. What does it output?
What to remember
- ✓A feature map is a spatial answer sheet: where does MY pattern appear? One filter produces one map.
- ✓A conv layer runs many filters in parallel → a stack of maps: the image translated from pixels into evidence.
- ✓The next layer convolves the stack, detecting combinations of answers, corners from edges, motifs from corners. That is the hierarchy, mechanized.
- ✓Filters are learned, not designed, and first layers reliably reinvent edge detectors resembling the brain’s V1.
- ✓Next: the maps are still huge, and exact pixel positions stopped mattering the moment detection happened. Time to shrink: pooling.