Key points
- A feature map is a picture made by taking the scores a small tile gave spot by spot and laying them out in their original arrangement.
- A spot with a big value is where the pattern that tile hunts for showed up clearly; a small value means it barely showed up at all.
- With several tiles, you get that many feature maps — one layer hands over a whole bundle of pictures, not just one.
- Climb the layers and the pictures get smaller while the count grows. Position blurs, meaning thickens.
- The next layer doesn't scan the original photo — it scans the feature map the layer before it made.
The analogy
Wind turbines stand in neat rows across a hillside. On the control room screen, a grid is drawn in exactly the layout the turbines stand in, and each square shows a number for how fast that turbine is spinning right now. Where a ridge blocks the wind, the number in that row drops off; where wind comes straight through, the number climbs.
Just jotting the numbers down on a sheet of paper tells you nothing. But put them in a grid that matches where the turbines actually stand, and the line where the wind gets blocked shows up as clearly as a picture, cutting right across the hillside. What the operator does isn't reading numbers one by one — it's reading the pattern on the screen.
The screen isn't just one, either. Put up screens side by side for the same moment, each measuring something different — rotation speed, vibration, temperature — and you end up with several layers stacked over the same hillside. Feature maps stack up the same way — several of them, each still holding its position.
In detail
Holding a position is the whole point
The key to understanding a feature map isn't "what is the value" — it's "where is the value." Scanning produces one score per spot on the photo, and if those scores get dumped into a sack with no regard for where they came from, they're useless. Laying each score onto a grid at exactly the spot it came from is what turns it into a picture.
That's why a feature map ends up similar in shape to the original photo. Its width and height stay the same or shrink depending on how far the tile steps and how edges are handled, but the top left is still the top left and the bottom right is still the bottom right. If an object sat on the right side of the photo, the spots with big values cluster on the right side of the feature map too.
This property is exactly what makes it possible to figure out where an object is. Boxing every object or cutting things out pixel by pixel works by tracing back the position that's kept intact in a feature map. Putting a single label on a whole photo usually merges the positions into one at the end and gathers only what's there.
It's a bundle, not a single sheet
A layer holds anywhere from dozens to hundreds of tiles, and since each one scans separately, you get that many results — a picture from the tile hunting for vertical lines, a picture from the tile hunting for a sudden color change, stacked up side by side.
Looking at the same spot across several of these pictures at once tells you something about that spot's character. A spot that's high in both the vertical-line picture and the horizontal-line picture is likely where two lines meet — a corner. The next layer uses exactly this: scanning with a thick tile that spans all the pictures at once, it looks at which pictures are lit up together and recognizes bigger parts from that.
It gets smaller and thicker going up
As it moves through layers, a feature map's shape changes in a predictable way. Every few layers, at the spots where pooling or a bigger stride comes in, the width and height are halved and the count is doubled — a common design. The layers in between often leave both size and count as they were. What starts wide and thin ends up small and thick.
What this shift means is clear enough. One square in an early feature map points to a fingernail-sized spot on the original photo, but one square in a later feature map summarizes a whole wide region of the photo at once. Position gets blurred, and in its place sits the meaning of "what was around here." By the last layer, only a few squares remain in width and height, and one square all but stands in for the entire photo.
The early ones read straight off by eye
Pull a feature map out and look at it as a picture, and the first layer is easy to make sense of — an outline-only photo, a photo where lines running one direction stand out — the traces of the original are still there. You can more or less guess what it caught, just by eye.
Further back, things change. It looks like a blotchy small grid, and it's not obvious at a glance why any particular square lit up, because there are few squares and the meaning is scattered across many pictures. Some pictures do respond to things you could name, like a face or a wheel, but finding them takes separate analysis. That's why people trying to see what a model based its judgment on don't look at the raw feature map — instead they overlay, in color, which spots contributed most to the conclusion, right on top of the original photo.
More precisely
A feature map is the output of a convolution layer. It's a block of values with three directions — width, height, and count — where the count matches the number of filters in that layer. Each value is the number left after adding a reference value to a sum of products and passing it through an activation function. Because it's common to use an activation function that pushes negative values down to zero, a real feature map has a lot of squares sitting at exactly zero.
The control-room analogy breaks down in one place. A number on the screen comes from one turbine, but a single square in a feature map isn't the result of one point on the original photo — it's the joint result of that point and the points around it. The deeper the layer, the wider the range of the original photo one square ends up covering, until a late square corresponds to a large chunk of the photo.
And a control screen is built for a person to read, but a feature map is a mid-way result built for the next layer to use. It can be drawn out as a picture, but it isn't laid out for a person to read comfortably.
Try it yourself
Common misconceptions
It's easy to think a feature map is a blurred version of the original photo, but actually it isn't a photo at all — it's a table of scores laid out spot by spot.
It's easy to assume one feature map holds one object, but actually a single one is just the response to one pattern, and an object shows up scattered across several of them.
It's easy to think the number of pictures always stays the same going up the layers, but actually each layer's tile count is set separately. Within one stage it often stays put, then grows when the pictures shrink.
One-line summary
A feature map is a picture that lays the scores a scan gave, spot by spot, out in their original arrangement, and climbing the layers it gets smaller and thicker as it trades position for meaning.
Spotted an error or have a better analogy? Suggest an edit · Last updated 2026-10-17