How a Machine
Learns to See.
Upload a photo and watch a real image-recognition model take it apart — pixel by pixel, layer by layer. No math, no jargon. Just you and a machine that has only ever learned two things: dogs, and cats.
A photo is just numbers.
Before a machine can tell a dog from a cat, the picture has to become something it can do arithmetic on.
This is all it gets.
One small photo, shrunk to a fixed 150×150 — the size this model reads. No labels, no context. Just colour.
Really, it’s a grid.
Zoom in and the photo is just a lattice of 22,500 little squares — 150 across, 150 down. Each square is one pixel.
And each square is three numbers.
How much red, green and blue it holds — each 0 to 255. Hover the image to read any pixel.
That’s the whole photo to the machine: 67,500 numbers. It has no idea what a “dog” is. Everything from here is patterns in these numbers.
The machine looks for patterns.
It passes your photo through a stack of layers — each hunting for something a little more complex than the last.
Start here again.
Same photo, same grid of numbers. Now we look inside the machine as it flows through.
It starts with edges.
The first layer is almost dumb. It only reacts to sharp boundaries — light meeting dark. Watch it trace the outline and ignore flat areas.
Edges become textures.
A few layers later, those edges combine into texture — fur, grass, the grain of things. The map gets coarser: each point now summarises a whole patch.
Textures become bigger patterns.
Deeper again, the patterns grow larger and more abstract — textures combining across whole regions of the photo. Push much deeper than this and they sharpen into actual parts, like an ear or a snout. Nobody ever defined those; the machine just learned which shapes predict “dog.”
No single layer understands anything. The understanding is in the stacking — simple patterns becoming complex ones, dozens deep, until the last layer can just say which it is.
Where is it looking?
The machine commits to an answer — but which part of the photo made up its mind? Let’s find out by hiding regions and watching its confidence wobble.
Scroll into view to run the experiment.
How did it learn?
It wasn’t programmed with rules about whiskers. It was shown 20,000 labelled photos, over and over. Here is its actual report card — scroll to read it.
At first, it learns fast.
Each step is one epoch — a full pass through all 20,000 photos. Early on, every pass makes it noticeably better.
Then it flattens out.
Around 96% the line goes flat. More studying stops helping. The honest score — on photos it has never seen — is the only one that counts, and it’s stuck.
So it changes tactics.
At the dashed line it gently unfreezes the deeper, borrowed vision machinery and lets it adjust a little. That’s fine-tuning — and it buys the last hard percent, up to 97%.
Then it stops itself.
It was allowed 20 more epochs but quit after 8 — because the honest score stopped improving. Training longer wouldn’t make it smarter, only start memorising.
During practice the machine wears a handicap: cropped, flipped, distorted copies, with part of its own brain switched off at random. That makes practice harder than the real test on purpose, so it can’t just memorise. Take the handicap off and it does even better.
Break it.
It will answer every question with total confidence — even one that makes no sense. It has only ever learned two words. So show it something that is neither.
It has exactly two words in its vocabulary: “dog” and “cat.” It physically cannot answer anything else.
You just watched a machine fail with total confidence.
Confidence is not correctness. A model can be 99% sure and dead wrong — it measures “how well does this match what I’ve seen,” not “is this true.”
A model only knows the world it was trained on. This one saw dogs and cats. The big ones saw more — but outside their training, they don’t know what they don’t know.
“I don’t know” has to be designed in. Nothing here ever says “that’s not a pet,” because no one built that option. If you ship AI, the question isn’t “how accurate is it?” — it’s “what does it do when handed something it was never trained for?”
From pixels, to patterns, to a confident mistake — that whole arc is the same one running inside the much larger AI systems you use every day. Now you’ve seen it happen, end to end, with your own photo.