Dreamento Dev Path · Tiny Workshop
A tiny workshop about how a 1D CNN finds patterns in a brain signal
Robo is a friendly robot who has never seen a brain signal before. Help teach Robo how to find a spindle burst hiding inside the squiggle — one step at a time.
To you, this looks like a wiggly line. To Robo, it's just a list of numbers — one number for every tiny moment in time.
Right now, Robo has no idea what a spindle looks like. Before Robo can find one, it needs a stamp shaped exactly like what it's hunting for — then it slides that stamp along the signal, checking everywhere: "does it look like me right here?"
Robo's spindle stamp
Match right here:
0%
The strip below fills in as Robo scans — brighter means a stronger match at that spot.
That heat strip has dozens of numbers in it — more detail than Robo actually needs. Pooling is Robo zooming out: it chops the strip into a few big chunks and keeps only the strongest match from each chunk, like shrinking a whole book chapter down to one sentence.
Before (every position):
After (zoomed out to 6 chunks):
A real CNN doesn't stop at one stamp. Robo keeps a whole box of them — each one hunting for something different in the exact same signal.
Spindle stamp
K-complex stamp
Spindle stamp's heat strip:
K-complex stamp's heat strip:
Robo starts with a bad, lazy stamp — barely shaped like anything. Six real snippets, one at a time: Robo guesses, you reveal the truth, and if Robo's wrong, it nudges its stamp a little closer to right. Watch the stamp shape change as you go.
Round 1 of 6
Robo's guess
—
Scanning with a stamp, zooming out, and practising with feedback is exactly how a real 1D CNN learns to find spindles in Chapter 6 — except a real one tries thousands of tiny stamp shapes automatically, thousands of times, instead of just one shape nudged by hand like you just did.