A 3-layer CNN trains on CIFAR-10 in your browser. Watch Conv 1 filters evolve from noise into edge and colour detectors, then step through any test image to see what each layer responds to.

Loading CIFAR-10 data…

Training loss

Conv 1 filters — 16 learned RGB detectors

Random noise at init → oriented edges and colour blobs after ~2 k steps.

Layer-by-layer activations

—
Input
32×32 RGB
After Conv 1
16 maps · 32×32
lines & edges
After Conv 2
20 maps · 16×16
textures & curves
After Conv 3
20 maps · 8×8
object parts