Interactive module
Convolutional Neural Network
A single conv filter learns to tell horizontal stripes from vertical stripes. Watch the kernel, feature map, and pooled output change in real time.
How a CNN works
- Kernel (filter) — a small 3×3 grid of weights that slides across the image computing dot products. It learns to “light up” for specific patterns.
- Feature map — the 6×6 grid of responses. Bright cells = the kernel fired strongly there.
- Max pooling — shrinks the 6×6 feature map to 3×3 by keeping only the strongest response in each 2×2 block. White-bordered cells show the winners.
- Dense layer — 9 weights combine the pooled values into a single score. Sigmoid squashes it to a probability of class 1 (vertical).
Press Train to start, or Step to advance one sample at a time.
step0samples seen
loss—lower = better
ŷ—≥0.5 → vertical
test loss—held-out sample
test acc—running average