CNN Lab
Draw a digit. A convolutional network sees the strokes, slides small filters across the image, and guesses which digit it is. The first-layer feature maps show which filters lit up on your drawing.
Network architecture
A picture (28×28 grayscale) flows left → right. Convolutions find features (edges → shapes), pooling shrinks the map so the network cares about the pattern, not the exact pixel, and the two dense layers vote on which digit it saw. About 12 000 trainable parameters — tiny by modern standards.
Draw a digit (0 – 9)
Draw with your mouse or finger. Try a 3, 5, or 9.
Prediction
Draw something and press Predict.
Feature maps
Bright pixels = the filter fired there. Layer 1 filters spot simple things (edges, curves). Layer 2 filters take Layer 1's responses as their input, so they combine those simple things into shapes — the classic "shallow → deep" progression.