Feature extractor. Three identical blocks: a 3×3 convolution with padding 1 (keeps the spatial size), BatchNorm, ReLU, then 2×2 max-pooling that halves height and width. Each block doubles the channels (3 → 32 → 64 → 128) while the map shrinks 64 → 32 → 16 → 8, so the network trades resolution for feature richness.

Classifier. The final 128×8×8 map is flattened to 8,192 values and passed through a 256-unit dense layer with ReLU (Dropout 0.5 regularizes it during training only) and a last dense layer to 10 logits; Softmax turns those into class probabilities. Almost all of the parameters sit in the 8192×256 dense layer.

In the walkthrough, every one of these ops is a stage of its own, drawn with the real activations of the selected patch.

Choose an image patch, then press Run to watch the network classify it