The story behind this piece
Rete Neurale is a real forward pass. A 28 × 28 handwritten seven is flattened into 784 inputs and pushed through three ReLU layers of 256, 128 and 64 units to a ten-way softmax: 242,762 parameters in all. The weights are He-initialised, and the output layer is fitted to this image by gradient descent, so the network answers "7" with about 93% confidence. Amber lines are positive weights and blue negative; their strength is the weight's size. The brightest paths are the largest contributions, weight times activation, that actually flowed on this pass. Dark rings are units that ReLU switched off. Dots mark the units too many to draw, and every layer width and equation is labelled.


