The network is a two-layer perceptron: your 28×28 drawing feeds 64 hidden neurons, which feed 10 output neurons — 51,000 learned connections, all of them running in your browser. The bars show how strongly each digit-neuron responds.
The heatmap is an attribution map (gradient × input): for the chosen digit, it asks every pixel "if you were a little darker, would the network believe in this digit more, or less?" Green pixels are evidence for the digit, red pixels are evidence against it. This is the same family of methods our SHERPA framework applies to EEG data: opening the black box by tracing a decision back to its inputs, in space and time.
Things to try: draw a 7, then click the bar for "1" — the red pixels show exactly which stroke rules the 1 out (usually the crossbar). Draw an ambiguous 4/9 and watch the evidence split. Note also what attribution does not tell you: it explains this one decision locally, not how the network works in general — a caveat that applies to every saliency method, in machine learning as in neuroimaging.