The 5×5 grid is a retina; every pixel is connected to all three output neurons (+, ×, ○). A connection's strength is drawn as a line — green excitatory, red inhibitory, thickness = strength. Classifying is one step: each output neuron sums its weighted inputs, loudest wins. Learning is one rule: when the network is wrong, strengthen the connections from active pixels to the correct neuron, weaken them to the one that answered. Rosenblatt proved this converges whenever a correct weighting exists.
Watch the weight maps on the right while it trains: each neuron's weights become a template of its shape — a receptive field, much like those Hubel and Wiesel were recording from visual cortex in the same years. Nobody draws the template; it precipitates out of corrections. After training, draw on the retina yourself — a sloppy ×, a half ○ — and see whether the network generalizes.
And its limits are as instructive as its powers: a single layer can only learn categories a straight boundary can separate (Minsky & Papert, 1969 — the critique that froze the field for a decade). The fix — stacking perceptrons and propagating the error backwards — is, at heart, the deep learning behind today's AI.