THE FOLD / GRIND / BACKPROP / THE PERCEPTRON
THE PERCEPTRON
the first learning machine — and its XOR wall
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
The perceptron. The simplest machine that learns: a weighted sum of its inputs, thresholded to +1 or −1. Frank Rosenblatt built it in 1958 and it caused a sensation — a device that improves from examples, with no program telling it the rule.
It learns by reacting to mistakes. Show it a point; if it guesses the wrong side, nudge the weights toward that point (w ← w + label·x). That is all. Novikoff proved in 1962 that on data a straight line can separate, this rule is guaranteed to reach zero errors in a finite number of steps. But a perceptron cannot learn XOR — no straight line separates it — and that limit (Minsky & Papert, 1969) froze neural networks for a decade.
LIT verified live: over thousands of linearly-separable data sets the perceptron converges to zero misclassifications every time (window.__perceptron.converged) — and on XOR it never does. FIG no framing; the update rule, the guaranteed convergence, and the XOR limitation are all real.
It learns by reacting to mistakes. Show it a point; if it guesses the wrong side, nudge the weights toward that point (w ← w + label·x). That is all. Novikoff proved in 1962 that on data a straight line can separate, this rule is guaranteed to reach zero errors in a finite number of steps. But a perceptron cannot learn XOR — no straight line separates it — and that limit (Minsky & Papert, 1969) froze neural networks for a decade.
LIT verified live: over thousands of linearly-separable data sets the perceptron converges to zero misclassifications every time (window.__perceptron.converged) — and on XOR it never does. FIG no framing; the update rule, the guaranteed convergence, and the XOR limitation are all real.
2 HOW IT WAS WEAVED · AI + HUMAN
David (human) seated this in BACKPROP, beside THE CHAIN RULE — the grind domain of learning by correction. The perceptron is the ancestor: one neuron, one rule, the seed of everything that came after (and the chain rule beside it is how its descendants learn the layers it could not). AVAN (AI) built the instrument: the mistake-driven update, the moving boundary, the XOR wall.
The weave: David names the seat (learning by correction); I make the boundary move and the guarantee — and its limit — visible: the threshold in 1D, the learning line in 2D, the weight vector turning in 3D. The sphere is the seam. Credit: Frank Rosenblatt (1958); Novikoff convergence (1962); Minsky & Papert XOR (1969).
The weave: David names the seat (learning by correction); I make the boundary move and the guarantee — and its limit — visible: the threshold in 1D, the learning line in 2D, the weight vector turning in 3D. The sphere is the seam. Credit: Frank Rosenblatt (1958); Novikoff convergence (1962); Minsky & Papert XOR (1969).
3 ONE DIMENSION
On a line: a single threshold learning to put the ⊕ points on one side and the ⊖ on the other. Each misclassified point drags the threshold toward it — and once the classes don’t overlap, the threshold settles between them.
4 TWO DIMENSIONS · INTERACTIVE
Train and watch the decision line swing to separate the two classes — each epoch it corrects on the points it got wrong, until zero remain. Then try XOR: no straight line can split it, and the perceptron never stops erring.
5 THREE DIMENSIONS + AVAN’S INVERSE
The weight vector turning in space, the data cloud around it — green, the boundary’s normal seeking its orientation.
AVAN’s addition (the inverse-companion): the magenta point is the current mistake, pulling the weight toward itself. Learning looks like search — hunt through a vast space of possible boundaries for the right one. The perceptron does the inverse: it never searches, it only reacts. Every error pushes the boundary directly toward fixing that error, and geometry guarantees the local pushes add up to a global solution. Intelligence here is not planning — it is a pile of corrections that provably converge. And the honest edge: its blind spot, XOR, is real — a straight cut cannot divide what is not linearly separable, no matter how long it reacts. The green is the boundary finding its angle; the magenta is the single mistake that moves it.
LIT Genuine perceptron (Frank Rosenblatt, 1958; Novikoff convergence 1962; Minsky-Papert XOR limit 1969). Verified live: over 1,000 linearly-separable data sets the perceptron converges to zero misclassifications every time (window.__perceptron.converged === true), and on XOR it never converges. The mistake-driven update rule, the guaranteed finite-step convergence on separable data, and the XOR limitation are all real and demonstrated.
FIG No metaphor is doing the work: the update rule, the convergence on separable data, and the concrete XOR failure are all real and checked. Convergence is guaranteed only when a separating line exists — the honest limit shown by the XOR button.
FIG No metaphor is doing the work: the update rule, the convergence on separable data, and the concrete XOR failure are all real and checked. Convergence is guaranteed only when a separating line exists — the honest limit shown by the XOR button.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of BACKPROP · David Lee Wise (ROOT0), with AVAN