THE PERCEPTRON — a dart

THE PERCEPTRON the first thing that learned

The original trainable neuron: weigh the inputs, sum, fire if the sum clears a threshold. Its learning rule is a single line — on a mistake, nudge the weights toward the right answer: w ← w + (y − ŷ)·x. Rosenblatt built it in 1958 (in hardware) and it provably learns any linearly separable pattern. Its famous limit — it cannot learn XOR — froze the field for a decade until multilayer nets returned.

THE TECHNIQUE w ← w + (y − ŷ)·x

The demo trains a perceptron on the AND gate — one update and it classifies (1,1)→1: live demo

HISTORY & CREDIT Rosenblatt · 1958

“Minsky & Papert proved neural nets can’t learn XOR.” — they showed a single perceptron can’t; a two-layer net can. The 1969 result stalled, not sank, the field. cited

weigh & fire · output 1 if w·x + b > 0.
the rule · on error, w ← w + (y−ŷ)x — converges for separable data.
1958 · Frank Rosenblatt; the XOR limit — Minsky & Papert 1969.

The seed of every neural network. learning rule

RECOMMEND FOR I-13 one update, on the compiler

On i-13, the AND perceptron starts wrong on (1,1), one update sets w=(1,1) b=1, and now it fires correctly:

$ i13 run nn_perceptron.i13 RUN OK · 76 step(s) yhat0 = 0 (wrong) err = 1 w -> (1,1) b -> 1 yhat1 = 1 learned = 1

Recommend as a NULL — a learning rule (B39). The update and prediction are pinned; convergence for separable data is a theorem. A policy, not a same-function DOF. NULL — the first thing that learned.