---
title: "Madaline Rule I (1962) could adapt only the first layer of Widrow's networks — the second layer stayed fixed, making it the ceiling of pre-backpropagation multilayer training"
type: "claim"
status: "budding"
audit_status: "primary-verified (Widrow & Lehr 1990, Proc. IEEE 78(9), read directly via pdftotext 2026-07-09 once the poppler dependency was repaired; the paper names Madaline Rule I as the earliest popular learning rule and states the 1960s Madalines had adaptive first layers with fixed second layers. Two-column scan merges columns, so the architectural facts are confirmed-present rather than lifted as an unbroken verbatim string; the clean verbatim quote below is the adjacent mechanism sentence.)"
source_url: "https://isl.stanford.edu/~widrow/papers/j199030years.pdf"
source_author: "Bernard Widrow and Marcian A. Lehr"
source_date: "1990-09"
source_tier: 1
source_quote: "The adaptive elements in the original Madaline structure used hard-limiting quantizers (signums), while the elements in the backpropagation network use only differentiable nonlinearities, or \"sigmoid\" functions."
provenance: "Promotion from 10-inbox/raw/2026-07-08-did-widrow-attempt-and-abandon-multilayer-network-training-around-1965-1966.md, 2026-07-09, Fable clean-lane promotion marathon; the capture's [unverified-mechanism] flag was cleared by a direct read of Widrow & Lehr 1990."
origin: "batch"
derived_from: ["20260708-0224-did-widrow-attempt-and"]
date_created: "2026-07-09T00:00:00.000Z"
tags: ["widrow","madaline","adaline","multilayer-perceptron","backpropagation","history-of-ml","credit-assignment"]
audits: ["2026-07-09 claude-fable-5"]
drafted_in: ["what-the-gradient-cant-see"]
---


Madaline Rule I (MRI), devised by [[entity-bernard-widrow|Bernard Widrow]]'s Stanford group around 1961–62,
is described in Widrow's own retrospective (Widrow & Lehr 1990) as "the earliest
popular learning rule" for networks of multiple adaptive elements. Its limit was
structural: the Madalines of the 1960s had an **adaptive first layer** of Adaline
units feeding a **fixed** second (logic) layer. Only one layer actually learned —
so MRI was not multilayer training in the modern sense, but a single trainable
layer wired into a hand-set downstream combiner.

The obstacle to going further was the non-differentiable element. Widrow & Lehr
state that "the adaptive elements in the original Madaline structure used
hard-limiting quantizers (signums), while the elements in the backpropagation
network use only differentiable nonlinearities, or 'sigmoid' functions." A
hard-limiting quantizer has no usable gradient, so the error at the output cannot
be propagated back to adjust a hidden layer — the [[entity-credit-assignment|credit-assignment]] barrier that
[[entity-marvin-minsky|Minsky]] and Papert only conjectured to be fatal ([[claim-perceptrons-multilayer-sterile-was-conjecture]]),
Widrow's group hit empirically. Their later attempts to adapt multiple layers
while keeping the simpler quantizers produced Madaline Rule II (1987) and III
(1988), not a 1960s solution.

This makes MRI the concrete high-water mark behind the stall the field is mapped
around ([[backpropagation-gap]]): the method to train hidden layers did not exist
until the differentiable-nonlinearity route ([[claim-rhw-1986-demonstration-not-invention]])
reopened it. Widrow himself did not return to neural nets until backpropagation
reached him in the mid-1980s ([[claim-widrow-abandoned-multilayer-training-until-1985-backprop]]).
See [[moc-backpropagation-origins]].