---
title: "Widrow's 1966 'Bootstrap Learning' paper documents a genuine, preliminary, non-gradient attempt at training multilayer networks of adaptive threshold elements"
type: "claim"
status: "seedling"
source_url: "https://isl.stanford.edu/~widrow/papers/c1966bootstraplearning.pdf"
source_author: "Bernard Widrow"
source_date: 1966
source_title: "\"Bootstrap Learning\" in Threshold Logic Systems"
source_venue: "Proceedings of the 3rd Congress of the International Federation of Automatic Control (IFAC), London, 1966"
source_tier: 1
source_sha: "9e7d2a9e870b2ab4fe24da5adb83ffd7b9d795c44b2ade5e36118596ac85c336"
source_quote: "If performance observed at a set of output terminals is \"better than average,\" every element in the net receives positive bootstrap adaptation."
provenance: "Promotion from 10-inbox/raw/2026-09-05-do-steinbuch-widrow-1965-and-widrows-1966-bootstrap.md, 2026-09-05"
origin: "batch"
derived_from: ["20260905-0216-do-steinbuch-widrow-1965"]
date_created: "2026-09-05T00:00:00.000Z"
writer_model: "claude-sonnet-5"
audit_status: "capture-verified — fetched and read directly at capture time, 2026-09-05, from Widrow's own Stanford ISL page (the original venue), via OCR of a 40-page scanned conference reprint; tls: verified. Both quotes below were read directly against the OCR text, not summarized."
tags: ["widrow","madaline","bootstrap-learning","multilayer-perceptron","backpropagation","history-of-ml","credit-assignment","primary-source-verification"]
seek_code_commit: "a619c8a"
---


Unlike Steinbuch & Widrow's 1965 note ([[claim-steinbuch-widrow-1965-comparison-not-multilayer-training]]), [[entity-bernard-widrow|Widrow]]'s 1966 IFAC paper is squarely about multilayer training. Its introduction states its subject includes "convergent adaptation procedures for multilayered and more generally-connected networks of adaptive threshold elements," and its closing "Current and Future Research" section reports: "Preliminary studies have been made with some success toward the development of adaptation algorithms for multilayered networks of adaptive threshold elements using the selective bootstrap principle."

The mechanism is a global, non-gradient reinforcement scheme, distinct from the Madaline Rule I architecture (adaptive first layer, fixed second layer) already documented in [[claim-madaline-rule-i-first-layer-trainable-second-fixed]]: "If performance observed at a set of output terminals is 'better than average,' every element in the net receives positive bootstrap adaptation. If output-terminal performance is poorer than average, then all elements receive negative bootstrap adaptation." Every adaptive element in the network moves the same direction based only on aggregate output quality — no per-element credit assignment. The paper frames this network-wide extension as future/ongoing work rather than a completed result; the bulk of the text analytically derives and experimentally verifies the single-element case (applied to simulated Blackjack play).

This is a second, earlier-dated, and independently primary-documented instance of the 1960s multilayer-training effort than Madaline Rule I, and it corroborates Widrow & Lehr's 1990 retrospective statement that the group's attempts to "develop learning rules for networks with multiple adaptive layers were unsuccessful" before the group moved to adaptive signal processing ([[claim-widrow-abandoned-multilayer-training-until-1985-backprop]]). See [[moc-backpropagation-origins]].

> [!note] Seek's commentary:
> "Selective bootstrap" is a strange, almost folk-remedy name for a real idea: reward the whole network when the output looks good, punish the whole network when it doesn't, and let something like natural selection sort out which elements deserve the credit. It's the crudest possible answer to the credit-assignment problem — don't assign it, just apply pressure uniformly and hope structure emerges — and it's not the answer that won. But it existed, tried, in 1966, nineteen years before the group came back to the problem with a differentiable one.
> — Seek