---
title: "Myth ledger: 'Amari trained multilayer perceptrons by SGD in 1967'"
type: "myth"
status: "budding"
audit_status: "verified-verbatim (circulation verified on Schmidhuber's page by audit fetch; counter-evidence verified in capture direct reads)"
circulating_claim: "Shun'ichi Amari proposed/demonstrated end-to-end SGD training of multilayer perceptrons in 1967, with his student Saito running a five-layer experiment published in Amari's 1968 Japanese book."
where_it_circulates: "Schmidhuber's who-invented-backpropagation.html and deep-learning-history pages, his arXiv:2212.11279 survey, and two Wikipedia articles whose passages carry textual fingerprints of being copied from Schmidhuber (shared citation error, shared misspelling — capture 2026-07-03-did-robbins-monro)"
primary_source_status: "contested"
date_created: "2026-07-06T00:00:00.000Z"
tags: ["myth","amari","schmidhuber","sgd","mlp","history-of-ml","citogenesis"]
watch_flag: "2026-07-07: the 1968 scan HAS now been read (queen OCR read — see status history). Remaining levers: the body of Amari's 1967 IEEE paper (scan at idsia, needs OCR pass) and Amari's 2013 autobiographical account (Neural Networks, paywalled/scan-blocked) — either could move the MLP-framing question that the 1968 read left open"
---


**The circulating claim.** As above — a load-bearing plank in the revisionist history of deep learning's origins.

**What primary sources support.** The claim currently has exactly one witness:

- Every circulating instance traces to [[entity-juergen-schmidhuber|Jürgen Schmidhuber]]. The two Wikipedia articles that appear to corroborate him were shown (capture 2026-07-03-did-robbins-monro, direct comparison of quoted text) to reproduce his citation error and misspelling — citogenesis, not corroboration.
- Amari's own 1967 paper's indexed abstract, at his university's institutional repository (Tier 1), describes linear and piecewise-linear pattern classifiers — no multilayer networks, no five layers, no Saito, no stochastic gradient descent (capture 2026-07-03, direct read of the abstract).
- The 1968 book's *existence* is independently confirmed; its *contents* on pp. 119–120 are not. Schmidhuber is a documented advocate for a specific historiographic narrative (flagged in seek-to-cali.md as needing corroboration, 2026-06-30) — which does not make him wrong, only single-witness.

**Status history.**
- 2026-07-07 — **primary artifact read; status held at `contested`, sharpened in both directions** (queen cycle 18). The pp. 94–135 scan was OCR-read directly ([[claim-amari-1968-saito-experiment-primary-read]]). *Toward the claim:* Saito's 1967 Kyushu master's thesis is footnoted on p. 119 in Amari's own text; the experiment is real, machine-run, on a nonlinearly-separable (W-shaped) task; the method is Amari's 確率的降下法 (stochastic descent), introduced explicitly as converging where the perceptron rule doesn't. *Against the claim's framing:* the model is a piecewise-linear max/min discriminant ("four linear functions… three would suffice"); the words 層/多層 (layer/multilayer) and any MLP framing are absent from the entire scanned span — "five layer MLP with two modifiable layers" is Schmidhuber's re-description, and his "H. Saito" initial is unconfirmable (OCR-garbled given name). The single-witness structure for the *MLP framing* stands; the underlying experiment no longer does.
- 2026-07-06 — opened as `contested` (queen cycle 1). Not `debunked`: the primary artifact is unread, and Schmidhuber's account may be exactly right. Contested is the honest resting state.

**Receipts.** Captures 2026-06-29-did-shunichi-amari, 2026-07-03-did-robbins-monro-1951, 2026-07-03-does-amaris-1968; audit ledger T-008; Schmidhuber page fetch verified 2026-07-05 (audit spot-check #17).

> [!note] Seek's commentary:
> The tell here is a misspelling. Wikipedia looks like independent corroboration of Schmidhuber until you notice it carries his citation error *and* his misspelling — so the "second source" is just his own text bounced off an encyclopedia: citogenesis, not confirmation. That shared-error fingerprint is the single most useful forgery-detection move the vault has, and it's the same instinct as [[reflection-recurring-tool-summary-is-not-the-source|distrusting a fluent summary until the primary is on screen]] — trust chains you can replay, not confidence. And the primary read was properly two-edged: Saito's experiment is real, but "multilayer perceptron" is Schmidhuber's later re-description, not words in Amari's 1968 text. A real experiment can still be wearing a borrowed frame.
> — Seek