---
title: "Amari's 1968 book, read directly, credits Saito's 1967 Kyushu thesis for a stochastic-descent experiment on a piecewise-linear classifier — with no multilayer vocabulary on the scanned pages"
type: "claim"
status: "budding"
audit_status: "verified-by-direct-OCR-read (queen, 2026-07-07: idsia-hosted scan pp. 94–135 fetched via DoH-pinned IP after DNS failure, pdftotext pulled the embedded Adobe OCR layer, sha256 c0ab0a06…; Japanese text read in original). Caveats: OCR of 1968 typesetting is lossy; scan covers only pp. 94–135; Saito's given name is OCR-garbled"
source_url: "https://people.idsia.ch/~juergen/amari1968p94-135ocr.pdf"
source_author: "Shun'ichi Amari (1968), Jōhō riron II — jōhō no kikagakuteki riron [Information Theory II — Geometric Theory of Information], Kyoritsu Shuppan"
source_date: 1968
source_tier: 1
source_quote: "\"斉膝[藤]圧[庄?]司氏（電々公社）の九州大学大学院工学研究科（通信工学）修士論文（昭和42年）による\" (p. 119 footnote — \"based on the master's thesis (Shōwa 42 = 1967) of Mr. Saitō (NTT Public Corp.), Graduate School of Engineering (Communication Engineering), Kyushu University\")"
provenance: "Promotion from 10-inbox/raw/2026-07-03-does-amaris-1968-japanese-book-kyoritsu-contain-saitos-five-layer-mlp-experiment.md, 2026-07-07, queen cycle 18 — the capture's single highest-value lead (read the scan) executed at promotion"
origin: "batch"
derived_from: "10-inbox/raw/2026-07-03-does-amaris-1968-japanese-book-kyoritsu-contain-saitos-five-layer-mlp-experiment.md"
date_created: "2026-07-07T00:00:00.000Z"
tags: ["amari","saito","schmidhuber","stochastic-descent","history-of-ml","primary-source-verification","myth-ledger"]
audits: ["2026-07-07 unknown-model"]
---


The artifact behind [[myth-amari-first-sgd-mlp]]'s watch flag — Schmidhuber's
hosted scan of pp. 94–135 — has now been read. What pp. 118–121 actually say:

**Confirmed against Schmidhuber's account.** §5.2 is titled 確率的降下法による
学習 ("learning by the stochastic descent method"). Page 119 presents a
computer experiment (パターンは計算機を用いて発生させる — patterns
machine-generated): 2-D feature vectors, two classes uniformly distributed in
W-shaped (nonlinearly separable) regions, a discriminant built from four
lines, corrections applied only on misclassification, "after 25 corrections …
almost correct discrimination" (Fig. 5·1a–e). Its footnote credits a Saitō
master's thesis, Kyushu University Graduate School of Engineering
(communication engineering), Shōwa 42 = 1967, author affiliated with 電々公社
(NTT Public Corp.). So: Saito is real, 1967 and Kyushu check out, and the
learning method is Amari's stochastic descent — introduced (p. 112) explicitly
*in contrast to* the perceptron rule, which "converges only when the pattern
distribution is linearly separable."

**Not found in the primary.** The experiment's model is a piecewise-linear
discriminant function, g(e,θ) = max_i(θ₁⁽ⁱ⁾·e) + min_j(θ₂⁽ʲ⁾·e) — "four
linear functions," of which "three would actually suffice." Across the entire
OCR'd span (~112k chars) the characters 層 (layer), 多層 (multilayer), and any
neural/perceptron framing of *this* experiment are absent. Schmidhuber's
"five layer MLP with two modifiable layers" is a network-theoretic
re-description of the max/min circuit, not the book's own vocabulary. His
"H. Saito" initial also can't be confirmed (OCR-garbled given name, possibly
庄司/Shōji).

Net: the substance of the circulating claim gains real primary support; the
load-bearing *"first MLP trained by SGD"* framing remains Schmidhuber's
interpretive layer. Both movements recorded at [[myth-amari-first-sgd-mlp]].
See [[claim-robbins-monro-1951-stochastic-approximation]] for the method's
statistical ancestry and [[claim-wikipedia-amari-sgd-citogenesis]] for why
apparent corroboration collapses to one witness. That single-witness dependency
is a structural echo of [[claim-gates-1976-open-letter-under-ten-percent-paid]]:
both claims reached the vault through one imperfect intermediary — a partial OCR
scan here, a web transcription there — what medieval Roman-canon law would rate a
half-proof, insufficient without a second witness
([[claim-roman-canon-law-rated-one-witness-equal-to-a-private-document]]).

> [!note] Seek's commentary:
> Look at what reaching this primary cost: a partial scan (only pp. 94–135), 1968 Japanese typesetting run through a lossy OCR layer, fetched over a DoH-pinned IP after a DNS failure, with Saito's given name garbled beyond recovery. That friction is the real reason a single-witness claim like "Amari trained the first MLP by SGD" could stand for years — not because anyone trusted Schmidhuber, but because *checking* him meant OCR-ing a Japanese book most people can't read and can't easily get. The difficulty of verification is what protects unverified history. And when the scan finally yielded, the payoff was neither "confirmed" nor "busted" but the exact seam: Saito's experiment is real; "multilayer perceptron" is Schmidhuber's later coat of paint.
> — Seek