Does Amari's 1968 Japanese book (Kyoritsu) contain Saito's five-layer MLP experiment, and what exactly does it report?

Short answer: The claim traces to exactly one origin — Jürgen Schmidhuber — repeated consistently across his own self-published pages and (apparently) his arXiv survey, and reproduced by two Wikipedia articles that cite him rather than confirming independently. The book's existence (title, publisher, year) is independently corroborated. The specific content attributed to it — that pages 119–120 describe a computer experiment by a student named Saito training a five-layer MLP with two modifiable layers to classify non-linearly-separable pattern classes — rests entirely on Schmidhuber's own reading of a scanned PDF he hosts (which this session could not read) plus an unpublished "personal communication" with Amari in 2021 for Saito's identity. No independent scholar, no accessible Amari autobiographical writing, and no library record of Saito's thesis was found to corroborate the specific mechanism. Central question: [unverified — could not confirm or deny after search].

Claim: Amari published a Japanese-language book in 1968, titled (in English translation) roughly "Information Theory II — Geometric Theory of Information," through Kyoritsu Shuppan

Claim type: Historical/bibliographical — uncontested, and independently corroborated by two sources that do not derive from each other.

Wikipedia's Shun'ichi Amari article lists in its "Key works" section: "A Geometrical Theory of Information (in Japanese), Kyoritsu, 1968." Independently, a mirror of Hiroshi Matsuzoe's academic survey "Half a century of information geometry, part 1" (published in the journal Information Geometry, Springer) cites: "Amari, S.: Information theory II – Geometric theory of information (Kyoritsu Shuppan, Tokyo) (1968) (in Japanese)." These two citations agree on publisher (Kyoritsu / Kyoritsu Shuppan), year (1968), language (Japanese), and the substance of the title, and come from genuinely separate lines of transmission (an encyclopedia bibliography vs. an academic survey's reference list) — neither one is a copy of Schmidhuber's page.

Sourcing floor check: Historical/definitional claim, uncontested; clears the Tier 3–4 floor comfortably, and is strengthened by two independent Tier-3 sources agreeing rather than one.

Claim: Schmidhuber asserts, in his own words, that the 1968 book (pages 119–120) contains computer-simulation results for a five-layer network with two modifiable layers, learning internal representations to classify non-linearly-separable pattern classes, implemented by Amari's student Saito

Claim type: Specific technical-mechanism claim — requires Tier 1–2 per the rubric. Note the object of this claim is precisely "what Schmidhuber asserts," which is directly sourced to his own primary-authored page; whether that assertion is itself accurate against the book's actual text is a separate, unresolved question (next claim).

Schmidhuber's "Who Invented Backpropagation?" page gives, as reference [GD2]: "S. I. Amari (1968). Information Theory—Geometric Theory of Information, Kyoritsu Publ., 1968 (in Japanese). OCR-based PDF scan of pages 94-135 (see pages 119-120). Contains computer simulation results for a five layer network (with 2 modifiable layers) which learns internal representations to classify non-linearily separable pattern classes." His companion reference [GD2a] reads: "H. Saito (1967). Master's thesis, Graduate School of Engineering, Kyushu University, Japan. Implementation of Amari's 1967 stochastic gradient descent method for multilayer perceptrons. (S. Amari, personal communication, 2021.)" His "Annotated History of Modern AI and Deep Learning" narrative page states the same claim in prose: "Amari's implementation (with his student Saito) learned internal representations in a five layer MLP with two modifiable layers, which was trained to classify non-linearily separable pattern classes."

The URL Schmidhuber cites as the primary artifact, https://people.idsia.ch/~juergen/amari1968p94-135ocr.pdf, was confirmed live this session: it is a real 4.8MB PDF consisting of scanned page images (2340×1654px JPEG frames per page) with Adobe Acrobat OCR-layer metadata, spanning what its filename claims is pages 94–135 of the source book — consistent in form with a genuine period scan. However, the tools available this session could not extract or read its text (WebFetch returned only raw binary/metadata; no OCR or PDF-text-extraction tool was available), so this session cannot confirm what the scanned pages actually say — only that Schmidhuber's cited artifact exists and is plausible in form.

Sourcing floor check: As a claim about what Schmidhuber himself says, this clears Tier 1–2 (his own words, on his own page, quoted verbatim, and mirrored consistently across two of his pages). As a claim about what the book itself actually contains, it does not independently clear the floor for a technical-mechanism claim: the only source is a single historian's paraphrase of a document neither this researcher nor any other source found could independently read, plus an unpublished personal communication for Saito's identity. [unverified-mechanism — needs primary].

Claim: Wikipedia's repetitions of the Saito/five-layer claim are not independent corroboration — both articles that carry it cite Schmidhuber's arXiv paper directly

Claim type: Claim about citation provenance, verifiable by direct comparison of footnotes.

The Shun'ichi Amari Wikipedia article states "The same year, Amari and his student H. Saito reported the first multilayer perceptron (MLP) neural network trained by SGD," with the footnote resolving to "Schmidhuber, Juergen (2022). 'Annotated History of Modern AI and Deep Learning'. arXiv:2212.11279." Separately, the History of artificial neural networks Wikipedia article states "According to Amari, in computer experiments conducted by his student Saito, a five layer MLP with two modifiable layers learned internal representations to classify non-linearly separable pattern classes," footnoted to the same Schmidhuber arXiv paper. Both Wikipedia articles, in other words, are downstream of the same single source rather than two lines of evidence — a pattern this vault has flagged before in adjacent captures on Amari/Schmidhuber claims (see 2026-07-03-did-robbins-monro-1951-originate-stochastic-gradient-descent.md, which found a shared page-range error and a shared misspelling between Schmidhuber's text and a Wikipedia article on a related claim, suggesting direct copying rather than independent research).

Sourcing floor check: This is a claim about citation structure, directly verifiable by comparing quoted footnotes — the comparison itself is the evidence, not a single tier-rated source.

Claim: No independent record of H. Saito's 1967 Kyushu University master's thesis was found

Claim type: Search-completeness note — a documented negative result, not a positive factual claim.

Targeted searches for Saito's thesis (by the title/subject Schmidhuber describes, by author name, by institution) returned nothing beyond restatements of Schmidhuber's own [GD2a] reference entry. No library catalog record, no independent citation of it in another paper, and no confirmation of the thesis's existence outside Schmidhuber's citation — which is itself explicitly sourced to "personal communication" with Amari in 2021, not a public or checkable document. Amari's own most likely accessible autobiographical account of this period, "Dreaming of mathematical neuroscience for half a century" (Neural Networks, 2013), could not be read this session: the ScienceDirect version is paywalled (HTTP 403) and a hosted PDF mirror is an image-based scan not text-extractable with the tools available.

Sourcing floor check: N/A — documented absence, recorded so a future researcher does not repeat the same unsuccessful search paths without knowing they were tried.

Further leads

• The single highest-value unresolved action is reading https://people.idsia.ch/~juergen/amari1968p94-135ocr.pdf directly (via OCR tooling, e.g. poppler/pdftotext or an image-capable model pass over the page images) — this would let a researcher check Schmidhuber's paraphrase against the actual scanned Japanese text on pages 119–120, which is the only thing that can truly settle the core question.
• Amari's 2013 Neural Networks autobiographical article, "Dreaming of mathematical neuroscience for half a century," remains unread in this vault's research chain (paywalled abstract; scanned PDF mirror not text-extractable). This is the most likely place to find Amari describing Saito's work in his own accessible words, independent of a personal communication to Schmidhuber.
• The arXiv version of Schmidhuber's claim (arXiv:2212.11279, full PDF) was not read this session beyond its abstract page; confirming the exact wording there (as opposed to only the self-published HTML mirrors) would strengthen the claim's tier, since arXiv is treated as Tier 1 per this vault's sourcing rubric while Schmidhuber's personal website pages are treated as Tier 2.
• No search this session surfaced any published critique or rebuttal specifically targeting the Saito/five-layer claim (as distinct from the broader, well-documented controversy over Schmidhuber's priority claims in general). This absence should not be read as confirmation — it may simply reflect how obscure and hard-to-check the underlying 1968 Japanese text is for most historians of the field.

Central question status: The book's existence and bibliographic identity are confirmed independently of Schmidhuber. The specific claim that it contains Saito's five-layer MLP experiment on pages 119–120 is sourced only to Schmidhuber (self-published pages, mirrored in prose form; likely also in his arXiv paper though not independently confirmed there this session) plus an unpublished personal communication with Amari for Saito's identity — with no independent corroboration found and the primary artifact itself unread. [unverified — could not confirm or deny after search].