capture promoted 2026-07-05

Does Griewank (2012) "Who Invented the Reverse Mode of Differentiation?" mention or discuss Shun'ichi Amari as a priority claimant?

What the secondary record suggests (circumstantial, not primary-confirmed)

Claim type: historical/bibliographic (about the contents and scope of a specific document — normally this would need Tier 1-2 confirmation from the document itself; that could not be obtained, so everything here is presented as inference from citation patterns rather than a settled fact).

Four independent documents that discuss, cite, or summarize Griewank (2012) were located and read in full. None of them link Griewank's paper to Amari in any way:

Jürgen Schmidhuber's "Deep Learning in Neural Networks: An Overview" (arXiv:1404.7828 / Neural Networks 61, 2015) cites Griewank (2012) exactly once, in a sentence about backpropagation's identity with reverse-mode automatic differentiation:

"BP is also known as the reverse mode of automatic differentiation (Griewank, 2012), where the costs of forward activation spreading essentially equal the costs of backward derivative calculation."

The same document cites Amari separately, in an unrelated list of early optimal-control/gradient authors:

"(e.g., Kelley, 1960; Bryson, 1961; Bryson and Denham, 1961; Pontryagin et al., 1961; Dreyfus, 1962; Wilkinson, 1965; Amari, 1967; Bryson and Ho, 1969...)"

The two citations are in different sections addressing different points; nothing in the document states or implies that Griewank's paper itself discusses Amari.
Schmidhuber's mailing-list post to the Connectionists list (2014-07-22, "Who invented backpropagation?") repeats the identical pattern: Griewank (2012) cited only for the BP/reverse-mode-AD equivalence, Amari (1967) cited separately in the same optimal-control author list, with no connective statement between them.
Baydin, Pearlmutter, Radul & Siskind, "Automatic Differentiation in Machine Learning: A Survey" (arXiv:1502.05767 / JMLR 18, 2018) cites Griewank (2012) once, for the general historical framing that backpropagation "has... a colorful history of having been reinvented at various times by independent researchers (Griewank, 2012; Schmidhuber, 2015)." This survey does not mention Amari anywhere in its full text.
Schmidhuber's most recent and most comprehensive historical survey, "Annotated History of Modern AI and Deep Learning" (arXiv:2212.11279, revised as recently as 2025-12-29) — checked via its Hugging Face paper-page summary — discusses Amari's 1967–68 work on stochastic-gradient-descent-trained multilayer networks at length, in a section separate from any discussion of Griewank, automatic differentiation, or the "who invented reverse mode" question. No connection between the two is drawn.

Why this is suggestive: Schmidhuber is, across all four documents, simultaneously the most prominent living advocate for crediting Amari with early priority in gradient-based learning for neural networks, and a close reader/citer of Griewank's reverse-mode-differentiation history. If Griewank's 2012 paper discussed Amari as a priority claimant, it would be a natural and useful citation for Schmidhuber's own advocacy — and he cites Griewank (2012) multiple times across a decade of writing without ever drawing that link. That absence, repeated across independent documents written in 2014, 2015, and as late as a 2025 revision, is circumstantial evidence that Griewank's paper does not discuss Amari.

Why this is not sufficient to settle the question: none of this is a direct reading of Griewank's paper. It is an inference from what other authors chose to cite alongside it. It remains possible that Griewank's paper mentions Amari in a footnote, aside, or bibliography entry that none of these four secondary sources had reason to surface.

What is independently known about the paper's likely scope

Claim type: historical/definitional (uncontested framing).

Search-engine-indexed summaries of the paper's abstract and body (not independently verified against the primary text, so treated cautiously) describe its subject matter as centered on the numerical-analysis and optimal-control lineage of reverse-mode differentiation: Nick Trefethen's inclusion of automatic differentiation among "the 30 great numerical algorithms of the last century," Gerardi Ostrowski's chemical-engineering process models (~1965), Seppo Linnainmaa's 1970 rounding-error work, Bert Speelpenning's 1980 thesis, and Paul Werbos's optimal-control-derived comparison of forward and reverse propagation. This is consistent with the paper being framed around numerical computing and optimal control, not neural-network history — which would make Amari (a neural-network theorist whose 1967 work is not itself about automatic differentiation or the specific reverse-mode/chain-rule algorithm) plausibly out of scope for the piece regardless of any priority-dispute angle. This scope characterization rests on secondary summaries only and is flagged [unverified — needs primary] for its specificity, though it is consistent (and not contradicted) by every source consulted.

Further leads

Retry direct access to Griewank (2012) when EMIS/CiteSeerX mirrors recover, or locate the paper via a library/institutional proxy, or obtain PDF-to-text tooling (poppler/pdftotext) in the research sandbox — the full parent volume PDF (https://www.zib.de/userpage/groetschel/publications/OptimizationStories.pdf) was successfully downloaded and just needs text extraction on pp. 389–400.
Griewank's own bibliography/reference list in the 2012 paper (not accessible this session) would settle the question outright — either Amari appears in the reference list or he does not.
Related sibling capture 2026-06-29-did-shunichi-amari-describe-a-form-of-gradient-descent-for-layered-networks-in-the-1960s already established that Amari's 1967 contribution is a general SGD method for layered networks, distinct from the specific reverse-mode/chain-rule algorithm that is Griewank's actual subject — this is an independent reason (beyond the citation-pattern argument above) to expect Griewank's paper may simply not have found Amari relevant to include, rather than having considered and rejected his claim.

Seek's commentary: This capture answers a narrower question than it looks like at first glance. The task was framed as "does Griewank discuss Amari," but the honest finding is "no primary evidence either way was obtainable, and the circumstantial evidence points toward 'probably not, because it's out of scope' rather than toward 'yes, and here's what he said.'" That's a meaningfully different and weaker finding than a confirmed no. Worth flagging for whoever promotes this: the access wall on Griewank (2012) has now failed two separate research sessions on two separate days via entirely different mirrors, which suggests this isn't bad luck but a structurally hard-to-reach document from this environment — it may be worth explicitly requesting a manual fetch (e.g., a human downloading the EMIS/zbmath copy or a library PDF and dropping it in the vault) rather than a third automated attempt.

· batch run 2026-07-05 — primary source (Griewank 2012) remains unreachable; question addressed via convergent secondary sourcing. See access log below. · raw markdown