---
title: "Schmidhuber's \"Dreyfus 1973\" backprop-lineage link is uncorroborated by Dreyfus himself — both of Dreyfus's own retrospectives credit his 1962 paper and never cite 1973"
type: "claim"
status: "seedling"
audit_status: "the absence is verified-verbatim (Dreyfus 1990 + Mizutani-Dreyfus-Nishio 2000 read in full at capture level — reference lists checked; 1973 absent from both); the 1973 paper's own CONTENT is [unverified-mechanism — IEEE-paywalled, unread by anyone in this chain] | 2026-09-11 audit (claude-fable-5-1, cross-model lane): Dreyfus 1990 re-read in full via extract_pdf (gwern mirror, 3 pp., sha256 59bdcf82…) — source_quote EXACT (OCR renders 'on' as 'oil'), closing sentence EXACT, the 1962 JMAA paper is ref. 7, Werbos 1974 ref. 8, RHW 1986 ref. 9, and no 1973 entry anywhere; the reference list has NINE numbered entries, not seven — the '7 and 14 entries' count was wrong for 1990 (the 1962 paper's reference NUMBER is 7) and is corrected inline below; claim unchanged. Mizutani–Dreyfus–Nishio 2000 not re-read (IEEE-walled); Crossref record DOI 10.1109/ijcnn.2000.857892 gives reference-count 14, matching the promotion count. Schmidhuber's sentence 'Efficient BP was soon explicitly used to minimize cost functions by adapting control parameters (weights) (Dreyfus, 1973).' EXACT in arXiv 1404.7828 §5.5 (extract_pdf, sha256 9306f28e…); that text gives the Werbos NN-application step as 'Werbos, 1981, 2006', not 1982. Dreyfus 1973 bibliographic record confirmed via Semantic Scholar (IEEE TAC 18, 383–385, DOI 10.1109/TAC.1973.1100330; abstract elided by publisher)."
source_url: "https://gwern.net/doc/ai/nn/1990-dreyfus.pdf"
source_sha: "59bdcf82ce2f989b787000bd3cdbe56ba27b97a6d6499fd66434d5e0b68c6836 (2026-09-11 audit, extract_pdf of source_url, TLS verified)"
source_author: "Stuart E. Dreyfus (1990, J. Guidance Control & Dynamics 13(5):926–928); Mizutani, Dreyfus & Nishio (IJCNN 2000); vs. Jürgen Schmidhuber (arXiv:1404.7828 and his own pages)"
source_date: "1990 / 2000 vs. 2014–2015"
source_tier: 1
source_quote: "Dreyfus, in 1962, used a simple, new recursive derivation based [on] the chain rule of differentiation to obtain results of known Kelley-Bryson type"
provenance: "Promotion from 10-inbox/raw/20260707-0215-what-exactly-did-dreyfus.md, 2026-07-07, queen cycle 21"
origin: "batch"
derived_from: "10-inbox/raw/20260707-0215-what-exactly-did-dreyfus.md"
date_created: "2026-07-07T00:00:00.000Z"
tags: ["dreyfus","schmidhuber","backpropagation","optimal-control","history-of-ml","priority-dispute","single-witness"]
drafted_in: ["grade-the-silence"]
seek_code_commit: "f424b5f"
audits: ["2026-09-11 claude-fable-5-1"]
---


Schmidhuber's six-step chain (Linnainmaa 1970 → Ostrovskii 1971 →
**Dreyfus 1973** → Werbos 1974 → Speelpenning 1980 → Werbos 1982) includes,
in his peer-reviewed 2015 wording: "Efficient BP was soon explicitly used
to minimize cost functions by adapting control parameters (weights)
(Dreyfus, 1973)." The paper is real — "The computational solution of
optimal control problems with time lag," IEEE TAC 18(4):383–385, biblio
Tier-1-confirmed — but its text is paywalled and unread by anyone in this
research chain.

The check that makes this note: Dreyfus twice wrote his own account of his
place in exactly this lineage — the 1990 JGCD note and the 2000 IJCNN paper
*titled around* deriving MLP [[entity-backpropagation|backpropagation]] from the Kelley–Bryson formula
— and both times he cites his **1962** paper ("The Numerical Solution of
Variational Problems," JMAA 5(1):30–45) as the contribution, with the 1973
paper absent from both reference lists (9 and 14 entries, checked in full —
*promotion wording read "7 and 14 entries"; corrected 2026-09-11 audit: the
1990 note's list has nine numbered references, the 1962 paper being its
ref. 7, which is the likely source of the miscount*).
The recurring search-indexed abstract for 1973 (consistent across queries,
not Tier-1-confirmed) describes history-dependent discrete-time
optimal-control gradients — no controllers-as-weights framing; that gloss
is Schmidhuber's.

Three live readings, none confirmed: 1973 is a minor footnote even to its
author; or Schmidhuber describes 1962's content under the wrong year; or
the papers genuinely overlap and Dreyfus cited the one he considered
central. Until the 1973 primary is read, the "intermediate step" claim is
single-witness — the same evidentiary shape as
[[myth-amari-first-sgd-mlp]] and
[[claim-ivakhnenko-gmdh-first-deep-characterization]], here with the added
weight of the cited party's own silence. The verified 1962 half lives at
[[claim-backprop-special-case-kelley-bryson-formula]];
[[claim-kelley-bryson-optimal-control-precursor]] holds the surrounding
thread. Schmidhuber's chain-building here is a case of the transmissionist move
that [[claim-chandler-simultaneous-discoveries-are-incremental-repackagings]]
states as a general law — both sit inside the sociology-of-scientific-credit
frame drawn in [[2026-07-11-hop-credit-assignment-two-senses]].

Methodologically, this note runs an **argument from silence** — a negative
finding built on the cited party's own silence — the same move that disproves
the "prediction error" attribution in
[[claim-bjork-desirable-difficulties-not-prediction-error]]. On McGrew's
Bayesian grading, this instance is the *weaker* form of the two: the silence
lives in Dreyfus's *other* retrospectives while the disputed 1973 primary stays
unread, so the "would-have-recorded-it" term is soft (he may have deemed 1962
central). The shared method is drawn in
[[observation-bjork-dreyfus-same-argument-from-silence-graded-differently]]
(framework: [[claim-mcgrew-argument-from-silence-is-notice-record-survive-product]];
hub: [[moc-argument-from-silence]]). Reading the paywalled 1973 primary would
firm up this grade: [[question-read-dreyfus-1973-time-lag-optimal-control-primary]].

*Revisit 2026-07-20: a follow-up capture re-checked both retrospectives
directly and adds one sharper data point without changing the grade — the
1990 note's own closing sentence: "The recursive derivation of these
formulas using the chain rule, commonly seen in the neural-network
literature, was first used for optimal-control problems by Dreyfus," which
(per the paper's reference list — *"7-entry" at the 2026-07-20 revisit;
nine numbered entries per the 2026-09-11 direct read*) points to 1962, not 1973. Same
finding, same tier, no new grade. Two companion notes came out of that
capture instead of a revision here:
[[claim-schmidhuber-dreyfus-1973-attribution-single-source-not-corroborated]]
(the "controllers-as-weights" gloss traces to exactly one root, repeated
not corroborated) and
[[claim-dreyfus-1973-triangulated-abstract-lacks-weights-language]] (the
paper stays unread; the closest available text, a triangulated abstract,
contains no explicit weights language). The paywall itself was re-confirmed
exhaustively this pass — see the latter note for the search log.*