---
title: "Schmidhuber's 2014 survey names Werbos's 1981 DOE gas-market application — not the pre-1974 Deutsch project — as 'the first NN-specific application of efficient BP'"
type: "claim"
status: "budding"
source_url: "https://arxiv.org/pdf/1404.7828"
source_author: "Jürgen Schmidhuber"
source_date: "2014-10-08 (v4)"
source_venue: "Technical Report IDSIA-03-14 / arXiv:1404.7828v4, The Swiss AI Lab IDSIA"
source_tier: 1
source_sha: "9306f28e5eb896d2684114e7f0068d70fb6baaa2e26bbab47f686078013d309b"
source_quote: "To my knowledge, the first NN-specific application of efficient BP as above was described in 1981 (Werbos, 1981, 2006). Related work was published several years later (Parker, 1985; LeCun, 1985, 1988)."
provenance: "Promotion from 10-inbox/raw/2026-10-09-does-any-source-connect-lecuns-1989-first-practical.md, 2026-10-11"
origin: "batch"
derived_from: "10-inbox/raw/2026-10-09-does-any-source-connect-lecuns-1989-first-practical.md"
date_created: "2026-10-11T00:00:00.000Z"
writer_model: "claude-sonnet-5"
tags: ["paul-werbos","jurgen-schmidhuber","backpropagation","history-of-ml","priority-vs-paternity","what-was-first","natural-gas-market-model"]
audit_status: "quote read directly via extract_pdf against source_sha above at capture time (2026-10-09); no independent re-fetch this promotion, per the no-network design. | 2026-10-12 cross-model audit (claude-fable-5): re-fetched arXiv:1404.7828 via extract_pdf (tls verified; sha256 matches capture 9306f28e…); source_quote verbatim in §5.5; v4 date 2014-10-08 and report number IDSIA-03-14 confirmed against the arXiv abstract page. CORRECTION: the body previously stated flatly that 'the 1981 application named is Werbos's own DOE-developed natural-gas-market model,' implying the survey names the gas-market application — it does not: the survey cites only '(Werbos, 1981, 2006)' (the 1981 IFIP sensitivity-analysis paper, per its bibliography), and 'gas' never appears in the survey's body text. The gas-market identification is a bridge from Werbos's own 1988 paper ('This generalization was applied in 1981 to a model of natural gas markets' — Tier 1, verified verbatim in [[claim-werbos-1988-doe-affiliation-and-gas-market-test-case]]). Body rewritten to carry the bridge explicitly; the title's compression stands as shorthand for the bridged claim, which remains supported."
seek_code_commit: "b085fa6"
quote_sweep: "PASS 2026-10-12 — source_quote verbatim (normalized) in the capture-time archive (sha256 9306f28e5eb8…), same-night model-free sweep"
verified_verbatim: "2026-10-13 — source_quote matched verbatim (normalized) against a direct fetch of source_url by seek_verify (no model involved)"
---


[[entity-juergen-schmidhuber|Jürgen Schmidhuber]]'s "Deep Learning in Neural Networks: An Overview" (arXiv:1404.7828v4) states its own "first" for backpropagation's application to neural networks: "To my knowledge, the first NN-specific application of efficient BP as above was described in 1981 (Werbos, 1981, 2006). Related work was published several years later (Parker, 1985; LeCun, 1985, 1988)." What the survey itself names is the *reference* — Werbos (1981), "Applications of advances in nonlinear sensitivity analysis," Proceedings of the 10th IFIP Conference, pp. 762–770, with Werbos's 2006 retrospective AD chapter alongside — not the application's domain: the word "gas" appears nowhere in the survey's text outside the Werbos (1988) reference title. Identifying that 1981 application as [[entity-paul-werbos|Werbos]]'s own DOE-developed natural-gas-market model is a bridge supplied by Werbos's own 1988 paper, which states "This generalization was applied in 1981 to a model of natural gas markets" ([[claim-werbos-1988-doe-affiliation-and-gas-market-test-case]], Tier 1, verified verbatim). Either way, the application Schmidhuber dates to 1981 is not the pre-1974 Karl Deutsch nationalism-forecasting project that Werbos's own 1988 paper separately names "the first actual application of backpropagation" (see [[claim-werbos-1988-credits-deutsch-project-as-first-actual-application-of-backpropagation]]).

This adds a fifth candidate to the vault's running list of "what was first" for backpropagation, alongside the Deutsch project (first actual application, per Werbos's own 1988 account), the 1974 dissertation (first documented), the 1975 TSP manual (first published command), and the 1968 *Cybernetica* paper (earliest germ). Unlike the other four, this fifth entry is not Werbos's own characterization of his career — it is an external historian's framing of one specific 1981 application, made 33 years later, in a survey that elsewhere never engages Werbos's own "first actual application" claim at all (see [[observation-lecun-1989-werbos-1988-deutsch-cosine-0881-pairing-is-analytical-contrast]]).

> [!note] Seek's commentary:
> Schmidhuber's survey is this vault's most-leaned-on witness for backprop's attribution history, and the entity page on him already carries the standing caveat: corroborated, never copied. This is a case where that caveat matters directly — nobody else in the record calls the 1981 gas-market run "the first NN-specific application" of anything; it is Schmidhuber's own framing, stated in his own survey, about a man who had already made a different, earlier priority claim about his own work. Both claims can be true. Neither one corroborates the other.
> — Seek