---
id: "20261009-0213-does-any-source-connect"
title: "Does any source connect LeCun's 1989 'first practical recognition application' to Werbos's 1988 Deutsch-project 'first actual application' claim?"
type: "capture"
status: "promoted"
promoted_to: ["30-notes/observation-lecun-1989-werbos-1988-deutsch-cosine-0881-pairing-is-analytical-contrast.md","30-notes/claim-schmidhuber-2014-names-1981-werbos-gas-market-model-first-nn-specific-application.md","50-questions/question-does-werbos-1994-roots-of-backpropagation-or-2006-retrospective-mention-lecun.md","40-entities/entity-karl-deutsch.md (dated update appended)","40-entities/entity-paul-werbos.md (dated update appended)","40-entities/entity-yann-lecun.md (dated update appended)","40-entities/entity-juergen-schmidhuber.md (dated update appended)","40-mocs/moc-backpropagation-origins.md (two notes + one question linked in)"]
not_promoted: ["Andrej Karpathy's 2022 characterization of LeCun's 1989 paper as 'the earliest real-world application of a neural net trained end-to-end with backpropagation' — arrived only via a WebSearch summary of a gwern.net mirror, never a direct fetch; per the quote-provenance rule this is not sourced and not load-bearing to any kept claim, so it stays an unread lead rather than a claim or a routed question.","Andrej Karpathy as an entity-page candidate — declined (UNSURE per the promotion test). His only appearance in this capture is the unverified quote above; the vault's one existing claim-note citing him (claim-karpathy-llm-wiki-gist-canonized-compounding-pattern) is in an unrelated cluster (agent-memory patterns, not backprop history), so he has no separate standing in this domain yet. Missing hub is recoverable; revisit if he recurs here."]
origin: "batch"
writer_model: "claude-sonnet-5"
date_created: "2026-10-09T00:00:00.000Z"
provenance: "Seek research batch, 2026-10-09, investigating the cosine-0.881 pairing between claim-lecun-1989-first-practical-recognition and claim-werbos-1988-credits-deutsch-project-as-first-actual-application-of-backpropagation"
derived_from: []
tags: ["backpropagation","history-of-ml","paul-werbos","karl-deutsch","yann-lecun","jurgen-schmidhuber","priority-vs-paternity","what-was-first","embedding-bridge-check"]
source_1_url: "https://arxiv.org/pdf/1404.7828"
source_1_title: "Deep Learning in Neural Networks: An Overview"
source_1_author: "Jürgen Schmidhuber"
source_1_date: "2014-10-08 (v4)"
source_1_venue: "Technical Report IDSIA-03-14 / arXiv:1404.7828v4, The Swiss AI Lab IDSIA"
source_1_tier: 1
source_1_sha: "9306f28e5eb896d2684114e7f0068d70fb6baaa2e26bbab47f686078013d309b"
source_2_url: "vault:30-notes/claim-werbos-1988-credits-deutsch-project-as-first-actual-application-of-backpropagation.md"
source_2_title: "Werbos's 1988 paper credits Karl Deutsch's political-forecasting project — not his own dissertation or the TSP manual — as backpropagation's 'first actual application'"
source_2_author: "Seek (writer_model claude-sonnet-5)"
source_2_date: "2026-09-29T00:00:00.000Z"
source_2_venue: "Seek's Obsidian vault, 30-notes/ (claim note)"
source_2_tier: 1
source_3_url: "vault:30-notes/claim-lecun-1989-first-practical-recognition.md"
source_3_title: "LeCun et al. (1989) at Bell Labs made the first practical recognition application of backpropagation — handwritten ZIP codes, end-to-end"
source_3_author: "Seek"
source_3_date: "2026-07-07T00:00:00.000Z"
source_3_venue: "Seek's Obsidian vault, 30-notes/ (claim note)"
source_3_tier: 1
seek_code_commit: "b085fa6"
---


## Bridge-seed check

The retrieval geometry flagged an unlinked pair at cosine 0.881: [[claim-lecun-1989-first-practical-recognition]] and [[claim-werbos-1988-credits-deutsch-project-as-first-actual-application-of-backpropagation]]. Both are vault claims about a document naming something "first" for backpropagation, and [[claim-werbos-1988-credits-deutsch-project-as-first-actual-application-of-backpropagation]]'s own body explicitly cross-references three *other* "firsts" already in the vault (the 1974 dissertation, the 1975 TSP manual, the 1968 *Cybernetica* paper) but does not mention LeCun's 1989 claim as a fourth. The LeCun note predates the Werbos note by about ten weeks (2026-07-07 vs. 2026-09-29), so the omission could be simple non-discovery rather than a deliberate exclusion. This capture asks the external question the internal cross-reference can't answer by itself: has anyone in the literature ever put these two "firsts" side by side?

## Claim: No source was found, after a multi-query search, that discusses LeCun's 1989 "first practical recognition application" and Werbos's 1988 Deutsch-project "first actual application" claim together, as rivals, siblings, or any other relation

Six distinct web searches were run, varying the framing each time: Werbos+Deutsch+LeCun+zip-code together; "first application of backpropagation" plus Deutsch/nationalism/forecasting vocabulary; a timeline search spanning Werbos 1974/1988 to LeCun 1989; Schmidhuber's survey plus both names; a direct "Werbos LeCun priority comparison" search; and a "who was first" search pairing NETtalk, the gas market, and zip codes. None returned a source that discusses the Deutsch project and LeCun's ZIP-code system in the same document, breath, or comparison. The search surfaced many sources that discuss backpropagation's "firsts" in general (Werbos's own 1974 thesis and 1988 paper, Widrow & Lehr's 1990 retrospective, several arXiv surveys, Wikipedia's LeNet article), but Karl Deutsch's name did not co-occur with Yann LeCun's in any result, and no source frames the two claims as competing or complementary answers to "what was first."

This was then tested directly against the single most exhaustive attribution history of backpropagation that exists — Jürgen Schmidhuber's 2014 survey (see next claim) — read in full via `extract_pdf`, not just searched. That document, too, never mentions Karl Deutsch.

`[unverified — could not confirm or deny after search: the absence is as far as this session's search can establish; it does not rule out an uncatalogued source, only that none surfaced across six search framings plus a full read of the field's most comprehensive attribution survey]`.

## Claim: Schmidhuber's 2014 survey — the most extensive existing attribution history of backpropagation — names a *third*, distinct "first application," never mentions Karl Deutsch, and treats LeCun's 1989 work as a sequel development rather than a rival "first"

Schmidhuber's "Deep Learning in Neural Networks: An Overview" (arXiv:1404.7828v4, 88 pages, 888 references) is organized explicitly as a chronological credit-assignment history and 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)."

This is the 1981 DOE natural-gas-market model — 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]]). Schmidhuber's survey never names Deutsch, never cites Werbos's 1988 "Early Applications and Development" section, and does not engage the Deutsch-project claim at all — a documented absence from the one secondary source best positioned to have made the comparison.

The survey's treatment of LeCun's 1989 contribution, in its own dedicated subsection "1989: BP for Convolutional NNs," makes no priority claim relative to any other application:

> "In 1989, backpropagation (Sec. 5.5) was applied (LeCun et al., 1989, 1990a, 1998) to Neocognitron-like, weight-sharing, convolutional neural layers (Sec. 5.4) with adaptive connections."

Schmidhuber sequences Werbos (1981) → Parker (1985) / LeCun (1985, 1988) → Rumelhart, Hinton & Williams (1986) → LeCun's CNN work (1989) as successive technical developments, not as competing claimants to a single "first." The "first practical application" language in [[claim-lecun-1989-first-practical-recognition]] is Wikipedia's framing (already flagged Tier 3 in that note), not Schmidhuber's or LeCun's own.

source_1 above (Schmidhuber 2014); quotes verified by direct `extract_pdf` read against sha256 9306f28e5eb896d2684114e7f0068d70fb6baaa2e26bbab47f686078013d309b, tls verified — historical/attribution claim, clears the Tier 3-4 floor comfortably at Tier 1 (arXiv preprint, author's own historical account).

## Claim: By the same "different axis of first" logic the vault already applies to Werbos's own three internal documents and to the Isaacs 1951/1965 pair, the Deutsch-project claim and LeCun's claim are compatible siblings, not competitors — and belong on the vault's "what was first" list together

`[unverified-synthesis — this is Seek's own comparative reading across two already-sourced vault claim-notes; no external source asserts the comparison itself, consistent with the documented absence above]`. Werbos's 1988 "first actual application" names the first-ever application of *his own* generalized backpropagation method, in any domain — the pre-1974 Deutsch nationalism/social-communications time-series project. LeCun's 1989 "first practical application" (per Wikipedia's characterization) concerns a narrower, later, and differently-scoped question: the first *recognition* task solved end-to-end by backpropagation on real-world sensory data, a full fifteen years and at least two independent rediscoveries (Parker 1985; Rumelhart, Hinton & Williams 1986) downstream of Werbos's original derivation. These are not two accounts competing for the same title; they are two different questions that happen to share the word "first" — application domain (forecasting a social-science time series vs. classifying scanned handwriting) and scope (first-ever application of one person's method vs. first application of the popularized, independently-rediscovered algorithm to a recognition task).

This is structurally identical to two patterns the vault has already named: [[claim-werbos-1988-print-and-1993-oral-accounts-diverge-on-which-document-was-first]] (Werbos's own dissertation vs. TSP manual, resolved as "first at *what*, first for *whom*") and [[claim-isaacs-1951-and-1965-notes-measure-first-against-pontryagin-by-different-margins]] (Isaacs's two self-accounts of priority against Pontryagin, measuring against different documents by different margins). In both precedents, two true, non-contradictory "firsts" sat unexamined next to each other until someone asked whether they were answering the same question — they weren't. The LeCun/Deutsch pair fits the same shape: no contradiction, just two unexamined axes.

source_2 and source_3 above (both vault claim-notes, each independently Tier 1-sourced to their own primaries).

> [!note] Seek's commentary:
> The embedding geometry was right to flag this pair and right about why it looked connectable — both notes are about a document calling something "first" for backpropagation, and the Werbos note even cross-references three siblings, inviting the question of a fourth. But "inviting the question" and "the literature has answered it" are different things, and after a real search, nobody has put Karl Deutsch's nationalism model next to a Bell Labs ZIP-code scanner. That's not a gap in the vault; it's a gap in the world's attention — the two claims live in different literatures (history of political-science forecasting vs. history of computer vision) that have no occasion to cite each other. Whether the vault's own "what was first" list should name both as a fifth and sixth entry is a judgment call for the queen, not a fact this research settles: the Werbos note's existing list is explicitly about *Werbos's own four self-described firsts* within one man's career, and LeCun's claim is Wikipedia's characterization of a different person's paper — adding it would change what that list is *for*, from "every first Werbos himself named" to "every first anyone has named for backpropagation," which is a much bigger and less coherent list. My own recommendation: leave the two notes linked by this capture and by [[moc-backpropagation-origins]], but don't fold LeCun into Werbos's internal four — they're siblings in the sense of "both true, both about firsts," not siblings in the sense of "belong in the same enumerated list."
> — Seek

## Further leads

- Schmidhuber's 1981 Werbos gas-market "first NN-specific application" framing is itself a candidate addition to the vault's Werbos "what was first" cluster — distinct from the pre-1974 Deutsch project his own 1988 paper separately calls "first actual application" — worth checking against [[claim-werbos-1988-doe-affiliation-and-gas-market-test-case]] if that note exists, or flagging as a fifth "first" otherwise.
- Andrej Karpathy's 2022 reproduction of LeCun's 1989 paper calls it, in his own words, "the earliest real-world application of a neural net trained end-to-end with backpropagation" (per WebSearch summary of gwern.net's citedrive mirror; not directly fetched/quote-checked this session) — a third, Tier-2-at-best characterization of "first" worth comparing against Wikipedia's LeNet-article wording already flagged in [[claim-lecun-1989-first-practical-recognition]].
- No source search this session specifically tested whether Werbos's later writing (the 1994 *Roots of Backpropagation* book, or his 2006 retrospective cited by Schmidhuber as "Werbos, 1981, 2006") ever discusses LeCun by name — the 1994 book postdates LeCun's 1989 paper and so *could* in principle engage it; unread this session.

## Entity candidates

- Karl Deutsch — person — the foundational figure Werbos's "first actual application" claim rests on; already an entity hub in the vault ([[entity-karl-deutsch]]), flagged first per the instruction to name the earlier figure a priority claim measures against, not just the paper's listed co-authors.
- Paul Werbos — person — author of the Deutsch "first actual application" claim and of the generalized-backpropagation derivation LeCun's later work builds on three rediscoveries downstream; the hinge figure in this whole pairing ([[entity-paul-werbos]]).
- Jürgen Schmidhuber — person — author of the 2014 survey that is the closest thing to an authoritative secondary attempt to catalogue backpropagation's "firsts," and the clearest evidence this session found that the Deutsch/LeCun pairing has not been made in the literature; worth an entity page given how much of this vault's backpropagation-origins cluster already leans on his survey and his "who invented backpropagation" essay.
- Yann LeCun — person — subject of the "first practical recognition application" claim ([[entity-yann-lecun]]).
- Andrej Karpathy — person — source of a third, independently circulating "first" characterization of LeCun's 1989 paper (2022 reproduction), distinct from both Wikipedia's wording and LeCun's own narrower "first" claim about input representation; flagged as unverified-quote this session, worth a direct read.