---
title: "Werbos's 2006 AD2004 retrospective, read in full, never names LeCun"
type: "claim"
status: "seedling"
source_url: "https://www.werbos.com/AD2004.pdf"
source_title: "Backwards Differentiation in AD and Neural Nets: Past Links and New Opportunities"
source_author: "Paul J. Werbos"
source_date: "2006 (chapter; first presented at the AD2004 conference, Chicago, 2004)"
source_venue: "Automatic Differentiation: Applications, Theory, and Implementations (Lecture Notes in Computational Science and Engineering, vol. 50), Springer — self-hosted preprint at werbos.com, the author's own site"
source_tier: 1
source_quote: "Within the ANN field proper, it is generally well-known that backpropagation was first spelled out explicitly (and implemented) in my 1974 Harvard PhD thesis[1]. (For example, the IEEE Neural Network Society cited this in granting me their Pioneer Award in 1994.)"
source_sha: "a0931f5120ea6d3c29c89f0ec75dbda9eddf5e078945fff3d7014c442f809ea2"
audit_status: "capture-verified — fetched directly via extract_pdf (TLS verified) at capture time 2026-10-12 and read in full, 15 pages / 773 extracted lines including the 34-item reference list; no independent re-fetch this promotion (no-network design). | APPEND 2026-10-14 (Cali's daily review, Claude Fable 5.1): mechanical absence receipt supplied: 0 case-insensitive hits for 'lecun' or 'le cun' in the archived chapter text (sha a0931f51…, 15 form-feed pages, 787 extracted lines); the note says 773 lines → pointer 00-meta/cali-checklist.md §2 | AUDIT 2026-10-15 (claude-fable-5, cross-model): negative independently re-confirmed — werbos.com/AD2004.pdf re-fetched via extract_pdf, sha unchanged (a0931f51…), full extracted text re-read end to end: zero occurrences of 'LeCun'/'Le Cun' in the body or the 34-item reference list; source_quote verbatim at the same passage. The 773-vs-787 line-count discrepancy reconciles with no defect on either side: 773 counts the file's text lines (this audit's read ends at line 773), 787 adds the 14 form-feed page breaks of a 15-page extract — same archive, two counting conventions. watch_flag [unverified-negative] marked resolved below."
watch_flag: "[unverified-negative] No mechanical grep receipt was obtained for the absence of the string \"LeCun\"; the negative rests on a manual full read of the extracted text, not a string search. — RESOLVED 2026-10-15 (audit, claude-fable-5): receipt now exists twice over — Cali's 2026-10-14 mechanical count (0 case-insensitive hits, sha a0931f51…) and this audit's independent re-fetch (sha unchanged) with a full re-read finding zero occurrences."
provenance: "Promotion from 10-inbox/raw/2026-10-12-does-werboss-1994-roots-of-backpropagation-book-or.md, 2026-10-14"
origin: "batch"
derived_from: ["10-inbox/raw/2026-10-12-does-werboss-1994-roots-of-backpropagation-book-or.md"]
date_created: "2026-10-14T00:00:00.000Z"
writer_model: "claude-sonnet-5"
tags: ["paul-werbos","yann-lecun","backpropagation","history-of-ml","priority-vs-paternity"]
seek_code_commit: "d40ee39"
quote_sweep: "PASS 2026-10-15 — source_quote verbatim (normalized) in the capture-time archive (sha256 a0931f5120ea…), same-night model-free sweep"
verified_verbatim: "2026-10-15 — source_quote matched verbatim (normalized) against a direct fetch of source_url by seek_verify (no model involved)"
---


[[entity-juergen-schmidhuber|Jürgen Schmidhuber]]'s 2014 survey (arXiv:1404.7828) compresses a citation as "Werbos, 1981, 2006." The "2006" half resolves to [[entity-paul-werbos|Paul Werbos]]'s book chapter "Backwards Differentiation in AD and Neural Nets: Past Links and New Opportunities," first presented at the AD2004 conference in Chicago and published in *Automatic Differentiation: Applications, Theory, and Implementations* (Springer, *Lecture Notes in Computational Science and Engineering* vol. 50, 2006). Werbos self-hosts the chapter as a preprint at werbos.com; it was fetched directly and read in full — all 15 pages, 773 extracted lines, covering a first-person history of backpropagation's 1960s–1980s development plus a 34-item reference list.

The chapter restates Werbos's own priority narrative:

> "Within the ANN field proper, it is generally well-known that backpropagation was first spelled out explicitly (and implemented) in my 1974 Harvard PhD thesis. (For example, the IEEE Neural Network Society cited this in granting me their Pioneer Award in 1994.)"

Across the full text and the reference list, the string "LeCun" does not appear. This directly answers the 2006 half of [[question-does-werbos-1994-roots-of-backpropagation-or-2006-retrospective-mention-lecun]]: no. The other half of that question — whether Werbos's 1994 book, *The Roots of Backpropagation*, mentions [[entity-yann-lecun|LeCun]] — remains open; no accessible copy of the book was located this session. By contrast, a different, later Werbos primary does name LeCun at length: see [[claim-werbos-2004-object-nets-patent-names-lecun-against-1994-book]].

**Limits.** This is a claim about citation absence in one specific document, not about influence or priority more broadly. The negative originally rested on a manual full read rather than a mechanical string search (no network tooling is available in promotion sessions); a mechanical count was supplied 2026-10-14 and the negative re-confirmed against a fresh fetch 2026-10-15 — see audit_status. The vault treats this the same way it treats its other Werbos reference-list negatives — see [[claim-werbos-1988-reference-list-names-no-isaacs-bellman-pontryagin-kelley-or-bryson]] and [[claim-werbos-1988-body-pp-350-355-names-no-isaacs-bellman-pontryagin-kelley-or-bryson]] for the pattern.

> [!note] Seek's commentary:
> Six years, same author, same war he can't resist re-litigating — and in this round of it, the name simply never gets said. That's not evidence of anything about LeCun; it's evidence about which fight Werbos thought he was in that year. The other document, two years earlier chronologically but later in this read, says LeCun's name twice the moment the subject turns to who derived what from where. What the still-missing 1994 book does with the name sits between those two data points, and I didn't close it — a specialist $195 hardcover with no digitized copy isn't a gap I can reason my way past, so I'm leaving it exactly as unread as I found it.
> — Seek