Paul Werbos

Harvard PhD (1974, applied mathematics, under Karl Deutsch and Yu-Chi Ho); derived backpropagation independently while trying to give mathematical form to Freud's model of psychic energy, and applied it to political-science forecasting in the accepted dissertation. One of several independent originators of the reverse mode; disputes the Bryson–Ho lineage claim in his own oral-history account. His 1968 Cybernetica paper (self-described earliest germ) entered the inbox 2026-07-06, unpromoted.

References

• myth-werbos-invented-backpropagation · backpropagation-gap · claim-linnainmaa-priority-not-paternity · claim-rhw-1986-reference-list-four-works

• Captures: 2026-06-29-who-was-paul-werbos, 20260706-0225-does-werboss-1968-cybernetica

• 2026-07-29: his peer-reviewed 1988 paper ("Generalization of Backpropagation with Application to a Recurrent Gas Market Model," Neural Networks 1:339–356) states in his own words that RHW's 1986 feedforward algorithm is "a special case of equations 7 and 8" of his own 1974 "dynamic feedback" framework — a formal priority argument made in print, not just in oral history (claim-werbos-1988-rhw-special-case-of-dynamic-feedback).

• 2026-07-31: a previously undocumented DOE/EIA career leg — that same 1988 paper's byline lists his affiliation as the U.S. Department of Energy, not a university, and its real-world test case was a DOE-developed Natural Gas Market Model (claim-werbos-1988-doe-affiliation-and-gas-market-test-case). He was not merely using that model: he was its named federal documentarian, having authored EIA's official November 1981 "equations and data sources" report on it, three years earlier, for the study evaluating the Natural Gas Policy Act of 1978's price decontrol (claim-werbos-authored-eia-natural-gas-market-model-documentation-1981).

• 2026-08-01: his own account of DARPA's mid-1970s demand for "crisis warning," not "pure theory" — the pressure behind the DOCSUB block — matches Larry Roberts's independently documented account of DARPA-wide pressure on program managers after the 1969 Mansfield Amendment, connecting his backpropagation-priority story to the vault's separate Vietnam-era defense-funding-politics cluster for the first time (claim-darpa-crisis-warning-pressure-on-werbos-matches-roberts-account).

• 2026-08-01: the advisor named on this page, Karl Deutsch, did more than supervise — after the committee rejected the Freud-based thesis he personally supplied the political-forecasting problem ("what causes war and peace between nations") that became the accepted 1974 "Beyond Regression," and solving it under a compute budget is Werbos's own stated reason backpropagation was generalized to recurrent form (claim-deutsch-redirected-werbos-thesis-generalized-backprop-to-recurrent).

• 2026-08-07: his two "missing document" stories are not independent sources of each other — the lost 1968 Cybernetica scan and the DARPA report blocked from DOCSUB both come from the same June 1993 Talking Nets interview (ch. 15, thirteen pages apart), and his own werbos.com retrospective explicitly cross-references that interview; what read as two corroborated episodes is one witness telling one story twice (observation-werbos-two-missing-documents-share-one-1993-witness).

• 2026-08-17: a bridge check found that the DOCSUB block and RHW's 1986 citation gap are not causally linked — the DOCSUB story entered the public record 7–12 years after the Nature paper, and the blocked DARPA report is a different document from the 1974 Harvard thesis RHW would have needed to cite (claim-docsub-episode-postdates-rhw-1986-nature-paper-by-years, claim-werbos-docsub-report-and-rhws-missing-1974-thesis-are-different-documents). His own explanation for his broader early obscurity, offered unprompted in the same interview, names a different mechanism than suppression: "I tried to cram a lot into individual papers... I think that's one of the reasons they weren't widely recognized," corroborated by Laveen Kanal's independent 1993 recollection that a title/framing problem, not DARPA, buried the 1974 report on his own desk (claim-werbos-attributes-early-obscurity-to-publishing-strategy-not-suppression).

• 2026-08-22: the shared-witness structure extends to a third leg — the DOCSUB/thesis document-identity claim also rests on the June 1993 Talking Nets sitting (the "which document" evidence sits at pp. 348–351 of the same chapter), making three "missing document" legs from one interview rather than two. It is, however, the only one of the three with independent corroboration: his own CV supplies the disambiguating 1974-PhD / 1975–1978-DARPA chronology without the interview's help. So the retrieval bridge between the document-identity note and the two-document shared-witness note is a real shared-source dependency, not a false-friend cosine (observation-werbos-document-identity-is-corroborated-third-leg-of-1993-witness-cluster).

• 2026-09-17 (propagation-repair): the 2026-08-07 line above was corrected in place. Was: the two "missing document" passages sit "sixteen pages apart". Now: "thirteen pages apart" (Cybernetica/Freud passage on p. 338, DOCSUB passage on p. 351, in a chapter that runs pp. 335–358), following the 2026-08-08 audit correction in observation-werbos-two-missing-documents-share-one-1993-witness. The shared-witness claim is unchanged.

• 2026-09-19: a retrieval probe paired his own "publishing strategy, not suppression" account against Rodney Brooks's dropped-citation episode at cosine 0.85; a targeted search found no documented contact between the two men — different decades, subfields, and mechanisms behind superficially matching "obscurity" vocabulary. Confirmed as an embedding false friend; see observation-brooks-werbos-cosine-085-pairing-is-embedding-false-friend.

• 2026-09-19: a separate bridge check gave his optimization work a role the vault hadn't recorded on this page — his ADP "optimization as a foundation for understanding consciousness and the mind" bet (claim-werbos-1997-optimization-consciousness-chapter, claim-optimization-as-theory-of-mind-two-lineages) sits at one pole of a three-way structure whose opposite pole is Gigerenzer's "ecological rationality is not optimization," with Anderson's rational analysis as the load-bearing hinge between them. No direct Werbos–Gigerenzer citation traffic was found; the link is mediated, not a direct exchange (observation-optimization-lineages-gigerenzer-pairing-is-confirmed-bridge-via-rational-analysis).

• 2026-09-22: a bridge check against Greenhalgh's 2003 China-population-policy article (cosine 0.84, via the TSP manual's sole 1977-dated citation in claim-klensin-romberg-1989-sole-source-of-tsp-manual-1977-date) found no connection in either direction — a targeted search for any link between Klensin, Romberg, Brode, Dunn, or the MIT Cambridge Project and Song Jian, Kwakernaak, Greenhalgh, or Chinese population-policy cybernetics returned nothing. Confirmed as an embedding false friend, the third independent test of this same US/UK-computing-vs-Dutch-Chinese-population-policy crossing to return that verdict (observation-greenhalgh-2003-klensin-romberg-1989-cosine-pairing-is-embedding-false-friend, observation-greenhalgh-olsder-werbos-crossing-confirmed-false-friend-three-independent-tests). The TSP-manual dating dispute this bridge check turned on — his own thrice-repeated 1975 self-citation against Klensin & Romberg's unsourced 1977 — remains open at question-werbos-tsp-manual-year-1975-vs-1977.

• 2026-09-22: the DOCSUB-chronology claim gained a role beyond the RHW-citation question this page already tracks — a bridge check against an unrelated NASA ozone-anomaly dating claim (cosine 0.84) found both perform the identical evidentiary move, making DOCSUB a genuine, previously unflagged member of the vault's diplomatics cluster (dated-primary-source comparison overruling a received attribution), the same genre Mabillon founded in 1681 (observation-docsub-and-mcpeters-are-unflagged-diplomatics-cluster-members).

• 2026-09-25: In the 1988 paper he equates his "dynamic feedback" with control theory's "time-dependent Lagrange multipliers" and calls it "very similar to earlier methods from control theory." His reference list names no Isaacs, Bellman, Pontryagin, Kelley, or Bryson; the documented control-theory lines are Athans & Kalb (1966) and Oak Ridge adjoint work. (claim-werbos-1988-equates-dynamic-feedback-with-control-theorys-time-dependent-multipliers, claim-werbos-1988-reference-list-names-no-isaacs-bellman-pontryagin-kelley-or-bryson)

• 2026-10-03: a bipartite hop checking whether the cosine pairing between observation-ml-ad-en-gedi-cosine-pairing-is-embedding-false-friend and observation-brooks-werbos-cosine-085-pairing-is-embedding-false-friend was a real meta-bridge ruled it a false friend, but reading past that verdict found he himself is a real, partial person-bridge one level down: the same man appears both as one of Griewank's five independent rediscoverers of reverse-mode differentiation (the structural evidence behind the vault's ML/AD mutual-unawareness claim) and as the protagonist of his own obscurity testimony (this page's 2026-08-17 entry) — two facts about him the vault had logged separately and never connected until now (observation-werbos-person-bridges-ml-ad-mutual-unawareness-and-his-own-obscurity-testimony).

• 2026-10-04 (promotion of a recovered capture): his obscurity testimony was also cosine-paired (0.83) against Isaacs's P-257 no-colleagues silence — a third distinct pairing against his testimony beyond Brooks (0.85) and the ML/AD-mutual-unawareness meta-pairing (above). Staying strictly within the two seed claim-notes, this is a plain embedding false friend, same shape as Brooks: his testimony is positive self-explanation, Isaacs's silence is weak on McGrew's notice term, and the two share a genre (own-voice account tested against a received story) without sharing a referent. But a separate, already-promoted capture chasing Isaacs's side past P-257 to his 1965 book preface found the two men's own self-explanations really do converge — both reach for mundane circumstance, not suppression, as the reason for their own reception gaps (observation-isaacs-and-werbos-both-reframe-publication-gaps-as-circumstance-confirmed-bridge). The pairing is false friend and confirmed bridge at once, depending on how far past the seed claim the chase goes — a scope-dependence not seen in his other false-friend pairings. (observation-werbos-isaacs-cosine-083-pairing-verdict-depends-on-which-primary-is-chased)

• 2026-10-11: a fifth, independently-sourced "first" entered the vault's running count of his career's priority claims — Schmidhuber's 2014 survey, in his own historical voice rather than Werbos's, names the 1981 DOE gas-market application (not the pre-1974 Deutsch project) "the first NN-specific application of efficient BP" (claim-schmidhuber-2014-names-1981-werbos-gas-market-model-first-nn-specific-application). The same bridge check found that survey — the field's most exhaustive existing attribution history — never engages his own "first actual application" (Deutsch) claim, and that no source anywhere puts his Deutsch-project claim next to LeCun's 1989 "first practical application" claim (observation-lecun-1989-werbos-1988-deutsch-cosine-0881-pairing-is-analytical-contrast).

• 2026-10-14: resolved the "Werbos, 1981, 2006" citation Schmidhuber's survey compresses — it is his AD2004 book chapter, read in full, which restates the 1974-thesis priority claim but never names LeCun (claim-werbos-2006-ad2004-chapter-never-mentions-lecun). A different, later primary — his 2004 "Object Nets" patent (US 6,708,160 B1) — does name LeCun twice, contrasting a "duality theory" derivation it attributes to LeCun's 1985 paper with his own, broader 1974 derivation (claim-werbos-2004-object-nets-patent-names-lecun-against-1994-book). His 1994 book, The Roots of Backpropagation, remains unlocated and unread by the vault — a manual-consultation item.

• 2026-10-16: that patent's specific attribution looks weaker when checked against LeCun's own account of his own work — LeCun describes his 1985 paper as a non-gradient "virtual target values" method and credits the rigorous Lagrangian/duality-adjacent derivation to his separate 1988 paper. The likeliest reading is that his patent's "LeCun (1985)" citation compresses two of LeCun's publications, three years apart, into one — a small, specific error-shape in a document written to argue his own 1974 priority, not yet confirmed against the unread 1985 primary itself. (claim-lecun-1985-virtual-target-values-not-gradients, claim-lecun-1988-lagrangian-derivation-not-1985)