Jürgen Schmidhuber
Director at IDSIA; LSTM co-inventor; prolific historian-advocate of deep learning's origins (who-invented-backpropagation.html, annotated-history arXiv:2212.11279). For this vault he is primarily a WITNESS — domain-expert, Tier 2 on his own pages — whose historiography is load-bearing for the backprop-origins cluster and must be independently corroborated per the frontmatter caveat.
References
- myth-amari-first-sgd-mlp (the citogenesis case) · myth-werbos-invented-backpropagation · claim-linnainmaa-field-numerical-analysis · claim-linnainmaa-thesis-identity · moc-backpropagation-origins
Updates
- 2026-08-29: His oft-quoted "the chain rule (1676), the first NNs (circa 1800), the first practical AI (1914)…" timeline phrase turns out to be drawn from his Abstract's compressed summary sentence, not from the paper's own section headings (claim-schmidhubers-timeline-phrase-comes-from-abstract-not-section-headers), and he hedges comparable priority claims inconsistently within the same document — unhedged for the 1914 Torres y Quevedo "first practical AI" claim, hedged ("perhaps") for Jacquard's loom programmers (claim-schmidhuber-1914-torres-y-quevedo-claim-unhedged-vs-jacquard-hedge).
- 2026-09-22: a second, different-in-kind instance of the same unflagged pattern — his chain-rule sentence ("It is quite remarkable that he is also responsible for the chain rule, foundation of 'modern' deep learning") states a category projection onto Leibniz's 1676 work with the same unhedged declarative confidence as the document's verifiable dates, never marking "deep learning" as a retrospectively applied label. Read against Nicely's 2011 FAQ, which does draw that line for its own memory-based claims, this is Skinnerian prolepsis (category-projection) left unflagged, distinct from the Torres y Quevedo case (a priority claim inconsistently hedged). See observation-schmidhuber-nicely-cosine-088-pairing-is-analytical-contrast-not-evidentiary-bridge.
- 2026-10-11: his 2014 survey ("Deep Learning in Neural Networks: An Overview," arXiv:1404.7828) served as the load-bearing negative witness in a bridge check — the most exhaustive existing attribution history of backpropagation never mentions Karl Deutsch and never engages Werbos's own 1988 "first actual application" claim, while separately naming its own "first" (Werbos's 1981 DOE gas-market application) for a different, narrower question. A further instance of the entity caveat above: his framing is useful testimony, taken alone, not corroboration of anyone else's claim. (observation-lecun-1989-werbos-1988-deutsch-cosine-0881-pairing-is-analytical-contrast, claim-schmidhuber-2014-names-1981-werbos-gas-market-model-first-nn-specific-application)
- 2026-10-14: the "2006" half of his compressed "Werbos, 1981, 2006" citation is now resolved and read directly — Werbos's AD2004 book chapter, "Backwards Differentiation in AD and Neural Nets" (Springer, 2006), which restates Werbos's own priority narrative but never names LeCun (claim-werbos-2006-ad2004-chapter-never-mentions-lecun).
- 2026-10-24: his 2014 survey functions as the positive channel on the Linnainmaa side, not just a negative witness — a read of Linnainmaa 1976's actual citer list finds the survey (and his "Deep Learning Timeline 1962-2013") is the narrow, identifiable vehicle carrying Linnainmaa's backpropagation priority into mainstream ML, while the bulk of the post-2010 citations are generic applied-AD field growth with no priority-dispute content (claim-schmidhuber-2014-survey-is-channel-for-linnainmaa-name-recognition-in-ml).