Yann LeCun

The vault's anchor for backprop's first practical success: his 1989 work on handwritten-digit recognition (Bell Labs, US Postal Service ZIP codes) is the reference point for "who first applied backpropagation to a real recognition task." He recurs across otherwise-unrelated clusters — DjVu image compression (the 1998 paper co-authored with Bengio, built to distribute scanned documents freely) and Barlow-twins self-supervised loss (redundancy minimization) — which makes him a genuine multi-thread node, not just a backprop name.

References

• claim-lecun-1989-first-practical-recognition · claim-djvu-1998-paper-coauthored-by-lecun-and-bengio · claim-barlow-twins-loss-minimizes-embedding-redundancy · claim-rhw-1986-reference-list-four-works
• Captures: 2026-06-30-who-first-applied-backpropagation-to-a-practical-recognition-task-and-when, 2026-07-09-which-appeared-first-lecun-et-als-nips-1989

Updates

• 2026-08-13: LeCun's own words (re-fetched directly and verified verbatim, resolving a prior [unverified-quote] flag) confirm DjVu's purpose was cheap distribution of scanned documents, demonstrated by freely releasing the NIPS proceedings at nips.djvu.org in 2000 with publisher permission — the demo site is now offline (DNS failure), but the free-access mission it demonstrated migrated to NeurIPS's own official site. (claim-djvu-created-to-freely-distribute-scanned-documents, claim-nips-djvu-org-domain-no-longer-resolves-2026, claim-neurips-absorbed-free-nips-proceedings-into-papers-nips-cc)
• 2026-10-11: a bridge check against Paul Werbos's 1988 "first actual application" claim found no source puts the two "firsts" side by side; Schmidhuber's 2014 survey — the field's most exhaustive existing attribution history — sequences LeCun's 1989 CNN work as a successive technical development after Werbos (1981), Parker (1985), and RHW (1986), making no priority claim of its own relative to any earlier application, consistent with but distinct from the Wikipedia "first practical application" wording already on this page. (observation-lecun-1989-werbos-1988-deutsch-cosine-0881-pairing-is-analytical-contrast)
• 2026-10-14: Werbos's own later writing engages LeCun directly, if unevenly — his 2006 AD2004 retrospective never names him, but his 2004 "Object Nets" patent does, twice, characterizing LeCun's 1985 derivation of backpropagation as reached via "duality theory from control theory (Bryson and Ho, 1969)" and contrasting it with Werbos's own 1974 derivation — an attribution made by an interested party, not yet checked against LeCun's own 1985 paper. (claim-werbos-2004-object-nets-patent-names-lecun-against-1994-book, question-does-lecuns-1985-derivation-match-werboss-duality-theory-characterization)
• 2026-10-16: his own research page, read directly, weighs against that patent's attribution — he describes his 1985 method as propagating "virtual target values," not gradients, with no control-theory framing stated, and credits the rigorous Lagrangian (duality-adjacent) derivation specifically to his 1988 paper, three years later. The 1985 primary text itself remains unlocated and unread, so the question stays open, but the likeliest reading is that Werbos's patent compresses two LeCun publications into one citation. (claim-lecun-1985-virtual-target-values-not-gradients, claim-lecun-1988-lagrangian-derivation-not-1985)