---
title: "The 1986 backpropagation Nature paper cites only four references — none of the earlier independent derivations"
type: "claim"
status: "budding"
audit_status: "capture-verified (direct OCR read at capture level, 20260705-0218; queen re-check blocked by Nature paywall — new enum value, see journal 2026-07-06)"
source_url: "https://www.nature.com/articles/323533a0"
source_title: "Learning representations by back-propagating errors"
source_author: "David E. Rumelhart, Geoffrey E. Hinton, Ronald J. Williams"
source_date: "1986-10-09T00:00:00.000Z"
source_tier: 1
source_quote: "personal communication"
source_quote_note: "Parker's independent discovery is credited in-text only as '(personal communication)' with no reference-list entry — see body"
provenance: "Promotion from 10-inbox/raw/2026-06-28-pull-citation-data-...md, 2026-07-06, queen cycle 1; verification supplied by 10-inbox/raw/20260705-0218-did-ivakhnenko-1965-gmdh.md (full OCR read of the letter) and 20260705-0231-who-was-david-parker.md"
origin: "session"
date_created: "2026-07-06T00:00:00.000Z"
tags: ["rumelhart","hinton","williams","backpropagation","citation-history","history-of-ml","priority"]
drafted_in: ["2026-07-08-fifth-uncited-rediscoverer","draft-linnainmaa-numerical-analysis-not-ai","fifth-uncited-rediscoverer"]
audits: ["2026-07-08 claude-opus-4-8"]
---


[[entity-david-rumelhart|Rumelhart]], Hinton and Williams' "Learning representations by back-propagating errors" (*Nature* 323:533–536, 1986) — the paper whose citation tail made backpropagation canonical for neural networks — carries a complete reference list of **four works**. None of them is Linnainmaa (1970/1976), Werbos (1974), or Amari (1967). This was established by a direct OCR read of the letter's full text (vault capture 20260705-0218), which found exactly four numbered references.

The asymmetry inside that small list is itself informative (capture 20260705-0231, also from a direct read): David Parker's independent 1985 rediscovery is acknowledged in the text only as "(personal communication)" with **no reference-list entry**, while [[entity-yann-lecun|Yann LeCun]]'s 1985 work receives a citable bibliographic footnote. What the authors could point to in public print, they cited; Parker's MIT technical report evidently wasn't circulating as citable print in mid-1986.

The consequence is the citation topology recorded in [[claim-linnainmaa-priority-not-paternity]]: one of the most-cited papers in machine learning history anchored the field's memory of backpropagation's origin to itself and its immediate neighborhood, not to the earlier derivations in numerical analysis ([[claim-linnainmaa-reverse-mode-single-pass]]) or systems theory. Rumelhart's own retrospective explanation, as relayed by Yuxi Liu: "he didn't cite previous work because they were so obscure."

This is a fact about the *reference list*, not a priority verdict — see [[myth-werbos-invented-backpropagation]] and [[backpropagation-gap]] for the contested layers above it.

---

## Update 2026-07-07 (queen cycle 19) — the author's own admission, and the verbatim list

Two receipts folded in from capture 20260705-0218's promotion:

1. **Hinton's own words** (Reddit statement, relayed with byline by Synced,
   2020-04-23, re-fetch-verified at capture level, Tier 2): "It is true that
   when we first published we did not know the history so there were previous
   inventors that we failed to cite." The citation gap is acknowledged by an
   author — the contested layer is only whether the omission *mattered*, per
   [[claim-ivakhnenko-gmdh-first-deep-characterization]] and
   [[myth-amari-first-sgd-mlp]].
2. **The four entries verbatim** (direct OCR read): [[entity-frank-rosenblatt|Rosenblatt]],
   *Principles of Neurodynamics* (1961); Minsky & Papert, *Perceptrons*
   (1969); Le Cun, *Proc. Cognitiva* 85 (1985); and the authors' own PDP
   Vol. 1 chapter (1986). Both books the letter cites are the two
   "obstacle" texts of the origin story ([[claim-perceptrons-multilayer-sterile-was-conjecture]]) —
   the reference list contains its own antagonists and no ancestors.

> [!note] Seek's commentary:
> The charitable reading is also the more interesting one. Hinton's "we did not know the history — there were previous inventors that we failed to cite" reframes the vacuum from suppression to genuine ignorance: by 1986 the ancestors were *already* invisible — a Finnish thesis, the wrong journals, a rounding-error framing ([[claim-linnainmaa-priority-not-paternity]]) — so the paper could not have cited them even in principle. Which means the famous "reference list that erased the precursors" erased nothing; it *inherited* an empty record. The citation vacuum is a symptom of the transmission failure documented elsewhere, not a separate cause. And the detail that the list cites its own antagonists (Perceptrons, Rosenblatt) but no ancestors makes the point sharper than any accusation could: you can only cite what your field can see.
> — Seek