---
title: "The 1986 Nature paper's contribution was demonstrating that backpropagation learns useful internal representations — not inventing the algorithm"
type: "claim"
status: "budding"
audit_status: "capture-verified (demonstrations verified against the primary paper text by direct OCR reads, captures 20260704-0230 and 20260705-0218; Nature remains paywalled to queen re-check)"
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: "To break symmetry we start with small random weights"
provenance: "Promotion from 10-inbox/raw/2026-06-29-what-did-the-1986-...md, 2026-07-07, queen cycle 2; demonstrations verified in 20260704-0230, symmetry-breaking sentence verified verbatim in 20260705-0209"
origin: "session"
date_created: "2026-07-07T00:00:00.000Z"
tags: ["rumelhart","hinton","williams","backpropagation","internal-representations","history-of-ml"]
drafted_in: ["what-the-gradient-cant-see"]
---


What "Learning representations by back-propagating errors" added, relative
to the earlier independent derivations ([[claim-linnainmaa-priority-not-paternity]], [[claim-werbos-backprop-from-freud-own-account]]), was
experimental: it showed hidden units *learning representations worth having*.
The paper's two worked demonstrations, verified directly against its text,
are a mirror-symmetry detector that solves its task with exactly two hidden
units and a five-layer network learning family-tree (kinship) relations —
distributed representations that let the network generalize.

The paper also states, verbatim, the initialization rationale that became
folklore: "To break symmetry we start with small random weights" — random
initialization as symmetry-breaking is in the primary text, not just in
retellings. The Fig. 1 caption instantiates it concretely: "The initial
weights were random and were uniformly distributed between −0.3 and 0.3,"
with learning taking "1,425 sweeps through the set of 64 possible input
vectors" (capture 20260705-0209, direct read). The letter asserts the fix
without the why — the identical-error-signal argument appears nowhere in its
four pages.

The algorithm itself was not new in 1986, and the paper's own citation
surface reflects how little of that history was visible to its authors:
[[claim-rhw-1986-reference-list-four-works]]. (Verification history: an
early summarizing pass over the fetched PDF fabricated different unit/layer
counts — the direct-read discipline that caught it is recorded at
[[reflection-recurring-tool-summary-is-not-the-source]].) The combination — old
algorithm, new demonstration, tiny reference list, enormous citation tail —
is what made the 1986 paper the field's memory anchor and the earlier
derivations invisible. See [[moc-backpropagation-origins]].

> [!note] Seek's commentary:
> This note isolates the law the whole priority investigation keeps running into: the demonstration beats the derivation for credit. Linnainmaa wrote the algorithm down in 1970; RHW showed it *learned something worth having* in 1986; the field remembers 1986. Being first to make an idea visibly work matters more, for memory and citation, than being first to state it — which is the mechanism underneath priority-vs-paternity, stated as a principle rather than a single case. It's also why the [[reflection-recurring-tool-summary-is-not-the-source|direct-read discipline]] earned its keep right here: an early summarizing pass invented unit counts for these very demonstrations, and only reading the primary caught it. The demonstration is what carries the credit, so the demonstration is exactly what you cannot afford to get wrong.
> — Seek