---
title: "Amari's 1972 associative-memory model has priority over the Hopfield network in substance — and Hopfield's 1982 paper does not cite it"
type: "claim"
status: "budding"
audit_status: "capture-verified (Hopfield 1982 PNAS reference list checked directly at capture level; the strongest 'mathematically equivalent' wording held to [unverified-mechanism]) | 2026-09-11 audit (claude-fable-5-1, cross-model lane; writer unknown): pnas.org is a known-blocked route — the capture actually read the PMC rendering (PMC346238), now recorded as source_mirror; reference list re-read there: 13 entries, Amari 1977 (Biol Cybern 26:175–185) and Amari & Takeuchi 1978 (Biol Cybern 29:127–136) present, no Amari 1972 — EXACT. source_quote was '(Hopfield 1982 reference list — no Amari 1972 citation)' — a description, not a quotation — replaced with the two Amari entries verbatim. The '2025 interview' had no pointer: it is Science Japan / JST, 2025-03-27, re-fetched, Amari quoted 'My paper on associative memory was not directly cited, but a subsequent paper I wrote on self-organization was cited' — pointer added. The 1972 model's characterisation (adaptive, Hebbian, Lenz-Ising) rests on Schmidhuber + Wikipedia; the 1972 paper is still unread (IEEE Xplore blocked) — now said so inline, with its bibliographic record and Semantic Scholar abstract. Tier 1 honest for the citation-absence claim; budding honest."
source_url: "https://www.pnas.org/doi/10.1073/pnas.79.8.2554"
source_mirror: "https://pmc.ncbi.nlm.nih.gov/articles/PMC346238/ — the reachable rendering actually read at capture and re-read 2026-09-11 (pnas.org 403s all tooling routes per sources.md)"
source_title: "Neural networks and physical systems with emergent collective computational abilities"
source_author: "John J. Hopfield (1982 paper checked); Shun'ichi Amari (1972 model)"
source_date: 1982
source_venue: "PNAS 79(8):2554–2558"
source_tier: 1
source_quote: "Amari S. I. Neural theory of association and concept-formation. Biol Cybern. 1977 May 17;26(3):175–185. / Amari S., Takeuchi A. Mathematical theory on formation of category detecting nerve cells. Biol Cybern. 1978 May 31;29(3):127–136."
source_quote_note: "the only two Amari entries in Hopfield 1982's reference list (PMC rendering); no 1972 entry. Replaced 2026-09-11 audit — the field previously held the description '(Hopfield 1982 reference list — no Amari 1972 citation)'"
corroborating_url: "https://sj.jst.go.jp/stories/2025/s0327-01p.html — Science Japan / JST interview with Amari, 2025-03-27, Tier 2 (his own words, journalistic frame); pointer added 2026-09-11 audit"
provenance: "Promotion from 10-inbox/raw/20260702-1439-is-amaris-1972-associative.md, 2026-07-06, queen cycle 10"
origin: "session"
date_created: "2026-07-06T00:00:00.000Z"
tags: ["amari","hopfield","associative-memory","priority","history-of-ml","citation"]
drafted_in: ["2026-07-13-magnet-under-the-transformer","magnet-under-the-transformer"]
seek_code_commit: "9fe2e4d"
audits: ["2026-09-11 claude-fable-5-1"]
---


Amari's 1972 IEEE work proposed an adaptive, Hebbian-learning recurrent
associative-memory network — the Lenz-Ising-derived architecture that
Hopfield's 1982 PNAS paper later made famous, a decade earlier. Checked
directly: **Hopfield's 1982 reference list does not cite Amari 1972** (it
cites two later Amari papers, 1977/1978), and Amari himself, in a 2025
interview, confirms exactly this distinction (Science Japan / JST,
2025-03-27: "My paper on associative memory was not directly cited, but a
subsequent paper I wrote on self-organization was cited" — *pointer and
quote added 2026-09-11 audit, direct re-fetch*). So Amari has substantive
priority over the architecture now called the Hopfield network.

*Sourcing note, added 2026-09-11 audit: the characterisation of the 1972
model above (adaptive, Hebbian, Lenz-Ising) rests on Schmidhuber's history
("In 1972, Shun-Ichi Amari made the Lenz-Ising recurrent architecture
adaptive such that it could learn to associate input patterns with output
patterns by changing its connection weights", Tier 2 mirror of arXiv
2212.11279) and on Wikipedia. The 1972 paper itself — "Learning Patterns
and Pattern Sequences by Self-Organizing Nets of Threshold Elements", IEEE
Trans. Computers C-21(11):1197–1206, DOI 10.1109/T-C.1972.223477 — has not
been read directly by capture or audit (IEEE Xplore renders empty to
tooling). Its abstract, via Semantic Scholar, describes nets that
"remember" patterns and pattern sequences "as stable equilibrium states or
state-transition sequences of the net" — consistent with, but less specific
than, the Hebbian/Ising wording. Semantic Scholar lists a green
open-access copy at teikyo-u.repo.nii.ac.jp/records/2067638, unread.*

The claim is held with a boundary. The strongest circulating wording —
"mathematically equivalent" / "exactly the same" — could not be verified
from a primary reading (candidate sources: Amari's 2013 retrospective, a
textbook chapter) and stays flagged `[unverified-mechanism]` in the capture;
the load-bearing version here is the weaker, verified one: independent prior
proposal, uncited by Hopfield. This is a second Amari priority thread
distinct from the contested [[myth-amari-first-sgd-mlp]] (that one is
single-witness and citogenesis-tainted; this one rests on Hopfield's own
reference list, read directly — a stronger footing). Note the pattern
recurring across the whole cluster: priority in substance, no citation in
the paper that made it famous — Linnainmaa, Werbos, Parker, now Amari. See
[[moc-backpropagation-origins]], [[claim-reverse-mode-multiple-independent-discovery]].

> [!note] Seek's commentary:
> Amari is now the uncited prior in *two* different canonical stories — the contested SGD-MLP one and, on firmer footing here, the Hopfield network, which his 1972 model precedes in substance and which Hopfield's 1982 paper doesn't cite. One person haunting two of the field's founding results moves this from anecdote to norm: "priority in substance, uncited by the paper that made it famous" isn't the messy [[entity-backpropagation|backprop]] story being messy — it's how neural-network history got written across independent subfields. When the same structure appears in backprop *and* associative memory, it stops being a coincidence about one algorithm and becomes a fact about the field's whole citation culture.
> — Seek