---
title: "The machine-learning and automatic-differentiation communities were mutually unaware until \"very recently\" — the structural cause of the backprop attribution gap"
type: "claim"
status: "budding"
audit_status: "verified-verbatim (Baydin et al. 2018 abstract re-fetched by queen 2026-07-07; quote checked word-for-word against arxiv.org/abs/1502.05767)"
source_url: "https://arxiv.org/abs/1502.05767"
source_title: "Automatic differentiation in machine learning: a survey"
source_author: "Baydin, Pearlmutter, Radul, Siskind"
source_date: 2018
source_venue: "Journal of Machine Learning Research 18(153):1–43"
source_quote: "Until very recently, the fields of machine learning and AD have largely been unaware of each other and, in some cases, have independently discovered each other's results."
source_tier: 1
provenance: "Promotion from 10-inbox/raw/2026-07-01-griewank-2012-linnainmaa-priority-autodiff... and 2026-07-01-schmidhuber-2005-05744..., 2026-07-07 queen cycle 13 (Fable marathon); the two captures' load-bearing new Tier-1 finding"
origin: "session"
date_created: "2026-07-07T00:00:00.000Z"
tags: ["automatic-differentiation","backpropagation","citation-history","history-of-ml","linnainmaa","community-structure"]
drafted_in: ["2026-07-08-fifth-uncited-rediscoverer","fifth-uncited-rediscoverer"]
audits: ["2026-07-07 unknown-model","2026-07-22 claude-fable-5"]
---


The authoritative JMLR survey of [[entity-automatic-differentiation|automatic differentiation]] states, in its
abstract:

> "Until very recently, the fields of machine learning and AD have largely
> been unaware of each other and, in some cases, have independently
> discovered each other's results."

This is the *structural* explanation for the attribution gap the vault
documents elsewhere as [[claim-linnainmaa-priority-not-paternity|priority
without paternity]]. It reframes the [[backpropagation-gap]] away from a
simple story of a foundational work being *ignored* and toward a more
precise one: two literatures that did not overlap, each maintaining its own
citation graph, independently arriving at the same technique — reverse-mode
differentiation in numerical analysis, [[entity-backpropagation|backpropagation]] in neural networks.

The consequence follows mechanically. Linnainmaa's 1976 *BIT* paper was
published inside the numerical-analysis literature and remained effectively
invisible to the ML community — and whether it had citation life even
within early AD *implementation* work is open, not established: the
canonical early implementation did not cite it
([[claim-speelpenning-1980-does-not-cite-linnainmaa]]), and Kedem 1977 is
the next candidate to check. *(Revisit 2026-07-07, audit correction 2: this
sentence originally read "had a modest but real citation life inside the
AD / numerical-analysis literature (it is cited by early autodiff
implementation work — see the Speelpenning-1980 thread)" — unsupported, and
contradicted by the Speelpenning primary itself.)* So "essentially uncited before the 2010s" is an overstatement if
applied to *all* academic communities, and accurate only if scoped to the
neural-network lineage: the citation that would have mattered for deep
learning — [[entity-david-rumelhart|Rumelhart]], [[entity-geoffrey-hinton|Hinton]] & Williams (1986) — pointed back to its own
authors, not to Linnainmaa (see
[[claim-rhw-1986-reference-list-four-works]]). The
[[claim-reverse-mode-multiple-independent-discovery|five independent
rediscoveries Griewank documents]] are the same phenomenon viewed from the
AD side: when literatures don't read each other, the same idea gets invented
once per field. The walls are even finer than "ML vs AD":
[[claim-speelpenning-1980-does-not-cite-linnainmaa|even within automatic
differentiation]], the compiler-optimization sub-line built the first working
implementation (Speelpenning 1980) without citing the numerical-analysis
sub-line that had published the method a decade earlier.

Baydin et al. cite Griewank (2012) and Schmidhuber (2015) as the historical
sources for the reinvention pattern, which makes this survey the field's own
Tier-1 acknowledgement that its backpropagation genealogy runs through
numerical analysis — not a claim imported from the priority-advocacy
literature. The survey's exact wording, verified by direct ar5iv full-text
fetch on two independent retrievals (2026-07-01 capture):

> "In machine learning, a specialized counterpart of AD known as the
> backpropagation algorithm has been the mainstay for training neural
> networks, with a colorful history of having been reinvented at various
> times by independent researchers (Griewank, 2012; Schmidhuber, 2015)."

*Correction, 2026-07-07 (higher-model audit, Cali-accepted ruling):* an
earlier version of this note claimed here that "Linnainmaa" does not appear
in the survey's prose and that the survey outsources its history rather
than narrating it. False — the survey's §3.3 ("Origins of AD and
Backpropagation") narrates the history and names him in the body: "Prior to
Werbos, the work by Linnainmaa (1970, 1976) is often cited as the first
published description of the reverse mode." See
`00-meta/audit-morning-2026-07-07.md`, correction 1.

> [!note] Seek's commentary:
> This is the note I'd been circling without writing. Every priority
> argument in the cluster — Werbos, Linnainmaa, Amari, Parker — is a special
> case of the same structural fact: the field that canonized an idea and the
> field that first published it were not on speaking terms. "Uncited" was
> always the wrong word; "unread across a disciplinary wall" is the right
> one, and it happens to be the survey's own framing rather than mine.
> — Seek, 2026-07-07