---
id: "20260701-1639-does-baydin-et-al"
title: "Does Baydin et al. (2018) JMLR survey explicitly discuss Linnainmaa's citation history?"
type: "capture"
status: "promoted"
promoted_to: ["claim-ml-and-ad-communities-mutually-unaware"]
not_promoted: ["the core negative finding (Baydin does not discuss Linnainmaa's citation history in prose) is not itself a shelf-worthy claim — it is folded into the note as the 'survey outsources the history' point; the 'colorful history' body quote was folded as verified corroboration"]
origin: "batch"
model: "claude-sonnet-5"
date_created: "2026-07-01T00:00:00.000Z"
provenance: "batch run, 2026-07-01"
derived_from: []
tags: ["automatic-differentiation","backpropagation","linnainmaa","citation-history","historiography-of-ml","baydin-2018"]
source_url: "https://ar5iv.labs.arxiv.org/html/1502.05767"
source_author: "Baydin, Pearlmutter, Radul, Siskind"
source_date: "2018-02-05T00:00:00.000Z"
source_tier: 1
---


## Core question

Does Baydin, Pearlmutter, Radul & Siskind, "Automatic Differentiation in Machine Learning: a Survey" (JMLR 18(153):1–43, 2018; arXiv:1502.05767) **explicitly discuss [[Seppo Linnainmaa]]'s citation history** — i.e., does it narrate when his 1970/1976 work was recognized, ignored, or rediscovered?

**Finding: No — not in the portions of the paper that could be directly retrieved and verified.** The survey's own text does not appear to contain a discussion of Linnainmaa's citation history at all. It cites Linnainmaa's 1976 paper in its bibliography (a standard reference), but the prose discussion of "why backpropagation has a tangled history" in the introduction cites two *other* papers — Griewank (2012) and Schmidhuber (2015) — as the sources for that history, rather than narrating it itself.

## What was directly verified (Tier 1 — primary text)

Multiple independent fetches of the paper's own text (arXiv HTML mirror at ar5iv, and the JMLR/arXiv PDF) were cross-checked. Two separate fetch passes of the introduction returned the identical sentence verbatim:

> "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)."

— Baydin et al. 2018, Introduction. source_url: https://ar5iv.labs.arxiv.org/html/1502.05767 · source_tier: 1 (direct fetch of the paper's own text, confirmed on two separate retrievals with identical wording).

This is the *only* passage found in the retrievable portion of the paper that touches on the "reinvention" history of backpropagation/reverse-mode AD. Notably, **it does not name Linnainmaa** — it defers the historical account to Griewank (2012) and Schmidhuber (2015) by citation rather than summarizing it in-text.

A follow-up direct query against the same fetched text ("does the word 'Linnainmaa' appear anywhere in this document at all — yes or no?") returned: **"No, this word does not appear anywhere in the provided document."** This covered the paper from the abstract/introduction through at least a section containing "Table 3" (i.e., well into the paper's methods/background material, including the passage discussing Wengert (1964) and Rumelhart et al. (1986) by name). source_url: https://ar5iv.labs.arxiv.org/html/1502.05767 · source_tier: 1.

Caveat on coverage: the fetch tooling used truncates very long pages, and the retrieval was confirmed complete only through roughly the paper's background/methods sections (up to a table captioned "Table 3"), not verified page-by-page through the conclusion and full bibliography list. Genuine effort was made to retrieve the remainder (direct PDF fetch, alternate mirrors, a proxy reader) but these either failed (HTTP 401/404), returned unparsable binary/compressed stream data, or hit the same truncation point. So it cannot be stated with 100% certainty that no later section (e.g., a conclusion or appendix) ever mentions Linnainmaa by name — only that no such mention was found in the roughly first half-to-two-thirds of the paper across multiple independent query angles.

## Bibliography fact (Tier 2 — structured metadata)

The Semantic Scholar API's reference list for this paper (arXiv:1502.05767) does include Linnainmaa's 1976 paper as a cited reference:

> {"title": "Taylor expansion of the accumulated rounding error", "year": 1976, "authors": [{"name": "S. Linnainmaa"}]}

source_url: https://api.semanticscholar.org/graph/v1/paper/arXiv:1502.05767/references — source_tier: 2 (structured bibliographic API, not the paper's own prose; adequate for the low-stakes, uncontested claim that the reference is *in the bibliography*, not for claims about what the prose *says*).

This confirms Linnainmaa (1976) is cited by Baydin et al., but a bibliography entry is not itself "discussion of citation history" — it is a standard citation, consistent with the paper's practice of citing Wengert (1964), Rumelhart et al. (1986), Griewank (2012), and Schmidhuber (2015) elsewhere without narrating each one's reception history.

## The "un-retrieved section" question

The task framing presupposes an "un-retrieved section" of Baydin et al. exists that discusses when Linnainmaa's work was recognized. Based on the direct-text evidence gathered here, **no such section was located**, and the paper's own historical framing sentence explicitly outsources the "colorful history" narrative to Griewank (2012) and Schmidhuber (2015) rather than telling it in Baydin et al.'s own words. It is possible the "un-retrieved section" being referenced by the earlier harvesting task is not in Baydin et al. at all, but in one of those two works it cites — Griewank, A. (2012), "Who Invented the Reverse Mode of Differentiation?", Documenta Mathematica, Extra Volume ISMP, pp. 389–400 (confirmed via WebSearch citation matching, source_tier: 3 for the citation metadata itself — title/venue/pages triangulated across CiteSeerX, ResearchGate, and Semantic Scholar listings, not yet directly read in full). This is flagged as a lead for follow-up, not a verified finding.

## Related but separately-sourced claim (NOT from Baydin et al. — do not conflate)

A different, independent source — Yuxi Liu's blog post "The Backstory of Backpropagation" — makes the specific claim that Linnainmaa's work was essentially uncited for decades:

> "A bibliography search shows that his thesis were basically never cited before the 2010s."

> "I checked all his English papers during the 1970s, and it seems only (Linnainmaa 1976) has been cited non-negligibly before 2000, and may have had some impact on automatic differentiation, but I cannot tell by how much."

source_url: https://yuxi.ml/essays/posts/backstory-of-backpropagation/ — source_author: Yuxi Liu (independent researcher blog) — source_tier: 3-4 (independent blog; author describes doing his own bibliography/citation search, but this is a secondary retelling, not a primary citation-count dataset).

This is a **quantitative/historical claim about citation counts**, which per the sourcing floor requires Tier 1-2 for the quantitative portion. It is recorded here only as context for why the harvested question exists, and is marked **[unverified-quant — needs primary]**: the underlying citation-count claim ("basically never cited before the 2010s") needs a primary bibliometric source (e.g., a direct Google Scholar/Web of Science citation-history pull for Linnainmaa 1970/1976) before it can be treated as established fact. This blog post explicitly is *not* Baydin et al., and confirming or denying whether Baydin et al. discusses this claim is exactly the un-resolved part of the core question above.

> [!note] Seek's commentary:
> The most interesting finding here isn't a yes/no on Linnainmaa — it's that Baydin et al. seem to treat the "who really invented backpropagation" question as already litigated elsewhere (Griewank 2012, Schmidhuber 2015) and decline to re-relitigate it themselves. That's a reasonable editorial choice for a technical survey, but it means anyone chasing "when was Linnainmaa recognized" should go to Griewank (2012) directly rather than expecting Baydin et al. to carry that historiographic weight. Also worth flagging for whoever picks this up next: the tooling available in this pass could not get a clean, complete read of the back half of a 43-page PDF/HTML document despite five distinct retrieval strategies. That's a capability gap worth remembering next time a "pull the exact section" task comes in — plain-text mirrors (ar5iv) work better than raw PDFs, but even those get silently truncated on long documents.

## Status of core question

**[unverified -- could not confirm or deny after search]** for the specific sub-claim about a not-yet-located later section of Baydin et al. discussing Linnainmaa's recognition timeline. What *could* be confirmed directly from the primary text: Baydin et al. cite Linnainmaa (1976) in the bibliography, do not appear to discuss his citation history in prose in the roughly two-thirds of the paper that was directly and repeatedly verified, and attribute the "reinvented at various times" historical narrative to Griewank (2012) and Schmidhuber (2015) by citation rather than narrating it themselves.