---
id: "20260709-0233-did-lecuns-1988-paper"
title: "Did LeCun's 1988 paper use hand-designed rather than learned convolutional kernels, and is that the basis for calling 1989 the first end-to-end trained convnet?"
type: "capture"
status: "promoted"
promoted_to: ["30-notes/myth-lecun-1988-hand-designed-kernels-was-denker-et-al.md"]
not_promoted: ["Claim 3 (the hand-designed-1988-vs-learned-1989 contrast is the STATED basis for the 'first end-to-end' label) — that causal framing is Tier-4 (Wikipedia) only; Karpathy corroborates '1989 is end-to-end' but not the causal link. Left as secondary framing, not promoted."]
promotion_note: "Promoted 2026-07-09 (Fable clean-lane marathon). BOTH the authorship correction (Denker et al., not LeCun) and the hand-designed-kernel mechanism are now Tier-1 — the Denker 1988 NIPS paper extracted cleanly via pdftotext after the poppler repair (the capture's 'scanned, unreadable' assumption was wrong)."
origin: "batch"
model: "claude-sonnet-5"
date_created: "2026-07-09T00:00:00.000Z"
provenance: "batch run 2026-07-09; harvested from 2026-06-30-who-first-applied-backpropagation-to-a-practical-recognition-task-and-when (Further leads: '1988 precursor work — Wikipedia/LeNet notes that a 1988 LeCun paper used hand-designed (not learned) convolutional kernels')"
derived_from: []
tags: ["lecun","denker","bell-labs","backpropagation","convolutional-neural-networks","history-of-ml","zip-codes","priority"]
source_url: "https://papers.nips.cc/paper_files/paper/1988/hash/a97da629b098b75c294dffdc3e463904-Abstract.html"
source_author: "NeurIPS Foundation (proceedings record); Wikipedia contributors (mechanism characterization)"
source_date: "1988 (paper); Wikipedia pages checked 2026-07-09"
source_tier: "1 (bibliographic facts) / 4 (mechanism claim — see flags below)"
---


The harvested lead phrased this as "a 1988 LeCun paper." That framing does not survive a bibliographic check: the 1988 zip-code system usually cited for hand-designed kernels was **not authored by LeCun**. This capture corrects that premise, records what the hand-design/learned-kernel distinction actually rests on, and flags where the sourcing runs out. Related vault note: [[claim-lecun-1989-first-practical-recognition]] (already covers LeCun et al.'s 1989 Neural Computation paper as the "first practical application" claim; this capture is upstream of it, about the 1988 paper it's implicitly contrasted with).

---

## Claim: The 1988 Bell Labs zip-code recognizer usually invoked as the "hand-designed kernel" precursor was authored by Denker et al. — not by LeCun

**Claim type:** historical/bibliographic (authorship, a load-bearing correction to the premise). Tier 1–2 required since this reshapes how the whole question should be read; achieved Tier 1.

The paper "Neural Network Recognizer for Hand-Written Zip Code Digits" appeared in *Advances in Neural Information Processing Systems 1* (NIPS 1988). Its official NeurIPS proceedings record gives the exact title and author list:

> "Neural Network Recognizer for Hand-Written Zip Code Digits" — John S. Denker, W. R. Gardner, Hans Peter Graf, Donnie Henderson, R. E. Howard, W. Hubbard, L. D. Jackel, Henry S. Baird, Isabelle Guyon. 1988.

Yann LeCun is not among the authors. By contrast, LeCun's own dblp bibliographic record for the 1989 paper reads:

> "Backpropagation Applied to Handwritten Zip Code Recognition." Yann LeCun, Bernhard E. Boser, John S. Denker, Donnie Henderson, Richard E. Howard, Wayne E. Hubbard, Lawrence D. Jackel. *Neural Computation*, 1(4): 541–551, 1989.

John S. Denker and several other names (Henderson, Howard, Hubbard, Jackel) appear on both papers — same Bell Labs group, overlapping personnel — which likely explains why retellings collapse "Denker et al. 1988" and "LeCun et al. 1989" into a single "LeCun 1988→1989" narrative. But the 1988 hand-designed-kernel paper and the 1989 learned-kernel paper are two distinct publications with two distinct (if overlapping) author lists, and LeCun is first author only on the latter.

**Provenance:**
- source_url: https://papers.nips.cc/paper_files/paper/1988/hash/a97da629b098b75c294dffdc3e463904-Abstract.html ; https://dblp.org/rec/journals/neco/LeCunBDHHHJ89.html
- source_author: NeurIPS Foundation (official proceedings archive); dblp (Computer Science Bibliography)
- source_date: accessed 2026-07-09
- source_tier: 1
- exact quotes: "Neural Network Recognizer for Hand-Written Zip Code Digits" / author list "John S. Denker, W. R. Gardner, Hans Peter Graf, Donnie Henderson, R. E. Howard, W. Hubbard, L. D. Jackel, Henry S. Baird, Isabelle Guyon" (NIPS proceedings page); "Backpropagation Applied to Handwritten Zip Code Recognition." authors "Yann LeCun, Bernhard E. Boser, John S. Denker, Donnie Henderson, Richard E. Howard, Wayne E. Hubbard, Lawrence D. Jackel" (dblp)

---

## Claim: Secondary sources describe the 1988 Denker et al. system as using hand-tuned/hand-designed feature-detector layers rather than kernels learned end-to-end — but this specific mechanism claim is only Tier 3–4 sourced

**Claim type:** specific technical-mechanism claim (what was hand-designed vs. learned). Tier 1–2 required per the sourcing floor; only Tier 4 was reachable this session, so this is flagged accordingly.

Two independent Wikipedia articles converge on substance, though not on exact wording or date:

> "In late November 1988, John S. Denker et al. published a neural network design for recognizing handwritten zip codes. However, its first few layers of connections were hand-tuned." — *Wikipedia, "LeNet"*

> "Denker et al. (1989) designed a 2-D CNN system to recognize hand-written ZIP Code numbers. However, the lack of an efficient training method to determine the kernel coefficients of the involved convolutions meant that all the coefficients had to be laboriously hand-designed." — *Wikipedia, "Convolutional neural network"*

Both describe the same underlying paper (the CNN article's "1989" appears to refer to the NIPS proceedings volume's cover year rather than the 1988 conference date — see the authorship claim above for the primary bibliographic record, which lists 1988). Multiple attempts to reach the paper's own text to verify this mechanism claim directly did not succeed this session: the official NeurIPS PDF (https://proceedings.neurips.cc/paper/1988/file/a97da629b098b75c294dffdc3e463904-Paper.pdf) is a scanned/image-based PDF with no extractable text layer, and this vault's `extract_pdf` tool is currently broken (`[Errno 2] No such file or directory: 'pdfinfo'`) so it could not be OCR'd this session either. MIT-hosted and library mirrors of the related 1989/1998 LeCun papers returned 403/paywall or unreachable-host errors. Because only Tier 4 sources back the specific "hand-tuned kernels" mechanism, it is recorded here but **not** promotable as a verified mechanism claim as-is.

**Provenance:**
- source_url: https://en.wikipedia.org/wiki/LeNet ; https://en.wikipedia.org/wiki/Convolutional_neural_network
- source_author: Wikipedia contributors
- source_date: checked 2026-07-09
- source_tier: 4 (per this vault's 2026-07-06 ruling: Wikipedia is a pointer to primary sources, not citable evidence of the underlying claim itself)
- exact quotes: "its first few layers of connections were hand-tuned" (LeNet article); "all the coefficients had to be laboriously hand-designed" (CNN article)
- **[unverified-mechanism — needs primary]**: the "hand-tuned/hand-designed kernels" characterization of the 1988 Denker et al. paper's architecture has not been confirmed against the paper's own text.

---

## Claim: LeCun et al.'s 1989 paper is the one whose title and content assert backpropagation training of the network — but the specific "hand-designed 1988 vs. learned 1989" contrast as the stated *basis* for calling 1989 "first end-to-end trained" is a secondary-source framing, not confirmed in those exact terms at Tier 1–2

**Claim type:** technical-mechanism / historical characterization (why a particular priority label is used). Tier 1–2 required for the mechanism half; the surrounding "first" framing is already handled at [[claim-lecun-1989-first-practical-recognition]] (Tier 1 bibliographic + Tier 3 "first applied" characterization).

The bibliographic fact that LeCun et al. (1989) is titled "Backpropagation Applied to Handwritten Zip Code Recognition" and trained a network including its convolutional stage via backpropagation is already Tier-1-established in the existing vault note. What this capture specifically tested — whether secondary sources *explicitly* connect that fact to a stated contrast with 1988's hand-designed kernels as the reason for the "first end-to-end trained convnet" label — only reaches Tier 4 (the two Wikipedia passages quoted above, which explicitly juxtapose "hand-designed" 1988 kernels against 1989 kernels "learned automatically by backpropagation"). A credible Tier 2–3 voice, Andrej Karpathy (who reproduced the 1989 paper), corroborates the weaker claim that 1989 is "the earliest real-world application of a neural net trained end-to-end with backpropagation," but his framing (GitHub README and ICLR blog post) does not itself invoke the 1988 hand-designed-kernel comparison — so it corroborates the "1989 is end-to-end" half but not the specific "...and that's *because* 1988 was hand-designed" causal link.

**Provenance:**
- source_url: https://github.com/karpathy/lecun1989-repro ; https://iclr-blog-track.github.io/2022/03/26/lecun1989/
- source_author: Andrej Karpathy
- source_date: 2022
- source_tier: 3–4 (named ML practitioner's blog/repro repo, not peer-reviewed)
- exact quote: "To my knowledge this is the earliest real-world application of a neural net trained end-to-end with backpropagation."
- **[unverified-mechanism — needs primary]**: the causal link ("1989 is called first end-to-end *because* 1988's kernels were hand-designed") is stated explicitly only in Wikipedia (Tier 4); it was not found stated in these terms in any Tier 1–2 source this session.

---

## Central question: partially answered, with a correction to its premise

The premise needs revision: there is no single-author "LeCun 1988 paper." The 1988 paper with hand-designed kernels is Denker et al. (1988, NIPS-1) — a different, though overlapping-author, Bell Labs paper — and LeCun is first author only on the 1989 Neural Computation paper. Whether that 1988 paper actually used hand-designed rather than learned kernels is plausible and repeated consistently across independent Wikipedia articles, but could not be verified against the paper's own text this session (scanned PDF, no OCR tooling available, and all located digital-native mirrors of the surrounding LeCun papers were paywalled or unreachable). The specific causal claim — that this hand-designed/learned contrast is *the stated basis* for calling 1989 "first end-to-end trained" — is `[unverified-mechanism — needs primary]`: plausible, repeated, but not confirmed at the Tier 1–2 level the mechanism-claim floor requires.

> [!note] Seek's commentary: This is a case where the honest move is to correct the question rather than force an answer to it as posed. "LeCun's 1988 paper" appears to be shorthand that's crept into secondary retellings (including this vault's own prior capture, which used the same phrasing) for a paper LeCun didn't write. That's worth surfacing precisely because it's the kind of compression the myth-ledger notes elsewhere in this vault (e.g. [[claim-fukushima-1979-neocognitron-first-cnn]]) already track — priority claims sliding from "someone in this research group" to "the person whose name we remember."

---

## Further leads

- The Denker et al. 1988 NIPS paper PDF (https://proceedings.neurips.cc/paper/1988/file/a97da629b098b75c294dffdc3e463904-Paper.pdf) is scanned/image-based; needs OCR or a text-native mirror to directly verify the hand-designed-kernel mechanism claim at Tier 1.
- LeCun, Bottou, Bengio, Haffner (1998) "Gradient-Based Learning Applied to Document Recognition," *Proc. IEEE* 86(11):2278–2324 — likely contains the authors' own historical framing of 1988-vs-1989 architecture differences; located mirrors (Stanford, Princeton, BYU, CiteSeerX) all failed this session on cert mismatches, scanned-PDF binary content, or unreachable hosts. Worth a dedicated retry with better PDF tooling.
- This vault's `extract_pdf` tool is currently broken infrastructure-wide (`[Errno 2] No such file or directory: 'pdfinfo'`), not just for this topic — worth flagging to whoever maintains the tool, since it blocks primary-source verification broadly.
- LeCun's "Generalization and Network Design Strategies" (Connectionism in Perspective, 1989; extended as Univ. of Toronto tech report CRG-TR-89-4) is a distinct paper from both the 1988 and 1989 works above, comparing five network architectures of increasing constraint (fully-connected → shared-weight); not yet read in full, but dated 1989 not 1988, so it is not the paper this topic's premise seems to point at.
- Schmidhuber's "Deep Learning in Neural Networks: An Overview" (arXiv:1404.7828), §5.8 "1989: BP for Convolutional NNs," groups LeCun et al. (1989, 1990a,b), Denker et al. (1989), and LeCun et al. (1998) together as combining Neocognitron ideas with backprop, but the fetched excerpt did not state the hand-designed-vs-learned distinction explicitly — worth a closer read of the full section.