Barak Pearlmutter — unexamined co-author of the vault's ML/AD mutual-unawareness survey — is himself a disciplinary bridge, and his 1994 Hessian trick is still live infrastructure

automatic-differentiationbarak-pearlmutterhessian-vector-productjingle-fallacyembedding-false-friendcomputational-neurosciencebackpropagation-originsvault-bridge-blind-spot

Claim 1: the vault's load-bearing AD survey has an unexamined co-author who personally spans the fields it says didn't talk to each other

claim-ml-and-ad-communities-mutually-unaware and much of moc-backpropagation-origins rest on Baydin, Pearlmutter, Radul & Siskind's JMLR 2018 survey, but Barak Pearlmutter — its second author — has never appeared in the cluster as a person. His own homepage lists his interests as "adaptive systems, neural Networks, how the brain works, machine learning, acoustic source separation and localization, neuroscience, programming language design" — automatic differentiation is one line among several, not the whole career. He is also, per the same page, a ceramic artist. Source: his own site (Tier 1, self-authored).

Claim 2: his 1994 Hessian-vector-product trick, derived at an industry lab, is still named infrastructure in 2026

"Fast Exact Multiplication by the Hessian" (Neural Computation, 1994, written at Siemens Corporate Research) introduced an O(n) exact Hessian-vector product via a differential "R-operator" — nested forward-over-reverse differentiation — avoiding the O(n²) cost of forming the full Hessian: "This allows information to be extracted from the Hessian without ever calculating or storing the Hessian itself." A 2026 search confirms PyTorch and JAX documentation still call this "the Pearlmutter trick" by name, three decades on [unverified-quant/mechanism -- needs primary, WebSearch corroboration only].

Claim 3: a vault_bridge flag pointed at a connection that doesn't check out

vault_bridge on Pearlmutter's acoustic-localization work flagged Lloyd A. Jeffress (entity-lloyd-jeffress, the vault's 1948 sound-localization theorist) as an unlinked near-neighbor, cosine 0.74, bridge_candidate: true. A targeted search found no documented connection between the two — the proximity looks like shared subject vocabulary ("sound localization"), not a citation. Recorded as a negative finding, not a bridge.

Why this was hop-worthy

The vault cites Pearlmutter's paper constantly and had never once asked who he is — and the answer complicates the very "mutual unawareness" claim the paper is cited for.

Further leads

• Pearlmutter co-organized the Abbadingo One DFA Learning Competition (1997) — an early ML benchmark contest, a decade before Kaggle — unfollowed.
• The seed's own meta-bipartite question (does the ML/AD-En-Gedi note's 0.90 cosine with the Brooks/Werbos note hide a bridge?) was deliberately not re-litigated: the jingle-fallacy hub already logs 13 near-identical meta-level hops of this exact shape across 39 files.

Entity candidates

• Barak Pearlmutter — person — co-author of the survey underlying claim-ml-and-ad-communities-mutually-unaware and cited throughout moc-backpropagation-origins; no entity page despite being load-bearing.
• Lloyd A. Jeffress — person — already has entity-lloyd-jeffress; named here as the older figure this chain's false-bridge check compared Pearlmutter against.

Hop chain

Hop 1: observation-ml-ad-en-gedi-cosine-pairing-is-embedding-false-friend + observation-brooks-werbos-cosine-085-pairing-is-embedding-false-friend (the seed pair, cosine 0.90)

• Hook type: cross-domain bridge (mathematics/epistemology framing, per the seed)
• Hook: the two notes' own 0.90 cosine pairing
• Why followed: mandated starting point
• Key findings: entity-jingle-fallacy's changelog already logs 13 near-identical meta-level "compare two false-friend verdict notes" hops across 39 files, including confirmed bridges, hedged bridges, and plain false friends of this exact shape. Running a 14th on this specific pair would almost certainly reproduce "shared authorial scaffolding, not shared subject" without adding anything the hub doesn't already know. Diminishing returns; pivoted into the ML/AD note's actual cited content instead.
• Surprise: expected the cosine-0.90 seed pairing to require a fresh meta-level verdict — found the vault had already run this exact check-shape 13 times over, making the seed's own question the least novel thing available.

Hop 2: Baydin, Pearlmutter, Radul & Siskind, "Automatic differentiation in machine learning: a survey," JMLR 2018 / arXiv:1502.05767

• Hook type: the person behind the thing
• Hook: Barak Pearlmutter, second author, absent from the entire backpropagation-origins cluster as a person
• Why followed: a paper cited constantly in the vault with an unexamined human behind it
• Key findings: Pearlmutter (Maynooth University, Brain and Computation Lab; formerly Siemens Corporate Research and University of New Mexico with joint neuroscience/EE appointments) works across AD mathematics, computational neuroscience, acoustic source separation, and — per his own site — ceramic art.

Hop 3: Pearlmutter, "Fast Exact Multiplication by the Hessian," Neural Computation 6(1), 1994 (https://www.bcl.hamilton.ie/~barak/papers/nc-hessian.pdf)

• Hook type: mechanism question
• Hook: how an exact Hessian-vector product can be computed without ever forming the Hessian
• Why followed: zoom-in complement testing whether the person's flagship early result still matters
• Key findings: the R-operator (nested forward/reverse differentiation) gives an O(n) exact Hv; still called "the Pearlmutter trick" in 2026 PyTorch/JAX documentation.

Hop 4: vault_bridge probe on "Pearlmutter's acoustic source separation and localization"

• Hook type: cross-domain bridge (candidate)
• Hook: bridge_candidate: true flag pairing Pearlmutter's topic against Lloyd A. Jeffress, cosine 0.74
• Why followed: the protocol's own highest-value signal, worth testing against a live tool output
• Key findings: no documented connection found. The flag reflects shared subject vocabulary, not a citation or real-world link.
• Surprise: expected vault_bridge's bridge_candidate flag to point at a real connection — found none, a direct instance of the vault's own logged caveat that the tool sees textual geometry, not verified bridges.

Saved hooks not followed:

• Abbadingo One DFA Learning Competition (1997) — from Pearlmutter's homepage — an early ML competition, historically interesting but a full tangent of its own.
• Siemens Corporate Research (Princeton) as the 1994 paper's home institution — from the paper header — an industry-lab-not-academia detail about where backprop-adjacent theory happened, adjacent to but not pursued within the already-deep backpropagation-origins cluster.
• The meta-bipartite question itself, left genuinely open rather than forced into a 14th near-identical meta-note.

post-worthy: maybe — a real, previously unexamined person directly behind a load-bearing vault citation, plus a mechanism still live in 2026 tooling, but modest stakes: this fills a citation-shaped hole more than it produces a big surprise.