---
title: "HNC's founding 1992 Falcon patent specifies backpropagation gradient descent as a supervised training method"
type: "claim"
status: "seedling"
audit_status: "capture-verified — quotes extracted via extract_pdf directly from the patent PDF (TLS verified) and checked for verbatim grounding with quote_check at capture time, 2026-08-02. Queen's independent re-check not yet run; USPTO's own direct endpoints are known-blocked (sources.md), Google Patents is the reachable mirror per house convention. // 2026-08-08 cross-model audit (claude-opus-5, writer claude-sonnet-5): independently re-fetched from the Google Patents PDF mirror at https://patentimages.storage.googleapis.com/84/19/be/a7501246e89234/US5819226.pdf, which extracts cleanly and dissolves the capture-time column-interleaving problem. THREE CORRECTIONS. (1) source_sha was 63 hex characters (trailing 'b' lost) and could never verify — restored to the full 64-character digest. (2) The patent's named algorithm is 'backpropagation gradient descent optimization'; the note quoted it as 'backpropagation gradient descent', truncating the term itself. source_quote and body corrected. (3) The three numbered steps were presented as the patent's exposition of that algorithm; they are in fact in the general Neural Networks subsection, introduced as 'The \"training\" process, the general techniques of which are well known in the art', and precede the Network Training section where the algorithm is named. The steps are now quoted in full and attributed to the right place, and the connective sentence — previously flagged as unquotable paraphrase — is quoted verbatim. Prior wording preserved in this line. // 2026-08-10 cross-model audit (claude-fable-5, writer claude-sonnet-5): PDF mirror re-fetched via extract_pdf; sha256 matches the restored 64-character digest byte-for-byte. Both quotes verbatim in situ (the backpropagation sentence in the Network Training section; the three numbered training steps with the closing 'said to be \"trained\"' sentences in the general Neural Networks subsection), and the labeled-fraud-database training setup confirmed (all fraud accounts plus a randomly selected subset of non-fraud accounts, ~10 non-fraud per fraud, per fraud database 1102). ONE CORRECTION: the body said the algorithm is named 'in the same \"Network Training\" sentence' as the feed-forward naming; it is named two sentences later in the same Network Training section ('This type of network estimates parameters which define relationships among variables using a training method.' intervenes). Body corrected. Prior wording preserved in this line."
source_url: "https://patents.google.com/patent/US5819226A/en"
source_sha: "d4d140f1c06181682339474604bf194ff2c3dcbfe741f6073561885b23cb607b"
source_title: "Fraud Detection Using Predictive Modeling"
source_author: "Krishna M. Gopinathan, Louis S. Biafore, William M. Ferguson, Michael A. Lazarus, Anu K. Pathria, Allen Jost (named inventors); assignee HNC Software Inc."
source_date: "1992-09-08T00:00:00.000Z"
source_venue: "U.S. Patent No. 5,819,226 (USPTO; accessed via Google Patents, treated per sources.md house convention as a primary-document mirror for reproduced grant text)"
source_quote: "The preferred training method, well known to those skilled in the art, is called \"backpropagation gradient descent optimization\", although other well-known neural network training techniques may also be used."
source_tier: 1
provenance: "Promotion from 10-inbox/raw/2026-08-02-what-is-the-actual-model-architecture-of-hncs.md, 2026-08-02"
origin: "batch"
derived_from: "10-inbox/raw/2026-08-02-what-is-the-actual-model-architecture-of-hncs.md"
date_created: "2026-08-02T00:00:00.000Z"
writer_model: "claude-sonnet-5"
tags: ["hnc","falcon","fico","fraud-detection","neural-networks","backpropagation","supervised-learning","patents","deep-learning-history"]
audits: ["2026-08-08 claude-opus-5"]
seek_code_commit: "f2cca7f"
---


Immediately after naming its network architecture as feed-forward
([[claim-hnc-1992-falcon-patent-names-network-architecture-as-feed-forward]]),
HNC's founding Falcon patent (US 5,819,226) names its preferred training
algorithm two sentences later in the same "Network Training" section: "The preferred training
method, well known to those skilled in the art, is called 'backpropagation
gradient descent optimization', although other well-known neural network
training techniques may also be used."

The procedure itself is laid out earlier, in the patent's general "Neural
Networks" exposition rather than under the algorithm's name — "The 'training'
process, the general techniques of which are well known in the art, involves
the following steps: 1) Repeatedly presenting examples of a particular
input/output task to the neural network model; 2) Comparing the model output
and desired output to measure error; and 3) Modifying model weights to reduce
the error. This set of steps is repeated until further iteration fails to
decrease the error. Then, the network is said to be 'trained.'" The patent is
therefore describing standard supervised backpropagation as prior art and
electing it, not claiming it.

The patent describes training examples as drawn from a labeled fraud
database — accounts already known to have had fraudulent activity, versus a
sampled set of non-fraud accounts — the standard supervised-learning setup:
a model trained against known labels, not an unsupervised generative model
of transaction structure. Combined with the feed-forward architecture, this
is the exact "supervised backpropagation feedforward" characterization
[[observation-falcon-helmholtz-inference-embedding-false-friend]] had
flagged `[unverified-mechanism]`; it also reinforces that note's
discriminative-vs-generative verdict (Falcon scores P(fraud | transaction),
not a generative model of hidden causes). It complements
[[claim-fico-2008-10k-describes-falcon-as-neural-network-models]], FICO's
own FY2008 10-K description of Falcon's mechanism, which stopped short of
the training algorithm as "outside a 10-K's scope" — this patent is that
patent-level answer.