US 10856815

US 10,856,815 B2 — Verification Summary

Search caveat up front: I found no Court of Appeals for the Federal Circuit (or district court) 2026 docket, appeal, or other litigation involving U.S. Patent No. 10,856,815. Searches for the number returned only the patent itself plus unrelated litigation. Separately, be aware of a literal-identifier trap: the string "10856815" is also the application serial number (10/856,815) of a different, unrelated patent — US 7,374,452 B2 ("Battery Accommodating Structure and Mobile Terminal," NTT DoCoMo). That is not the patent you asked about. My summary below covers only US 10,856,815 B2, whose application number is 15/752,276.

Bibliographic data

Abstract (as granted)

Given functional imaging data acquired while a subject reads a text passage, a reconstruction of the text passage is produced. Linguistic semantic vector representations are assigned (1301) to words, phrases or sentences to be used as training stimuli. Basis learning is performed (1305), using brain imaging data acquired (1303) when a subject is exposed to the training stimuli and the corresponding semantic vectors, to learn an image basis directly. Semantic vector decoding (1309) is performed with functional brain imaging data for test stimuli and using the image basis to generate a semantic vector representing the test imaging stimuli. Text generation (1311) is then performed using the decoded semantic vector.

Plain-language overview of the three independent claims

Claim 1 (method). Capture a first set of functional brain images while a subject is exposed to training text (the text being assigned semantic vectors). Decompose that first imaging data into a weighted combination of basis images, each basis image being a spatial pattern of brain activation. For each training stimulus, map that weighted combination to values in each dimension of the stimulus's semantic vector. Then capture a second set of brain images while the subject is exposed to a test stimulus; identify a linear combination of the basis images within that second data; and decode a semantic vector for the test stimulus by estimating the weight of each basis image in that linear combination. Finally, generate a text output from the decoded semantic vector.

Claim 5 (system). The apparatus counterpart: a linguistic semantic vector representor that assigns semantic vectors to training text; a basis learner that decomposes the first (training) imaging data into a weighted combination of basis images (each a spatial activation pattern) and maps that combination to semantic-vector dimension values; a semantic vector decoder that identifies a linear combination of basis images in a second (test) imaging dataset and decodes a semantic vector by estimating each basis image's weight; and a text generator that produces text output from the decoded vector.

Claim 12 (method). A narrower/re-framed method variant. It begins from assigning multidimensional linguistic semantic vector representations to training text, then decomposes first-set imaging data (acquired in the presence of that text) into a weighted combination of basis images, maps that combination to semantic-vector dimension values, identifies a linear combination of basis images in second-set imaging data, decodes a semantic vector by estimating each basis image's weight, and generates text based on the semantic vector. Notably, claim 12 recites the imaging data as already acquired (no "capturing" step) and recites the output only as "text."

Dependent claims. 2 (semantic vectors based on context); 3 (output is a natural-language representation of the second dataset); 4 (vectors are multidimensional); 6 (training text = words/phrases/sentences); 7 (sentence vector by averaging word vectors); 8 (basis learner identifies common characteristics for overlapping semantic vectors); 9 (ridge regression); 10 (recurrent neural network text generator); 11 (output is a sequence of words forming a predicted sentence); 13 (training text = words/phrases/sentences); 14 (generated text = word/phrase/sentence).

Working notes / uncertainties

• Claim-specification framing difference: the specification describes a forward model (learn basis matrix B mapping known semantic vectors Z to voxel patterns X, i.e., X ≈ ZB) and a reverse decoding step. The granted independent claims recast this as "decomposing" the brain image into weighted basis images and "mapping" those weights to semantic-vector dimensions. I have not verified the prosecution history that produced this wording — treat that observation as descriptive, not a legal conclusion.
• Named prior art cited on the face: US 2013/0184558 A1; US 2013/0246322 A1; US 2015/0269139 A1; US 2008/0270120 A1; US 2011/0301431 A1; US 2013/0178733 A1; CN 104391963 A; US 2018/0092567 A1; and non-patent literature including Pereira et al., "Generating Text from Functional Brain Images" (Front. Hum. Neurosci., 2011), Fyshe et al. (ACL 2014), Anderson et al. (NeuroImage 2015), Simanova et al. (Cereb. Cortex 2014), Shinkareva et al. (NeuroImage 2011), Friston SPM (2007) and Smith FSL (2004). The Pereira 2011 paper by a co-inventor appears as the closest "X"-category reference in the EP search report.
• Not found: any Federal Circuit or other court docket in 2026 (or otherwise) specifically asserting or reviewing US 10,856,815. I cannot confirm litigation activity for this patent; the Siemens-related suits surfaced in searching were unrelated patents/technologies.
• Data above is taken from the authoritative full patent text supplied (Google Patents, https://patents.google.com/patent/US10856815/en) plus cross-checked EPO/UK register records for the EP family member.

I'll search for litigation involving US patent 10856815.

Let me search more targeted sources.

Litigation Search Results for U.S. Patent No. 10,856,815

Bottom line: I found no evidence of any litigation (district court, PTAB, ITC, or Federal Circuit appeal) involving U.S. Patent No. 10,856,815. Based on the searches I was able to run across patent-litigation portals and general web sources, there is no known case to report.

What the patent is

U.S. Patent No. 10,856,815 ("Generating natural language representations of mental content from functional brain images"):

• Assignee / patent owner: Siemens Medical Solutions USA, Inc. (original assignee; inventors Francisco Pereira, Bin Lou, Angeliki Lazaridou)
• Priority date: October 23, 2015; filed: October 21, 2016 (PCT/IB2016/001517, US national phase 15/752,276); granted: December 8, 2020
• Family members: EP3364868B1, CN108135520B, WO2017068414A2
• Government interest: Research was government-supported (ODNI/IARPA via AFRL, contract FA8650-14-C-7358), and a confirmatory license was recorded to The Government of the United States as represented by the Secretary of the Air Force (dated February 15, 2021). This is a Bayh-Dole-style government-interest recordation, not litigation.
• Status: Active; adjusted expiration ~June 20, 2037. Records show only routine prosecution (a non-final office action Jan. 2020, a final rejection June 2020, and issue-fee payment Oct. 2020) and a 4th-year maintenance fee (May 2024).

What my searches covered

I searched general web and patent-litigation-oriented sources (including Unified Patents' patent/litigation pages, CourtListener/Justia materials, PACER-adjacent documents, and PTAB-related documents surfaced through the web searches) for "10856815," "10,856,815," "US10856815," and Siemens + brain-imaging terms. Results returned:

• The patent's own prosecution/publication records (Google Patents, Justia, EPO register).
• Unrelated hits sharing the digits "10856815" — e.g., a Brazilian corporate registration number (JPV Participações Ltda.), a Chinese trademark application no. 10856815, a UK company (Supreme Joinery Ltd., company no. 10856815), an application number for US-7374452-B2 (Fujitsu/NTT Docomo battery patent), and a Max Planck article URL fragment.
• Similarly numbered but different patents appearing in litigation in other cases (e.g., U.S. Patent No. 9,858,215 asserted in Netlist v. Samsung, E.D. Tex. 2:22-cv-00293-JRG; U.S. Patent No. 8,606,856 in Uniloc v. Microsoft). Per your instruction, I did not treat these as the same patent.

No docket, complaint, IPR petition, ITC complaint, or appellate decision naming U.S. Patent No. 10,856,815 was located.

Important caveats

• My searches are not an exhaustive, docket-level check of PACER, Docket Navigator, or Unified Patents' full litigation portal; I was unable to directly pull a per-patent litigation record from those databases. It is possible a case exists that did not surface in these searches.
• To confirm a clean negative with higher confidence, the definitive next steps would be: (1) the Unified Patents litigation portal entry for US-10856815 (https://portal.unifiedpatents.com/patents/patent/US-10856815-B2); (2) PACER / CourtListener full-text search for "10,856,815" across all districts; (3) the USPTO PTAB API/Patent Center "litigation" tab; and (4) the USPTO assignment/legal-status record (which currently shows only the assignment and the Air Force confirmatory license).

If you'd like, I can run additional targeted searches (e.g., specific districts or the Siemens entity as plaintiff) to try to close this out further.

I'll start by checking whether any AIA trial proceedings exist for this patent beyond what the structured ODP block reports.

No hits yet. Let me run a few more targeted checks (PTAB E2E / Unified Patents / litigation that might spawn an IPR).

Let me do a final check for any district court litigation or PTAB docket entry tied to this patent number.

Proceedings overview

Total AIA trial proceedings on file for US 10,856,815: 0. The canonical structured block (USPTO Open Data Portal) reports no IPR, PGR, or CBM proceeding, and my independent web checks surfaced nothing to contradict it — so the breakdown is 0 active / 0 claims invalidated / 0 claims sustained / 0 settled / 0 institution denials, and I found no Federal Circuit appeal touching this patent. Bottom-line defensive posture: the patent has never been tested at the PTAB, so there is no claim-cancellation shortcut available to a defendant, no estoppel record to lean on, and no adverse FWD to cite. All 14 claims stand as issued (independent claims 1, 5, and 12, with dependents 2–4, 6–11, and 13–14), and any IPR/PGR would have to be built from scratch by whoever files it.

A caveat on how I reached that conclusion: my searches returned several near-miss results that are not this patent and should not be mistaken for activity on it — e.g., the "'568 Patent" in Samsung v. Genghiscomm, IPR2025-00788, is US 10,389,568 (OFDM/DFT spreading codes), and the Centripetal "'856" / "'806" IPRs and the VirtaMove and Neurent/Aerin matters are unrelated. No source connected any proceeding to U.S. Patent No. 10,856,815 or to application 15/752,276.

Proceedings

None to report. There are no proceeding numbers to list, no judge panels, no institution decisions, no Final Written Decisions, no settlements, and no appeals. Per the operating rules for this task, I am not inventing proceeding numbers to populate this section.

Strategic summary

Claim status. US 10,856,815 issued 2020-12-08 with 14 claims and has not been narrowed by any AIA trial. Independent claim 1 (method: capture training-set fMRI while a subject is exposed to training text → decompose into a weighted combination of basis images → map the weighted combination to semantic-vector dimensions → capture test-set fMRI → identify a linear combination of basis images → decode a semantic vector by estimating each basis image's weight → generate text output), independent claim 5 (the mirror system claim, reciting a linguistic semantic vector representor, basis learner, semantic vector decoder, and text generator), and independent claim 12 (method reciting "multidimensional linguistic semantic vector representations") are all UNTESTED at the PTAB, and therefore SUSTAINED as issued. That is not the same as upheld — it means no adjudicator has yet weighed in. Dependents 2–4 (§context-based vectors; natural-language output; multidimensional vectors), 6–11 (words/phrases/sentences; averaging word vectors; common characteristics for overlapping vectors; ridge regression; recurrent neural network; word-sequence output), and 13–14 (claim 12 variants) are likewise UNTESTED. Note the significance of the ridge-regression limitation in claim 9 and the RNN limitation in claim 10 — they are the two most obvious narrowing fallback positions if an IPR is ever filed against claims 1/5/12, and they sit in dependent claims that a petitioner would have to take on separately.

Estoppel landscape. Because no IPR or PGR has ever been instituted against this patent, § 315(e)(2) estoppel is a blank slate. No petitioner, real party in interest, or privy is barred from raising any § 102 or § 103 ground in a district court or ITC action. If you are being asserted against today, the full universe of prior art — including the eight references cited on the face of the patent and the fifteen non-patent references in the file wrapper — remains available, subject only to the ordinary limits of invalidity pleading. That said, several of the most on-point prior art references were already before the examiner or are cited on the face of the patent, which raises § 325(d) risk for any future petitioner: the Board routinely exercises discretion to deny petitions that recycle art the Office already considered. The file-wrapper NPL includes Fyshe et al. ("Interpretable Semantic Vectors from a Joint Model of Brain- and Text-Based Meaning," ACL 2014), Anderson et al. ("Reading visually embodied meaning from the brain," NeuroImage 2015), and — most notably — Pereira et al., "Generating Text from Functional Brain Images," Frontiers in Human Neuroscience (2011), which shares an inventor (Francisco Pereira) with the patent itself. That last reference is a publicly available printed publication that predates the 2015-10-23 priority date by four years and is squarely on the claim-1 decoding-and-text-generation workflow; it is the natural § 102/§ 103 starting point, but a petitioner should expect a § 325(d) / Advanced Bionics argument against building a ground on art already cited in the IDS.

Pattern signals. No petitioner has filed anything — first, second, or otherwise. There is no defensive aggregator in the chain (no Unified Patents, no RPX, no Open Invention Network filing). The patent owner, Siemens Medical Solutions USA, Inc., has not had occasion to defend at the PTAB and therefore has no appeal history on this patent. Two additional facts shape the commercial reality: (1) the patent carries government rights — it was made with IARPA/ODNI support via AFRL under contract FA8650-14-C-7358, and a confirmatory license to the U.S. Government (Secretary of the Air Force) was recorded 2021-07-12 with an effective date of 2021-02-15; and (2) the patent has a large, active family (EP 3364868 B1, CN 108135520 B) and remains in force with an adjusted expiration of 2037-06-20, and a 4th-year maintenance fee paid 2024-05-07. Siemens is a large operating entity, not a monetization vehicle — which explains the absence of IPRs better than any technical factor does. Well-asserted patents attract IPRs; patents that are never asserted do not. The absence of PTAB activity here is therefore most consistent with the patent never having been asserted against a well-resourced competitor, rather than with it being unassailable.

Recommended next steps

If you are a defendant and the patent has claims invalidated: it does not. There is no FWD to link to, no cancellation to quote, and no disposition to lean on. Do not represent to a court or to opposing counsel that any claim of US 10,856,815 has been canceled — no tribunal has canceled anything. Verify the current claim set yourself at the USPTO Patent Center (application 15/752,276) before relying on the 14-claim set reproduced above.

If active proceedings are pending: none are. There is no institution-decision deadline, no oral hearing, and no statutory one-year FWD clock running on this patent. If in the coming months you or your client wants to create that clock, a petition must be filed first, and post-Fintiv the Board's § 314(a) discretion turns on the state of any parallel district court case.

If no PTAB activity exists — it doesn't — say so plainly, and read the signal. Three practical consequences:

1. File-wrapper due diligence before filing anything. Pull the full prosecution history from Patent Center and confirm which of the eight cited references and fifteen NPL items were actually applied versus merely listed. Only references substantively applied by the examiner are safe from § 325(d); references merely cited on the face of the patent are fair game, but a petitioner must brief the distinction.
2. Front-load the Pereira 2011 and Fyshe 2014 analysis. Both are pre-priority printed publications, both were cited to the Office, and both go to the core decode-then-generate architecture. Whether a § 103 ground built on them is "the same or substantially the same" as what the examiner considered under § 325(d) is the threshold fight, and it should be scoped before you spend on expert declarations.
3. Watch the family, not just the U.S. patent. The EP and CN counterparts are alive and were subject to their own prosecution (the record notes a Chinese Office Action dated 2020-03-31 and an extended European Search Report dated 2019-05-21). Foreign prosecution histories often surface art and claim-scope concessions that are useful U.S. invalidity evidence. Separately, I saw an unverified indication that the EP patent may have lapsed in at least one national jurisdiction — I could not confirm that from an authoritative register and you should not rely on it without checking the EPO Register directly.

Verification links for anything you file: USPTO PTAB E2E (proceeding search), USPTO PTAB Decisions, and CourtListener (any CAFC docket). As of the ODP's most recent ingest, all three are empty for US 10,856,815.

I'll research the assignment record. Let me run several targeted searches.

Let me dig into the assignment records and inventor backgrounds.

Ownership & Assignment Analysis — US 10,856,815 B2

Scope note / literal-identifier caveat. Assignment databases key off both patent numbers and application serials. Searching the bare string "10856815" returns records for a different patent — application serial 10/856,815 = US 7,374,452 B2 (NTT DoCoMo, "Battery Accommodating Structure and Mobile Terminal"). That record is unrelated to the patent analyzed here. All findings below correspond to US 10,856,815 B2, application 15/752,276, and, where relevant, to PCT/IB2016/001517. Search the Assignment Center by application 15/752,276 to reproduce.

Data limitation up front. I was able to retrieve the two recorded assignments via the Google Patents legal-events and assignment records supplied in the authoritative text (reel/frame, dates, assignors, assignees, conveyance type). I was not able to retrieve the correspondent of record (the attorney/agent who filed each recording) from available sources — the Assignment Center summary pages for reels 044908/0521 and 056832/0625 were not accessible to me in this session. I flag that explicitly rather than inferring.

Inventors

Unusual-pattern notes.

• Staggered execution dates on a single assignment. Reel 044908/0521 records signing dates spanning 2016-11-16 to 2017-02-22 — i.e., the three inventors signed at different times, and the assignment was not recorded until 2018-02-13, roughly 15 months after the earliest signature and after the PCT filing (2016-10-21). This is typical of a prosecution-stage chain-of-title cleanup, not a transaction.
• One inventor was an intern, not a full-time employee. Lazaridou's co-inventorship is consistent with her documented internship at the Princeton lab developing "a system based on recurrent neural networks for the decoding of brain activation signal into sequences of words" — the same subject matter as claim 10's recurrent-neural-network text generator. Inventorship by an intern is a benign but notable departure from the "all-employee-inventors" norm.
• Two of three inventors later left the Siemens orbit (Pereira to an academic/government research role; Lazaridou to a major AI lab). That is inventor mobility, not evidence of a portfolio fire-sale: the assignee retained title, and no inventor-side reversion was recorded. I found no evidence of an inventor exodus within 12 months of filing that would presage a sale.

Original assignee

Siemens Medical Solutions USA, Inc. (a Pennsylvania corporation; recorded as "SIEMENS MEDICAL SOLUTIONS USA, INC., PENNSYLVANIA" in the 2018-02-13 assignment; principal address Malvern, PA, with the relevant research lab in Princeton, NJ).

• Primary line of business: medical imaging hardware and software — the company is part of the Siemens Healthineers group (MRI, CT, molecular imaging, imaging software). The patented subject matter (decoding fMRI activation into natural-language text) was developed in a Siemens medical-imaging/computer-vision research lab.
• Did they ship a product embodying the claims? No evidence found. The technology is research-stage: it was funded by IARPA via the Air Force Research Laboratory under contract FA8650-14-C-7358 (a government research contract, i.e., pre-commercial), and no Siemens commercial product implementing the claimed fMRI→text pipeline was identified. Siemens is unquestionably an operating company, but that does not equate to a shipped product on these claims.
• Current status: operating. Siemens Medical Solutions USA, Inc. remains a U.S. operating subsidiary within Siemens Healthineers AG (carved out of Siemens AG, IPO'd 2018, Frankfurt: SHL). Not in bankruptcy; no dissolution, no Chapter 7/11, no fire-sale. The patent file is Active, with 4th-year maintenance fee paid 2024-05-07 and adjusted expiration 2037-06-20.

Assignment timeline

Two assignments are recorded against application 15/752,276 / patent 10,856,815. Neither is an NPE transfer.

1. Executed 2016-11-16 to 2017-02-22 / recorded 2018-02-13 — Reel 044908/0521

• Conveyance: Assignment of assignor's interest
• Assignor: Lou, Bin; Pereira, Francisco; Lazaridou, Angeliki
• Assignee: Siemens Medical Solutions USA, Inc. (Pennsylvania)
• Correspondent: Not retrievable from available sources. I could not open the reel/frame detail page in this session; I will not guess at the recording attorney. Flag: this is the single most important field for the NPE-forensics question and it is the one I could not confirm.
• Context: Initial inventor-to-employer assignment establishing chain of title during prosecution (recorded while the U.S. national-stage application was pending). Routine, not transactional.

2. Effective 2021-02-15 / recorded 2021-07-12 — Reel 056832/0625

• Conveyance: Confirmatory License (government license)
• Assignor: Siemens Medical Solutions USA, Inc.
• Assignee: The Government of the United States as represented by the Secretary of the Air Force (Ohio)
• Correspondent: Not retrievable from available sources.
• Context: Contractually mandated confirmatory license to the U.S. Government flowing from the IARPA/AFRL research funding (contract FA8650-14-C-7358). This is a license/interest recordation, not a transfer of title — Siemens remained and remains the owner. Under Bayh-Dole practice the government takes a non-exclusive license (and march-in rights), which is exactly what "confirmatory license" records.

No post-issuance ownership transfer exists. There is no assignment to any LLC, aggregator, or third party. On the record as it stands, Siemens Medical Solutions USA, Inc. is the current owner of record, with the U.S. Government holding a confirmatory license.

Timeline diagram

timeline
    title Ownership of US 10856815
    2015 : Priority text filed Oct 23
         : Second provisional Nov 2
    2016 : PCT filed Oct 21
         : Inventors sign assignment
    2017 : Final inventor signature Feb 22
    2018 : Assignment recorded Reel 044908 0521
    2020 : Patent issued Dec 8
    2021 : Govt license recorded Reel 056832 0625
    2024 : 4th year maintenance fee paid May 7

NPE / troll-pattern signals

Verdict

Insufficient data (no records, or only the original assignment).

The only recorded transfer of title for US 10,856,815 is the inventors' assignment to Siemens Medical Solutions USA, Inc., reel 044908/0521 (executed 2016-11-16 to 2017-02-22; recorded 2018-02-13). The only other record, reel 056832/0625 (effective 2021-02-15; recorded 2021-07-12), is a confirmatory license to the U.S. Government (Secretary of the Air Force) tied to IARPA/AFRL contract FA8650-14-C-7358 — an interest recordation, not an ownership change. Because there is no post-issuance assignment to analyze, the NPE framework yields a null result: every troll-pattern signal is negative or unverifiable, and the original operating-company assignee still owns the patent. The one field that could have sharpened the correspondent-recurrence analysis — the recording attorney on each reel — could not be verified, which is the only reason this is not a still-more-confident "clean operating-company hold" call.

Verify at: USPTO Assignment Center — https://assignmentcenter.uspto.gov/ (legacy: https://assignment.uspto.gov/patent/index.html). Search by application 15/752,276 (not the bare number "10856815," which collides with the unrelated NTT DoCoMo application 10/856,815). Cross-reference Google Patents legal events at https://patents.google.com/patent/US10856815/en.

Confidence. Reel/frame numbers, dates, assignors, assignees, and conveyance types are taken from the authoritative patent record and are high confidence. The correspondent-of-record field is not verified — treat any correspondent attribution for this patent as an open item to confirm on the Assignment Center detail pages for reels 044908/0521 and 056832/0625.

I'll search the USPTO database and related sources for the specific patent number 10856815 (= US 10,856,815 B2) and its cited references. Let me run targeted searches.

Let me gather details on the remaining cited references.

Prior-Art Analysis for U.S. Patent No. 10,856,815 (application 15/752,276)

Identifier confirmation (literal). The string "10856815" in the issued-patent sense is US 10,856,815 B2 — "Generating natural language representations of mental content from functional brain images," granted 2020-12-08, filed 2016-10-21 as app. 15/752,276 (PCT/IB2016/001517 → WO 2017/068414 A2), priority 2015-10-23 / 2015-11-02. I have kept to that patent only. (The same digits appear as an application number for an unrelated battery patent; that is not treated here.)

Search scope note. I worked from the authoritative full text you supplied (Google Patents, https://patents.google.com/patent/US10856815/en) and cross-checked individual references via Google Patents, FreePatentsOnline, Justia, EPO/PISE and the EP search report (EP3364868A4/B1). Two of my reference look-ups were cut off by a tool-step limit, so the descriptions of US 2011/0301431 A1 and US 2018/0092567 A1 rely on the bibliographic data in the patent itself plus partial results; treat those two as needing one confirmatory pull. All "§ 102" assignments below are preliminary analyst judgments, not legal conclusions — under § 102 a single reference must disclose every claimed element arranged as in the claim.

Threshold dates (AIA governs — filed after 2013-03-16)

• Critical date: 2015-10-23 (provisional 62/245,593). 2015-11-02 is the second provisional.
• § 102(a)(1): patents/printed publications/public uses on sale before 2015-10-23.
• § 102(a)(2): U.S. patent applications (published or granted) with an effective filing date before 2015-10-23, even if published later.

A. Patent citations on the face of US 10,856,815 (8 references)

In the Google Patents "Patent Citations" table, an asterisk marks a reference cited by the examiner; unmarked entries were cited by the applicant/third party. (That legend is stated explicitly for the neighboring "Cited By" tables; I apply it here with that caveat.)

1. US 2013/0184558 A1 — Gallant, Naselaris, Kay, Prenger (The Regents of the University of California). Apparatus and method for decoding sensory and cognitive information from brain activity.

• Filed 2012-12-21 (app. 13/725,893); published 2013-07-18; priority 2009-03-04; granted as US 9,451,883 B2 on 2016-09-27.
• Examiner-cited (*). https://patents.google.com/patent/US20130184558A1/en
• Description: Builds an encoding model that maps a stimulus's feature-space representation to voxel responses, then inverts it (a "decoding distribution," e.g., a multivariate Gaussian / MAP approximation) to rank and reconstruct the stimulus or mental state. Feature spaces explicitly include a "semantic" space built by SVD of a normalized term-document matrix (and WordNet, motion-energy, scene-category, Gabor). Claims recite multiple brain-imaging modalities (EEG/MEG/fMRI/fNIRS/SPECT/ECoG) and linear regression.
• Potential § 102: Available as § 102(a)(1)/(a)(2) art (published 2013). Its disclosure is the closest patent reference to the "decode a semantic feature from brain activity" idea, bearing on claims 1, 5, 12 (semantic decoding from imaging) and the modality-recited limitations. However, it teaches an encoding-model inversion, not the claim-1 step of decomposing imaging data into a weighted combination of basis images and mapping those weights to semantic-vector dimensions. So it is better cast as § 103 art combined with Pereira 2011 than as a clean § 102 anticipation of an independent claim. It does not address natural-language text generation, so claims 11/14 are untouched.

2. US 2008/0270120 A1 — Pestian et al. (Children's Hospital Medical Center). Processing text with domain-specific spreading activation methods.

• Filed 2008-01-04 (app. 12/006,813); published 2008-10-30; provisional 60/878,718 filed 2007-01-04; granted as US 8,930,178 B2 (2015-01-06) and continued (US 9,477,655; 10,140,288; 10,713,440).
• Examiner-cited (*). https://patents.google.com/patent/US20080270120A1/en
• Description: Natural-language processing of free text using spreading activation over semantic networks (UMLS concepts, conditional-probability link weights, decay/threshold constraints). Neuro-cognitive framing (recognition/semantic/episodic memory models).
• Potential § 102: Properly § 102(a)(1) art (published 2008). It addresses the text/semantic-representation half of the invention (assigning quantitative semantic structure to text — cf. claim 1's "training text is assigned semantic vectors" and claim 7's averaging concept in spirit) but contains no brain-imaging and no image-basis disclosure. Best treated as § 103 context for the semantic-text side; not an anticipatory reference for any independent claim.

3. US 2013/0178733 A1 — Langleben. Functional brain imaging for detecting and assessing deception and concealed recognition, and cognitive/emotional response to information.

• Priority 2001-06-15; published 2013-07-11.
• Description: fMRI-based detection of deception / concealed recognition / cognitive-emotional response to information.
• Potential § 102: § 102(a)(1) art, but directed at classification/detection, not generative decoding into semantic vectors or text. Relevant only to the "capturing functional brain imaging data" preamble of claims 1/5/12; not anticipatory. Background/§ 103 at best.

4. US 2011/0301431 A1 — The Board of Trustees of the Leland Stanford Junior University. Methods of classifying cognitive states and traits and applications thereof.

• Priority 2010-06-05; published 2011-12-08.
• Description: Multivariate classification of cognitive states and traits from brain imaging.
• Potential § 102: § 102(a)(1) art on the general "decode mental/cognitive state from brain images" concept; the reference's framing is classifier/classification of discrete states, whereas claim 1 requires a weighted basis-image decomposition mapped to semantic-vector dimensions. Background/§ 103; not anticipatory. (Bibliographic data from the patent face; one confirmatory pull advisable.)

5. US 2013/0246322 A1 — Cept Systems GmbH. Methods, Apparatus and Products for Semantic Processing of Text.

• Filed 2012-04-06; published 2013-09-19; priority 2012-03-15.
• Examiner-cited (*). https://patents.google.com/patent/US20130246322A1/en
• Description: Trains a self-organizing map to cluster documents by semantics, builds a "pattern dictionary" mapping each keyword to a SOM region, translates keyword sequences into pattern sequences, and trains a second (hierarchical/recurrent, e.g., HTM/MPF) neural network on those sequences for text classification/prediction.
• Potential § 102: § 102(a)(1) art bearing on the text-generation/neural-network limitations — notably claim 10 ("text generator based on a recurrent neural network") and the semantic representation of text (claims 7/13). No brain imaging → not anticipatory of claims 1/5/12; genuine § 103 candidate for the RNN-text-generation dependent claim.

6. US 2015/0269139 A1 — International Business Machines Corporation. Automatic Evaluation and Improvement of Ontologies for Natural Language Processing Tasks.

• Priority 2014-03-21; published 2015-09-24.
• Examiner-cited (*).
• Description: Automated ontology evaluation/improvement for NLP.
• Potential § 102: § 102(a)(1) art on semantic-text resources; peripheral. Not anticipatory; § 103 background only.

7. CN 104391963 A — 北京中科创益科技有限公司 (Beijing Zhongke Chuangyi Technology). Method for constructing correlation networks of keywords of natural language texts.

• Priority 2014-12-01; published 2015-03-04.
• Description: Builds keyword correlation networks from natural-language text.
• Potential § 102: § 102(a)(1) art (public before the critical date) on semantic/keyword representation of text. No imaging, no basis-image decoding → not anticipatory; § 103 context only. (Note: Chinese-language publication; machine translation required for any element-by-element comparison.)

8. US 2018/0092567 A1 — National Institute of Information and Communications Technology (NICT). Method for estimating perceptual semantic content by analysis of brain activity.

• Priority 2015-04-06; published 2018-04-05.
• Examiner-cited (*).
• Description: Estimates perceptual semantic content from brain activity.
• Potential § 102: Because it published after the 2015-10-23 critical date but has a pre-critical effective filing date (2015-04-06), it is available only as § 102(a)(2) art (U.S. application effective as of its filing date). It is the closest patent reference on the "decode semantic content from brain activity" axis and is potentially relevant to claims 1, 5 and 12; but absent a text-generation or basis-image-decomposition disclosure it is more realistically § 103 art. One confirmatory content pull is advisable (my look-up was truncated).

B. Non-patent literature on the face (the most probative art)

The EP search report (EP 3364868 A4/B1) categorized the following, and these — not the patents — are where anticipation pressure is highest:

Pereira, Lou, Lazaridou-era work: F. Pereira et al., "Generating Text from Functional Brain Images," Frontiers in Human Neuroscience 5:72 (2011-08-23), DOI 10.3389/fnhum.2011.00072. Marked [X] in the EP search — i.e., "particularly relevant if taken alone."

• Description: Reports generating text from fMRI by decoding semantic feature vectors of concrete-noun stimuli and mapping them to words/sentences (a co-inventor, Pereira, is a named inventor here).
• Potential § 102: Printed publication >1 year before the critical date → § 102(a)(1)/§ 102(b). This is the single reference most likely to be pressed as anticipatory of the core of claims 1, 5 and 12 (decode a semantic representation from functional brain images and emit text). Because it is the inventors' own prior work, expect an inventor-origination / § 102(b)(2)(A)-(B) or § 103-from-own-work discussion rather than a simple anticipation if litigation arose. In any element-by-element analysis, the claim-1 "decompose into a weighted combination of basis images" and "map weights to semantic-vector dimensions" steps are the limitations most plausibly missing from Pereira 2011, which is why the granted claims were probably drafted around them (see the framing note in the earlier summary).

Anderson, Bruni, Lopopolo, Poesio, Baroni, "Reading visually embodied meaning from the brain: Visually grounded computational models decode visual-object mental imagery induced by written text," NeuroImage 120:309-322 (2015), DOI 10.1016/j.neuroimage.2015.06.093. Also marked [X].

• Description: Decodes visual-object mental imagery cued by written text using image-based computational models.
• Potential § 102: Published 2015; if before the effective critical date it is § 102(a)(1) art. Bearing on claim 3 / claim 12 (text stimuli; natural-language representation of imaging data). Combines with Pereira under § 103.

Fyshe, Talukdar, Murphy, Mitchell, "Interpretable Semantic Vectors from a Joint Model of Brain- and Text-Based Meaning," Proc. ACL 2014 1:489-499, DOI 10.3115/v1/P14-1046 — joint brain/text semantic-vector model. § 102(a)(1) art (2014). Directly relevant to the semantic-vector-assignment steps (claim 1 preamble; claims 2, 7).

Simanova, Hagoort, Oostenveld, van Gerven, "Modality-Independent Decoding of Semantic Information from the Human Brain," Cerebral Cortex 24(2):426-434 (2014), DOI 10.1093/cercor/bhs324 — decoding semantic information independent of modality. § 102(a)(1) art; supports a § 103 case on claims 1/5/12.

Shinkareva, Malave, Mason, Mitchell, Just, "Commonality of neural representations of words and pictures," NeuroImage 54(3):2418-2425 (2011) — shared word/picture neural representations. § 102(a)(1) art; background on the training-stimulus side.

Friston et al., Statistical Parametric Mapping: The Analysis of Functional Brain Images (Academic Press, 2007) and S. M. Smith et al., "Advances in functional and structural MR image analysis and implementation as FSL," NeuroImage 23:S208-S219 (2004) — general fMRI analysis frameworks (GLM, motion correction). These underpin the GLM/HRF-deconvolution embodiment in the specification (FIGS. 8-11) and the "6-degree motion parameters" passage; § 102(a)(1)/§ 103 background, not anticipatory.

C. Bottom line — most relevant prior art

1. Most relevant overall: Pereira et al. 2011 ("Generating Text from Functional Brain Images") — designated [X] (relevant alone) by the EPO and authored in part by a named inventor; the primary § 102(b)/§ 103 reference against the decode-to-text core of claims 1/5/12.
2. Second, and the strongest patent reference: US 2013/0184558 A1 / US 9,451,883 B2 (UC Regents; Gallant et al.) — semantic feature-space decoding of brain activity; § 102(a)(1)/(a)(2) art, but an encoding-model inversion rather than claim-1's basis-image decomposition → chiefly § 103.
3. Best § 102(a)(2)-only reference: US 2018/0092567 A1 (NICT) — earlier effective filing date (2015-04-06), later publication.
4. Text-generation / neural-network limitation (claim 10): US 2013/0246322 A1 (Cept Systems) — SOM + recurrent/HTM neural network for semantic text processing.
5. Remaining face citations (Pestian US 2008/0270120; IBM US 2015/0269139; CN 104391963; Langleben US 2013/0178733; Stanford US 2011/0301431) are background/§ 103 references — none discloses the combination of (i) functional brain imaging, (ii) basis-image decomposition of that data, (iii) mapping decomposition weights to semantic-vector dimensions, and (iv) text generation from the decoded vector, so none appears anticipatory of independent claims 1, 5 or 12 standing alone.

Caveats

• Anticipation is element-by-element; my § 102/§ 103 characterizations are preliminary and not legal conclusions.
• I could not fully verify the contents (as opposed to the bibliographic data) of US 2011/0301431 A1 and US 2018/0092567 A1 before hitting a tool-step limit — recommend one confirmatory full-text pull each.
• The examiner-vs-applicant "cited by" asterisk convention in the Google Patents citation table should be double-checked against the USPTO face of the patent (Form PTO-892/1449) before relying on it.
• Nothing here reflects prosecution-history estoppel; the granted-claim basis-image wording vs. the specification's X ≈ ZB forward-model framing (flagged in the earlier summary) is a prosecution-history question I did not resolve.

I'll ground this analysis in the prior art referenced in the patent's Prior Art section. Let me verify the content of the key references.

Excellent — Pereira 2011 is highly material. Let me verify the other cited prior-art references.

Obviousness Analysis Under 35 U.S.C. § 103 — U.S. Patent No. 10,856,815 B2

Analysis date: April 26, 2026 (per this task's instruction; note the previously generated sections in this file are timestamped 2026-09-28 — a metadata inconsistency worth flagging, though it does not affect the substance below).

Scope and method. This analysis builds on, and does not repeat, the previously generated Patent Summary and Litigation Summary. Those sections found no litigation (district court, PTAB, ITC, or Federal Circuit) involving this patent. This is therefore a hypothetical validity analysis performed against the prior-art references identified on the face of the patent and in its Prior Art section. I have prioritized the authoritative full patent text supplied, and have verified the cited references by search where possible. Where I rely on partial disclosure (abstract or selected specification passages) rather than a full-text review, I say so.

1. Legal framework

The § 103 inquiry follows Graham v. John Deere Co., 383 U.S. 1 (1966): (1) scope and content of the prior art; (2) level of ordinary skill; (3) differences between the prior art and the claims; (4) secondary considerations. Under KSR Int'l Co. v. Teleflex Inc., 550 U.S. 398 (2007), a claimed combination is obvious where (a) the elements were known in the art, (b) there was a predictable or finite number of identified solutions, and (c) a POSITA had a reasonable expectation of success. An articulated reasoning with a rational underpinning (e.g., same field of endeavor, same problem, predictable use of a known technique) is required — In re Kahn / KSR.

Critically for this patent, the entire inventive concept is a pipeline of well-known mathematical techniques: latent semantic vector representation (from corpus statistics), matrix factorization / ridge regression (from neuroimaging multivariate analysis), and sequence generation from a vector (from NLP/ML). The specification itself concedes each piece: "ridge regression is used to factorize"; "Other factorization algorithms may be used, such as different types of regularization… L1 regularized logistical regression and singular value decomposition"; "GloVe… Skip-Thought vectors"; "the text generation module is based on training a Recurrent Neural Network… Other machine learning and data mining may be used." These are admissions of known techniques.

2. The patent's claimed subject matter (for the comparison)

Reproduced from the previously generated summary — the three independent claims share this core:

Note the framing gap flagged in the earlier summary: the specification describes a forward model (X ≈ ZB) and a reverse decode, while the granted claims are worded as "decomposing" the image into weighted basis images and "mapping" the weights to semantic dimensions. That recasting does not change the § 103 analysis because the same factorization (X ≈ ZB, solved as per-voxel regressions) underlies both framings.

3. Level of ordinary skill in the art (POSITA)

A POSITA would have a graduate degree (M.S./Ph.D.) in cognitive neuroscience, biomedical imaging, or machine learning, or equivalent experience, with working knowledge of (i) fMRI/MEG/EEG preprocessing and general linear models, (ii) multivariate pattern analysis / regularized regression, (iii) distributional semantic models (LSA, LDA, word embeddings), and (iv) basic neural-network sequence modeling. The patent's own cited NPL — Friston SPM (2007) and Smith FSL (2004) — are the standard toolboxes such a person would use; both were in the record and both are conventional. This is a high-skill, fast-moving, academically crowded art, which cuts toward obviousness of incremental combinations.

4. The prior art of record and what each teaches

4.1 Pereira et al. 2011 — the primary reference (and the closest)

"Generating Text from Functional Brain Images," Pereira, Detre & Botvinick, Front. Hum. Neurosci. 5:72 (2011), DOI 10.3389/fnhum.2011.00072 — by co-inventor Francisco Pereira. This is cited on the face of the patent as an "X"-category reference in the EP search report and appears in the NPL citations (https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2011.00072).

Verified disclosures (from the full text and its Appendices, retrieved above):

• fMRI was collected as participants read names of concrete items ("training text") — satisfying the "stimuli comprising training text" element.
• A semantic representation was built from text data (LDA topic-probability vectors over a corpus) — i.e., multidimensional semantic vectors assigned to the text, context-based by construction.
• The paper's Appendix A.3, "Basis image decomposition," states: "Each example x will be expressed as a linear combination of basis images b₁, …, b_K of the same dimensionality, with the weights given by the topic probability vector z." It defines X (n×m images), Z (n×K topic probabilities), and B (K×m basis images).
• Learning basis images = per-voxel regression: "X(:,j) … are predicted from Z using regression coefficients β = B(:,j)" — explicitly with a ridge term ("Any situation where linear regression was infeasible … addressed by using a ridge term with the trade-off parameter set to 1"; maximize ‖X:,j − Zβ‖² + λ‖Z‖²).
• Decoding a held-out image = "Predict topic probabilities given example images and basis images": "predicting the topic probability vector z … is a regression problem where x′ is predicted from B′ using regression coefficients z′."
• Text output: "we were able to generate from each one a collection of semantically pertinent words (e.g., 'door,' 'window' for 'Apartment')" via p(w|θ) = Σₖ p(w|topic_k)θₖ — a text output derived from the decoded semantic vector.

This is essentially the patent's Figures 5A–5C and the specification's own equations, published ~4 years before the priority date.

4.2 Fyshe et al. 2014 (ACL) — brain- and text-based semantic vectors

"Interpretable Semantic Vectors from a Joint Model of Brain- and Text-Based Meaning," Fyshe, Talukdar, Murphy & Mitchell, Proc. ACL 2014, pp. 489–499 (verified above). Teaches JNNSE, a matrix-factorization model that jointly learns semantic vectors from corpus statistics and fMRI activation recorded while people read words, producing interpretable, sparse, non-negative dimensions, and shows the resulting vectors predict brain activity and vice versa. Directly supplies: semantic vectors grounded in both text and brain data; interpretable dimensions ("chair" → furniture/supervisor topic clusters); and the mapping structure between semantic dimensions and voxels.

4.3 Anderson et al. 2015 (NeuroImage) — decoding meaning induced by written text

"Reading visually embodied meaning from the brain…" Anderson, Bruni, Lopopolo, Poesio & Baroni, NeuroImage 120:309–322 (verified above). Re-analyzes fMRI acquired while subjects read object names, applies text-based and image-based semantic models, and decodes representations using representational-similarity analysis. Establishes that the mental content triggered by reading words is decodable with computational semantic models — the exact "text stimulus → semantic vector → brain pattern" linkage the patent relies on.

4.4 US 2018/0092567 A1 — NICT (Nishimoto & Kashioka)

"Method for estimating perceptual semantic content by analysis of brain activity" (US 15/564,071; JP priority 2015-04-06). Verified disclosures relevant to the claims:

• Associates a semantic-space representation of a training stimulus with the measured brain-activity output ("associating a semantic space representation of the training stimulation and the output of the brain activity detection means in a stored semantic space").
• For a novel stimulus, "The output of the brain activity detection means … is decomposed as, for example, a linear synthesis of the output … induced by the training stimulation … and thereby a perceptual semantic content in response to the novel stimulation can be obtained as a linear synthesis… On the basis of a coefficient of the linear synthesis and the association …, a probability distribution in the semantic space … can be obtained." — i.e., decompose new brain data into a weighted linear combination of training brain-image patterns and read off the semantic-space representation from the weights.
• Embodiment 2 applies an LDA topic model and "a sentence can be estimated by a method like LDA" — a text/sentence output from the decoded semantic content.

Timing: JP priority 2015-04-06 predates the patent's 2015-10-23 priority; the US national-stage publication is therefore available as AIA § 102(a)(2) prior art (assuming the relied-upon disclosure is supported by the JP priority document, which is likely given the PCT claims that priority). It is cited on the patent's face, so the examiner had it.

4.5 US 2013/0184558 A1 (now US 9,451,883 B2) — Gallant, Naselaris, Kay & Prenger, UC Regents

"Apparatus and method for decoding sensory and cognitive information from brain activity." Verified disclosures: an encoding model maps a stimulus into a linearizing feature space (including a semantic feature space built from a term-document matrix and SVD — f_semantic = U·w), fits the feature-to-voxel relationship linearly per voxel, then inverts/transforms the encoding model into a decoding model to map novel brain activity back into the feature space and reconstruct the stimulus using a prior/decoding database. This is the reference family the patent's own Background discusses as "generative brain decoding," and it supplies the explicit encode → invert → decode-into-semantic-space → reconstruct architecture.

4.6 Other references of record (title-level; used for secondary combinations)

• US 2013/0246322 A1 (Cept Systems GmbH) — "[Methods/]Apparatus and Products for Semantic Processing of Text": NLP semantic processing of text.
• US 2015/0269139 A1 (IBM) — automatic evaluation/improvement of ontologies for NLP tasks.
• US 2008/0270120 A1 (Pestian) — processing text with domain-specific spreading-activation methods (semantic text modeling).
• US 2011/0301431 A1 (Leland Stanford Junior Univ.) — classifying cognitive states and traits from brain data.
• US 2013/0178733 A1 (Langleben) — functional brain imaging for detecting/assessing cognitive responses to information.
• CN 104391963 A — constructing correlation networks of natural-language-text keywords (corpus co-occurrence modeling).
• Shinkareva et al. 2011 (NeuroImage 54:2418) — commonality of neural representations of words and pictures (modality independence of semantic representation).
• Simanova et al. 2014 (Cereb. Cortex 24:426) — modality-independent decoding of semantic information.
• Friston, SPM (2007) and Smith, FSL, NeuroImage 23:S208 (2004) — the standard GLM/neuroimaging pipelines the patent's specification invokes (deconvolution, GLM regressors).

Confidence note: 4.1–4.4 were verified against full text/quoted passages. 4.5 was verified against the granted US 9,451,883 B2 text and its claims (as reproduced by multiple sources). 4.6 is relied on at title/abstract level and used only as corroborating art for elements independently disclosed by 4.1–4.4.

5. Grounds of rejection

Ground I (strongest): Claim 1 is obvious over Pereira 2011 alone; at most in view of the admitted known text-generation technique.

Element-by-element mapping of claim 1 to Pereira 2011:

Motivation / rationale: Pereira 2011 is by the same inventor, in the identical field, addressing the identical problem ("generate text about the mental content reflected in brain images"), using the identical mathematics the patent recites. For the system claim 5, Pereira's four functional modules (semantic model, basis learning, decoding regression, word generation) map onto the claimed representor / basis learner / decoder / text generator. For claim 12, every element is likewise present.

The only plausible gap is that claim 1's "text output" and (dependent) claims 11/14's "sentence" may exceed Pereira's word-list output. That gap is closed by NICT US 2018/0092567 ("a sentence can be estimated by a method like LDA") and by the admitted known technique of sequence-generation-from-a-vector (the patent itself: "the text generation module is based on training a Recurrent Neural Network"; claim 10). Under KSR, using a known sequence generator (RNN) to lengthen word output into sentences is the predictable use of a known technique — and the applicant's own specification concedes it.

Expected result: predictable; Pereira already produced a functioning decode-and-generate pipeline; substituting an RNN/LDA sentence generator for the word-probability readout changes the output format, not the principle of operation.

Ground II: Claims 1/5/12 obvious over Pereira 2011 in view of NICT (US 2018/0092567)

Even setting aside characterization issues with Pereira's stimuli, NICT independently supplies the two elements claimants might argue are missing from Pereira:

• NICT supplies "identify a linear combination of basis images in the second set…, decode… by estimating a weight of each basis image" ("linear synthesis of the output … induced by the training stimulation" with coefficients) and "text/sentence" output (LDA).
• Pereira 2011 supplies the text stimuli and the explicit ridge-regression basis-decomposition formulation.

Motivation: both references are in the same field (fMRI semantic decoding), solve the same problem (estimating arbitrary/novel content beyond a fixed stimulus set), and use the same technique (linear decomposition of brain images onto a semantic space). Combining them is the straightforward union of two known solutions with no change in their respective principles of operation. KSR, 550 U.S. at 417.

Ground III: Claims 1/5/12 obvious over Gallant (US 2013/0184558) in view of Pereira 2011 and Fyshe 2014

• Gallant supplies the general architecture: stimulus → linearizing feature space (including a semantic space via SVD of a term-document matrix) → per-voxel linear encoding model → invert to decode → map novel brain activity into the feature space → reconstruct. It expressly targets novel content ("It cannot be used to classify images that belong to novel … classes" — the stated deficiency being solved).
• Pereira 2011 replaces Gallant's "rank and combine top-k database items" reconstruction with a basis-image factorization that does not require an exhaustive stimulus database — precisely the deficiency the patent's own Background identifies and purports to solve. That is an express motivation to modify Gallant.
• Fyshe 2014 supplies brain-and-text jointly learned, interpretable semantic dimensions and confirms bi-directional predictivity between semantic vectors and voxel patterns.

Motivation: Gallant's specification states its framework is "agnostic to the precise nature of the brain measurements" and applicable to "any sensory or cognitive brain system"; Pereira/Fyshe supply the text-semantics-based feature space and the factorization that avoids Gallant's database limitation. Same field, same problem, complementary known techniques → obvious to try with reasonable expectation of success.

Ground IV (supporting): Claims 1/5/12 obvious over Anderson 2015 in view of Fyshe 2014 and NICT

Anderson establishes that fMRI acquired while a subject reads written text encodes decodable semantic structure that text-based computational semantic models capture, and that multimodal (text + image + brain) models improve decoding. Fyshe provides the joint brain/text semantic vector factorization. NICT provides the "decompose new activity into a linear synthesis of training patterns → probability distribution in semantic space → sentence" step. Together these teach the claimed pipeline; the motivation is the shared objective of recovering linguistic meaning from text-evoked brain activity.

6. Dependent claims

Claim 9 is essentially pre-disclosed by Pereira 2011. Claim 7 is an averaging step expressly identified as conventional in the patent's own text.

7. Secondary considerations (objective indicia)

There is no evidence in the record of secondary indicia that would rebut a prima facie case:

• No evidence of commercial success, licensing, or industry acquiescence tied to the claims (the earlier sections found no litigation and no identified product).
• No evidence of copying of the claimed invention by others.
• No evidence of unexpected results — the claimed pipeline is the union of known techniques whose combination the art already performed (Pereira 2011 is the pipeline).
• Any "long-felt need" argument would be weak: Pereira 2011 and Gallant already addressed the same need; the patent's asserted advance over Gallant (avoiding a stimulus database) is exactly the advance Pereira 2011 already made.
• Government interest: the invention arose under IARPA/AFRL contract FA8650-14-C-7358, with a confirmatory license to the U.S. Government (recorded 2021-07-12). This is a funding/licensing fact, not an objective indicium of non-obviousness.

8. Consolidated assessment

Bottom line. On the record before me, the independent claims 1, 5, and 12 of U.S. 10,856,815 are vulnerable to a § 103 obviousness challenge, and — more pointedly — claim 1 appears to be substantially disclosed by Pereira et al. 2011, which is prior art under AIA § 102(a)(1) (published Aug. 2011, more than one year before the Oct. 23, 2015 priority date, and not saved by the § 102(b)(1)(A) inventor-disclosure grace period). The Pereira 2011 reference is by co-inventor Francisco Pereira and was before the examiner, yet the granted claims recite precisely the "decompose into a weighted combination of basis images → map weights to semantic dimensions → estimate weights on new data" mathematics that Pereira's Appendix A.3 sets out. That is a significant validity-exposure finding.

The strongest defensive arguments available to the patentee would be:

1. Claim differentiation / narrow construction — arguing that claim 1's "text output" and dependent claims 11/14 require sentences (RNN-generated natural language), which Pereira's word lists do not literally disclose. But NICT's LDA-sentence estimation and the admitted RNN technique blunt this.
2. Non-analogous / different problem — arguing Pereira used concrete nouns with line drawings, not free text passages. Anderson 2015 (text-induced decoding) and Fyshe 2014 (reading single words) close this.
3. Teaching away — relying on the patent's own criticism of Gallant's database requirement. This backfires, because Pereira 2011 supplies the database-free alternative.

9. Caveats and items I could not verify

• I did not obtain full claim charts of every one of the eight cited patents/applications; Grounds III/IV (Gallant, Fyshe, Anderson, NICT) are grounded in verified specification/abstract passages and the granted Gallant claims, while § 4.6 references (Cept, IBM, Pestian, Stanford, Langleben, CN 104391963) are relied upon at title/abstract level only and are corroborative rather than load-bearing.
• NICT date assumption: I have assumed the relied-upon subject matter (linear-synthesis decomposition, semantic-space probability distribution, LDA sentence estimation) is supported by the JP 2015-077694 priority document. If it were shown to lack that priority support, NICT's effective date would shift to its 2016-04-05 PCT filing — after the patent's 2015-10-23 priority — and NICT would drop out as prior art (leaving Ground II unavailable, though Grounds I, III, IV survive).
• No litigation exists to have produced claim-construction rulings or invalidity findings; the earlier Litigation Summary confirms this. The above is therefore a prospective, hypothetical § 103 analysis, not a prediction of any adjudicated outcome.
• Two issues are out of scope here but bear on overall validity and should be separately assessed: § 101 (these claims recite a computational pipeline over mental content, raising the Alice/Mayo two-step question), and § 112 (the specification's Appendix text closely tracks Pereira 2011 — an enablement/utility question for "decoding mental content" claims). I flag these only; I have not analyzed them.
• No contradiction found between this analysis and the previously generated Patent Summary and Litigation Summary; the only inconsistency noted is the metadata date discrepancy in § 1 above.