.jpg) |
| Arpita Bhattacharyya |
 |
| Luigi Distefano |
.jpg) |
| Kelly Horn |
At the USPTO's fourth AI/Emerging Technologies Partnership meeting on Sept. 27, one of the salient discussion topics was training for patent examiners to enhance the quality and efficiency of examination of AI technologies across various technology areas.
But as examiners in different art units are increasingly examining patent applications involving AI, one question that is becoming ever more important for examiners and patent applicants alike is the amount of disclosure of AI operation, including source code, required to receive patent protection.
Such considerations of disclosure requirements are timely and much needed because at present, in both the U.S. and in Europe, AI inventions are facing unclear standards for adequate written description under Title 35 of the U.S. Code, Section 112(a), and sufficiency of disclosure under Article 83 of the European Patent Convention.
Although the level of detail required to satisfy these disclosure requirements differ in the two jurisdictions, they both require sufficient implementation details to show that the inventor had possession of the invention in the U.S. and a person skilled in the art reading the disclosure can carry out the invention in Europe.
Depending on the nature and scope of the claims and the complexity and predictability of the AI models used, patent applications for AI inventions may require more implementation details than traditional computer-implemented inventions, including disclosure of the types of training data and methods used to train the AI systems.
U.S. applicants must provide sufficient implementation details for AI inventions.
Although the patenting of AI-based technology is guided by the same legal framework as other types of inventions, certain pitfalls unique to this technology may be encountered during prosecution.
Patent applicants in the U.S. should be mindful of the type of support required in the patent application for AI-related inventions to comply with the written description requirement under Section 112(a).
Failure to include the proper support in the originally filed application may preclude issuance of a patent even if the other patentability requirements, including eligibility, novelty and nonobviousness, are met.
The Patent Trial and Appeal Board has repeatedly found that if applicants claim a desired algorithmic result, the patent specification must provide specific implementation details tied to that result to demonstrate that the inventor had possession of the underlying algorithms.
Consider, for example, the 2020 decision in Ex Parte Buhrmann, where the PTAB affirmed an examiner's finding of lack of written description for claims directed to authenticating entities or individuals engaging in electronic transactions.[1]
The PTAB found inadequate support in the specification for how to achieve the claimed result of determining a so-called risk score because the specification did not demonstrate that the inventor had possession of an algorithm for determining the claimed risk score.[2]
Instead, the specification merely disclosed use of weighted parameters, with the weighting factors being given "values that represent the importance of the parameters."[3]
While the specification provided examples of how the factors may be weighed, those disclosures were found to be too generic to demonstrate that the inventor had possession of a specific algorithm for achieving the claimed result of determining a risk score, and therefore the pending claims were found to lack written description support.[4]
A similar argument can also be made for lack of enablement if the specification does not describe how a claimed result is achieved.
The PTAB has found inadequacy in the written description even where the specification explicitly relied on the use of known AI algorithms and techniques.
In Ex Parte Allen case in 2020, for example, the claims were directed to using AI to create a patient-specific medication listing and modifying the list to remove certain medications based on a threshold analysis.[5]
Like in Ex Parte Buhrmann, the PTAB found that the specification did not provide adequate disclosure of how the claimed result was achieved.[6]
The PTAB reached this conclusion even though the specification explicitly disclosed use of known "artificial intelligence logic, such as natural language processing (NLP) based logic" and "machine learning logic."[7]
The PTAB found this description generic and insufficient to support the claimed features because it did not explain how the AI-based techniques could be "used to achieve the particular desired results claimed (i.e., determining whether to remove a medication from the medication list based on the enumerated factors)."[8]
Although the specification in Ex Parte Allen provided more implementation details than in Ex Parte Buhrmann, the written description still did not pass muster under Section 112 because it failed to provide a concrete algorithm for any particular implementation.[9]
The generic disclosure in the specification of how one could weigh and aggregate scores and the specification's reliance on known AI algorithms were found to be inadequate in describing the claimed method for determining whether the medication listing data structure was to be modified.[10]
In contrast, the PTAB has approved written descriptions that disclose use of known AI algorithms, provided sufficient implementation details are included for how to use that algorithm to achieve the claimed result.
For example, in Ex Parte Kirti in 2021, the claims were directed to training and using machine learning models with user data to determine membership in groups.[11]
The PTAB found the disclosure sufficient because it provided adequate implementation details to train and use the generic machine learning algorithm to achieve the claimed result.[12]
The PTAB also found that the specification adequately disclosed the claimed training sets and the training inputs and outputs of the model, such that a skilled artisan could build the claimed cluster models based on the disclosure, thus satisfying the written description requirement.[13]
Although these PTAB decisions shed some light on the level of implementation detail required to satisfy the written description requirement, there is still a lot of uncertainty, particularly when the applied AI model is not well known.
Based on the decisions discussed above, it appears that the USPTO is leaning toward requiring specific implementation details, including description of the training data sets and the AI algorithm, such that a skilled artisan can build the claimed system for achieving the claimed result based on the disclosure.
Only time will tell whether these disclosure requirements will ease as AI becomes more prevalent and certain aspects of AI become general knowledge to those skilled in the art, or whether the U.S. will move more toward the even stricter written disclosure requirements in Europe.
AI inventions face stricter written disclosure scrutiny in Europe compared to the U.S.
Despite the rapid increase in AI-related patent applications filed with the European Patent Office in recent years, the body of relevant case law of the boards of appeal is relatively sparse.
The decisions issued by the Technical Boards of Appeal, or TBA, demonstrate that the EPO typically applies a stricter standard than the USPTO when assessing sufficiency of disclosure.
For example, in decisions T 1539/20 of November 2022, T 0606/21 of February 2023, and T 1526/20 of March 2023, the TBA refused respective patent applications for lack of sufficiency of disclosure, while the USPTO granted the counterpart U.S. patent applications.
T 1539/20 was an appeal from the Examining Division's refusal of European Patent Application No. 12814717.0, which related to a method of monitoring performance of an application system distributed across a plurality of network connected nodes. This included "mapping a given distributed application system to a hierarchical model."[14]
The TBA refused the application, finding that it did "not contain any information explaining how a skilled person could implement the mapping process in software."[15]
Although the application disclosed that automated learning systems could be used for implementing the mapping process, the TBA reasoned that "this vague hint alone does not enable the skilled person to implement such automated knowledge discovery."[16]
The application did not describe any specific kind of training data or how the learning system could extract the necessary information from the monitored application system.
T 0606/21 was an appeal from the Examining Division's refusal of European Patent Application No. 7847781.6, which related to a method for evaluating predictions of trajectories by autonomous vehicles, including generating a Deep Neural Network, or DNN.
The TBA refused the application because it did not explain how the DNN could output a similarity score indicating whether a predicted trajectory was close to an actual trajectory when the actual trajectory was undefined, or how the DNN could provide a reliable output when the DNN was fed with incomplete data.
T 1526/20 was an appeal from the Examining Division's refusal of European Patent Application No. 15163789.9, which related to a computer-implemented liveness testing method for distinguishing between live faces and impersonations based on 2D images.
The TBA concluded that not all required image processing steps were described, and consequently expressed doubt whether the invention as disclosed in the application would reliably achieve the desired technical effect.
To comply with the stricter assessment of sufficiency of disclosure before the EPO, a patent application must explain what technical problem the underlying AI-related invention solves and how.
This requires describing the AI architecture and algorithms enabling the AI-related invention to solve the stated technical problem, and often will require characterizing the training data to enable the AI-related invention to be reproduced.
A preeminent decision on AI-related inventions by the TBA is T 0161/18 of May 2020, which explained the requirement to characterize training data — e.g., by describing the type of training data — when claiming specific applications of AI.
T 0161/18 was an appeal from the Examining Division's refusal of European Patent Application No. 06804383.5, which related to a method for determining cardiac output by using a trained neural network to transform an arterial blood pressure curve measured at a peripheral region to an equivalent aortic pressure.
The weighting values of the neural network were obtained by supervised training of the neural network.
Although the patent application disclosed the importance of using a wide spectrum of possible input values during training — obtained from patients of different sexes, ages, constitutional types and health conditions — to avoid overspecialization of the network, the application failed to characterize or provide examples of the data inputs.
The TBA refused the application, explaining that failure to characterize or provide examples of the training data meant that a skilled person could not reproduce the trained neural network.
T 1191/19 of April 2022 was an appeal from the Examining Division's refusal of European Patent Application No. 08877672, and expanded on the TBA's guidance for sufficiency by requiring disclosure of the AI architecture in addition to characterizing the training data.
The invention related to an AI-based neurorehabilitation method for predicting personalized neuroplasticity interventions using a meta-learning scheme having a patient database to obtain relevant classifiers.
The TBA commented that the application did not disclose any example sets of training data or validation data, nor "the minimum number of patients from which training data should be compiled to give a meaningful prediction and the set of relevant parameters."[17]
The TBA further commented on the lack of information disclosing "the structure of the artificial neural networks used as classifiers, their topology, activation functions, end conditions or learning mechanism."[18]
Concluding that at "the level of abstraction of the application, the available disclosure is more like an invitation to a research program,"[19] and not a sufficient disclosure, the TBA refused the application.
Ultimately, the EPO assesses whether a claimed invention is sufficiently disclosed on a case-by-case basis, taking into consideration the technical contribution the invention makes to the state of the art. The discussed TBA decisions highlight the importance of outlining clearly how the technical problem addressed by the AI is solved.
While training datasets do not need to be disclosed, the training data must be sufficiently characterized to enable the training process to be reproduced. The AI architecture and algorithms may need to be disclosed, while target data structures accounted for at every stage of their processing.
In conclusion, the assessment of sufficiency of disclosure requires determining if a person of ordinary skill in the art. Or a person of ordinary skill in the art, could, based on the disclosure and the common general knowledge, be able to reproduce the invention as claimed.
Since the USPTO and the EPO interpret the characteristics of the person of ordinary skill in the art differently, it is not surprising that the two offices often reach different conclusions on the sufficiency of written disclosure.
The person of ordinary skill in the art envisioned by the USPTO is not an automaton and has greater creative abilities than the EPO's person of ordinary skill in the art.
This difference in how a skilled artisan is viewed by the USPTO and the EPO may affect the level of detail required to satisfy the written description or sufficiency of disclosure requirements before each office.
Arpita Bhattacharyya is a partner, and co-leader of the AI industry working group, at Finnegan Henderson Farabow Garrett & Dunner LLP.
Luigi Distefano is an associate at the firm.
Kelly Horn is an associate at the firm.
Finnegan partner Yelena Morozova, and technical specialists Filip Wach and Samir Ismail, contributed to this article.
The opinions expressed are those of the author(s) and do not necessarily reflect the views of their employer, its clients, or Portfolio Media Inc., or any of its or their respective affiliates. This article is for general information purposes and is not intended to be and should not be taken as legal advice.
[1] Ex parte Michael F. Buhrmann, Charles L. Dennis, & Jeffrey Brennan, No. 2020‑001504, 2020 WL 4383650, at *4 (P.T.A.B. July 27, 2020).
[2] Id. at **3-4.
[3] U.S. Patent Application No. 13/752,271, Jan. 28, 2013, Specification, [0056].
[4] Ex parte Buhrmann, 2020 WL 4383650, at **3-4.
[5] U.S. Patent Application No. 15/339,973, Mar. 27, 2020 Appeal Brief at pp. 68-69, 71-73.
[6] Ex parte Corville O. Allen, Timothy A. Bishop, Albert A. Chung, & Elizabeth A. Scheiber, No. 2020-005211, Decision on Appeal at p. 56 (P.T.A.B. Dec. 2, 2021).
[7] U.S. Patent Application No. 15/339,973, Nov. 1, 2016, Specification, [0057].
[8] Ex parte Allen, Decision on Appeal at p. 57.
[9] U.S. Patent Application No. 15/339,973, Nov. 1, 2016, Specification, [0160].
[10] Ex parte Allen, 2020 WL 4383650, at **58-59.
[11] Ex parte Rituraj Kirti, Sue Ann Hong, & Leon R. Cho, No. 2020-000527, 2021 WL 2103251, at *1 (P.T.A.B. May 21, 2021).
[12] Ex parte Kirti, 2021 WL 2103251, at *4.
[13] Ex parte Kirti, 2021 WL 2103251 at *4.
[14] T 1539/20 Reasons for the Decision 3.1
[15] T 1539/20 Reasons for the Decision 3.2
[16] T 1539/20 Reasons for the Decision 3.3
[17] T 1191/19 Reasons for the Decision 4.1
[18] Id.
[19] Id.
For a reprint of this article, please contact reprints@law360.com.
