Within several weeks, the Federal Circuit has held that two Stanford haplotype estimation patent applications lack patent eligibility because their claims are directed to abstract mathematical calculations. The applications covered improved techniques for predicting which collection of alleles are associated with a single parent. In support of patent eligibility, Stanford argued that the claims led to increased accuracy in predicting which genes are associated with a parent. Despite acknowledging this increased accuracy and its benefits, the court held that the increased accuracy in prediction amounted to an improvement in the computational accuracy of an abstract mathematical algorithm. Such an improvement is not deemed as an enhancement to a technological process, which is eligible, but is instead deemed as an enhancement to an abstract idea, which is patent ineligible.
Like the innovation underlying Stanford's patent applications, machine learning (ML) involves computer-implemented probabilistic models that are used to generate a prediction from complex input data. Consequently, claims directed to ML may also be at risk of being rejected as a patent ineligible abstract idea. This risk is far more significant for claims directed to generally-applicable ML than for claims directed to ML specifically applied to another technological area.
Examples of ML specifically applied to another technological area include image and facial recognition and intrusion and imminent failure detection in a computer system. Examples of generally-applicable ML include improvements to fundamental ML algorithms and ML design aided by ML, such as ML aided feature engineering, ML aided hyperparameter value exploration and optimization, and ML aided machine algorithm selection.
Improved predictive accuracy is very often a primary goal and benefit of many generally-applicable ML innovations. The Federal Circuits findings on the relevance of improved predictive accuracy to patent eligibility is thus particularly pertinent to generally-applicable ML innovations.
Patent applications directed to generally-applicable ML should describe features that support patentable eligibility to overcome potential rejections as a patent ineligible abstract idea. While the subject matter of the Stanford haplotype estimation applications may not be viewed as ML, the Federal Circuit's decisions show that, if a ML claim is deemed to be directed to abstract mathematical algorithms or calculations, improved predictive accuracy may not save the claim from patent ineligibility. To achieve patent eligibility, patent applications directed to generally-applicable ML should also describe other factors that support patent eligibility, as further discussed below.
Stanford appealed from final rejections of patent application nos. 13/445,925 ("925 application") and 13/486,982 ("982 application"). The court addressed the 925 application in In Re Board of Trustees of The Leland Stanford Junior University, Case No. 20-1012 (Fed. Cir. Mar. 11th, 2021) herein Stanford I, and the 982 application in In Re Board of Trustees of The Leland Stanford Junior University, Case No. 20-1288 (Fed. Cir. Mar. 25th, 2021), herein in Stanford II.
The claims of both applications cover an improved method for haplotype phasing, namely determining which alleles are associated with the same chromosome. In other words, "haplotype phasing is a process for determining the parent from whom alleles—i.e., versions of a gene—are inherited."
To reject both applications' claims as patent ineligible, the Patent and Trademark Appeal Board ("Board") applied the two step Alice inquiry. For step 1, the Board found the claims were directed to abstract ideas "in the form of mathematical concepts." For step 2, the Board found the claims did not include additional limitations that provide an inventive concept that transforms the abstract idea into patent eligible subject matter.
In the appeal to the Federal Circuit, the court applied the two step Alice inquiry to affirm the Board's decision. The court found that the claims were directed to abstract "mathematical calculations and statistical modeling" that are patent ineligible.
In both circuit decisions, Stanford argued the claims are not directed to an abstract idea because the claims improved predictive accuracy. The court rejected this argument.
In Stanford I, Stanford argued that main claim 1 is not directed to an abstract idea because claim 1 "ascertains more haplotype information than was previously possible." "While the `trio' method may be able to provide long-range haplotype phasing for approximately 80% of heterozygous positions, the method of the present invention provides accurate, long-range phasing at 97.9% of all heterozygous positions." The court acknowledged the importance and benefit of the improved predictive accuracy of the method. Nevertheless, the court held that the improvement amounted to no more than improving the output of an abstract mathematical calculation, which is insufficient by itself for patent eligibility. "Even accepting the argument that the claimed process results in improved data, we are not persuaded that claim 1 is not directed to an abstract mathematical calculation." (emphasis added)
In Stanford II, Stanford argued that the "increase in haplotype prediction accuracy renders claim 1 a practical application rather than an abstract idea." Again, in effect, the court found improving the output of an abstract mathematical calculation is insufficient by itself for patent eligibility. "Stanford's cited cases do not support its argument because the cases involve practical, technological improvements extending beyond improving the accuracy of a mathematically calculated statistical prediction." (emphasis added) "Unlike the technological improvements made in those cases, the improvement in computational accuracy alleged here does not qualify as an improvement to a technological process; rather, it is merely an enhancement to the abstract mathematical calculation of haplotype phase itself."
In step two of the Alice inquiry, the court found that the claims fail to include additional limitations that provide an inventive concept that transforms the abstract idea into patent eligible subject matter. With respect to improved predictive accuracy, Stanford argued in Stanford II that the Board "fail[ed] to consider all the elements of claim 1 as an ordered combination." "[I]t is the specific combination of steps recited in claim 1 … that provides the increased accuracy over other methods."
The court rejected this argument for several reasons. One reason is "That a specific or different combination of mathematical steps yields more accurate haplotype predictions than previously achievable under the prior art is not enough to transform the abstract idea in claim 1 into a patent eligible application."
In Stanford II, Stanford made an argument that takes advantage of an important factor supporting patent eligibility, which is improved computer performance. Specifically, Stanford argued "that one claimed advance is greater efficiency in computing haplotype phase." The court acknowledged the pertinence of this point, highlighting the relevance of "whether the claimed advance alleged in the written description demonstrates an improvement of a technological process or merely enhances an ineligible concept."
However, the court stated it was prevented from considering this argument. "Stanford has forfeited its argument that greater computational efficiency renders claim 1 patent eligible by failing to raise it before the Board."
Generally, the United States Patent and Trademark Office and the courts find claims to be patent eligible when the claims are directed to ML that is sufficiently applied to another technological area beyond the computer implementation of the ML. Claims to generally-applicable ML, which are generally only directed to the computer implementation of ML, may need to rely on another factor for patent eligibility.
Stanford I and II make clear that improved predictive accuracy is not a factor that supports patent eligibility by itself. It is important to note that the predictive accuracy Stanford argued in support of patent eligibility was not a generalized improvement in predictive accuracy but was instead a very specific improvement limited to the domain of haplotype phasing for "heterozygous positions." Innovations in generally-applicable ML yield improvements in predictive accuracy that are broadly applicable. Stanford I and II show that a strategy that limits improved predictive accuracy of a generally-applicable ML innovation to a specific domain in the claims is a strategy that may not prevent the claims from being deemed to be directed to an abstract idea or mathematical calculation.
The Federal Circuit has held that a specification's description of improved computer performance that is tied to a claim demonstrates that the claim is not directed to an abstract idea and is patent eligible. Enfish, LLC v. Microsoft Corp., 822 F.3d 1327 (Fed. Cir. 2016) "Our conclusion that the claims are directed to an improvement of an existing technology is bolstered by the specification's teachings that the claimed invention achieves other benefits over conventional databases, such as increased flexibility, faster search times."
Innovations for generally-applicable ML generally provide improved computer performance. Application specifications directed to generally-applicable ML should explain how improved computer performance is directly achieved. A general assertion of improved computer performance may not be sufficient. An application specification should explain, in detail, how steps and operations pertinent to key claim limitations constitute a computational method that runs faster or stores less data compared to prior methods that achieve similar purposes, thereby achieving improved usage for specific computer resources.
Also very helpful are features that exploit hardware or architectural aspects of computer systems. An example of this kind of feature is improved amenability for parallel execution by separate threads or by separate computers.
Generally-applicable ML often involves combinations of known ML techniques. In such cases, an application specification should explain in detail how the combination yields improved computer performance.
It should be noted improved computer performance may not help in jurisdictions other than the United States. For example, in the European Patent Office, unless a claim directed to ML satisfies one of two safe harbors, the claim is treated as a patent ineligible mathematical method. Improved computer performance may not render a mathematical method patent eligible in the European Patent Office.
Disclaimer: This article is purely a public resource of general information that is intended, but not guaranteed, to be correct and complete. It is not intended to be a source of solicitation or legal advice and is for informational and entertainment purposes only. The information is not intended to create, and receipt does not constitute, an attorney-client relationship. The laws of different jurisdictions may be implicated, and facts and circumstances can vary widely. Therefore, the reader should not rely or act upon any information in this article, but should instead seek legal counsel for individualized legal advice. For more information, please contact a firm attorney through www.hickmanbecker.com.