Authored by: Jeremy J. Gustrowsky
A recent decision by the United States Patent and Trademark Office (USPTO) Appeals Review Panel has important implications for artificial intelligence (AI) and machine learning patents. The case involved U.S. Patent Application No. 16/319,040, which describes a method for training machine learning models so they can learn new tasks without forgetting previously learned information—a challenge known as “catastrophic forgetting.” The Board had previously rejected the claims under 35 U.S.C. § 101, arguing that the invention was directed to an abstract idea, but the Appeals Review Panel (ARP) has now vacated that rejection. For background, the USPTO retired the Precedential Opinion Panel process and established the ARP, and the Director Review process. The ARP may be convened at the Director’s discretion to review decisions of the Patent Trial and Appeal Board (PTAB or Board) in ex parte appeals, re-examination appeals, and reissue appeals.
The key issue in this review was whether the claimed invention was simply an abstract mathematical concept or if it provided a real, practical improvement to computer technology. The claims involved calculating the importance of each parameter in a machine learning model and then using that information to help the model learn new tasks while retaining its performance on old ones. The Board initially found that the claims recited mathematical calculations, which are generally considered abstract ideas and not eligible for patent protection.
However, the applicant argued—and the Panel agreed—that the invention went beyond just math. The Panel noted that the claims were not merely about performing calculations, but about improving how machine learning models operate. Specifically, the invention helps AI systems use less storage and maintain performance across multiple tasks, which are technical improvements in the field of computer science. The Panel emphasized that software innovations can be patent-eligible if they improve the functioning of computers or other technology, referencing well-established precedent.
The panel referred to the Federal Circuit’s decision in Enfish, LLC v. Microsoft Corp. as the leading precedent on the eligibility of technological improvements under Section 101. In particular, the panel highlighted that Enfish recognized software can make non-abstract improvements to computer technology, just as hardware improvements can, and that the eligibility determination should focus on whether the claims are directed to an improvement to computer functionality rather than an abstract idea. The panel also cited McRO, Inc. v. Bandai Namco Games America Inc. as another example of a case where claims were found patent-eligible because they improved the functioning of a computer or other technology. These cases were referenced to support the conclusion that the claims at issue in the present claims reflected a technical improvement to how machine learning models operate, and thus were not merely directed to an abstract idea.
The Panel criticized the earlier Board decision for being too broad and for equating all machine learning innovations with unpatentable algorithms. They stressed that examiners should not dismiss AI inventions as abstract ideas without carefully considering whether the claims actually improve technology. The Panel also pointed out that issues of novelty and obviousness (under §§ 102 and 103) are the proper tools for limiting patent protection, rather than relying too heavily on § 101.
In the end, the Panel vacated the § 101 rejection, finding that the claims integrated the abstract idea into a practical application and thus were not directed to an abstract idea. However, the application still faces challenges under § 103 for obviousness. This decision is a positive sign for inventors in the AI space, as it offers another example of the kind of technical improvements in machine learning that have a good chance of being patentable.