Authored by: Jeremy J. Gustrowsky
A recent decision from the Patent Trial and Appeal Board (PTAB) highlights the challenges of patenting advances in artificial intelligence and machine learning. In the appeal concerning U.S. Patent Application No. 16/319,040, filed by DeepMind Technologies Limited, the Board affirmed the examiner’s rejection of claims directed to a method and system for training a machine learning model on multiple tasks without forgetting previous tasks. The Board not only agreed with the examiner’s findings of obviousness but also introduced a new ground of rejection, finding the claims ineligible for patent protection under 35 U.S.C. § 101.
The invention at issue describes a system for sequentially training a machine learning model on different tasks, aiming to optimize performance on new tasks while preserving performance on earlier ones. This is achieved by assigning importance weights to model parameters based on their relevance to previous tasks and using these weights in a penalty term during subsequent training. The Board analyzed whether this approach was novel and non-obvious in light of several prior art references, including patents and academic papers on machine learning techniques.
DeepMind argued that the cited references did not teach or suggest key aspects of their invention, such as assigning parameter importance based on a posterior probability distribution. However, the Board found that the combination of prior art references, particularly the teachings of Gordon (US 2014/0101090 A1; published April 10, 2014) and Mehanna (US 2016/0092786 A1; published March 31, 2016), rendered the claims obvious. The Board also rejected DeepMind’s argument that the references were not analogous art, concluding that all the cited references were within the same field of endeavor—machine learning.
In a significant move, the Board issued a new ground of rejection under § 101, finding that the claims were directed to an abstract idea—specifically, mathematical algorithms for training machine learning models—and did not include any inventive concept that would transform them into patent-eligible subject matter. The Board noted that the claims merely recited generic computer components and did not improve the functioning of a computer or any other technology. As a result, the claims were found to be ineligible for patent protection.
Interestingly, one judge concurred in part, agreeing with the § 101 rejection but disagreeing with the finding of obviousness, arguing that the examiner had misinterpreted certain aspects of the prior art. Nevertheless, the overall outcome was that DeepMind’s patent application remains rejected, with the applicant given the option to amend the claims or request rehearing.