Authored by: Jeremy J. Gustrowsky
In a recent decision, the Patent Trial and Appeal Board (PTAB) denied a request for rehearing from DeepMind Technologies Limited regarding their U.S. Patent Application No. 16/319,040. The application, which focuses on improving machine learning models to handle multiple tasks without losing performance, faced rejections under both Section 101 (patent eligibility) and Section 103 (obviousness). DeepMind had argued that their invention offered technical improvements in continual learning, but the Board was not convinced.
DeepMind’s main argument centered on how their system enables a single machine learning model to be trained on multiple tasks sequentially, without needing to store separate models for each task. They claimed this approach reduces system complexity and storage needs, and specifically addresses the problem of “catastrophic forgetting”—a common challenge in continual learning where a model forgets previous tasks when trained on new ones. The Board, however, found that these features did not amount to a technological improvement, but rather described a mathematical concept typical of machine learning.
On the issue of obviousness, DeepMind challenged the Board’s interpretation of key terms in their claims, such as “parameter,” and argued that the cited prior art did not teach or suggest their claimed invention. The Board clarified that even if there was disagreement about the interpretation of certain terms, the prior art references—particularly one called Gordon—were sufficient to support the rejection. The Board also pointed out that DeepMind’s arguments were repetitive and did not identify any points that had been overlooked or misunderstood in the original decision.
Regarding patent eligibility under Section 101, the Board reaffirmed its view that the claims were directed to an abstract idea, specifically a mathematical calculation for training a machine learning model. The Board noted that simply applying an abstract idea to a particular technological environment, such as machine learning, does not make it patentable. They emphasized that the claims did not provide a specific technological solution or improvement to an existing process, but instead described a new use of a known concept.
In summary, the PTAB found that DeepMind’s arguments did not reveal any errors or oversights in the original decision. The Board maintained its rejections under both Section 101 and Section 103, and the request for rehearing was denied. This decision highlights the ongoing challenges in obtaining patent protection for innovations in artificial intelligence and machine learning, especially when claims are viewed as abstract ideas or obvious in light of existing technology.