On September 26, PTAB's Appeal Review Panel (ARP) issued a key decision addressing patent eligibility for patent application requests filed by Google's Deepmind Technologies, directed at how ML models were trained. This decision eliminates the preliminary basis for rejection under 35 USC §101 and confirms that when the claim in question is considered in whole, it is not directed towards patent-essential abstract ideas, but instead reflects technical improvements in the field of machine learning.
The claimed invention
The patent application at the heart of this decision concerns computer-implemented methods for training ML models on multiple tasks. This method determines the importance of model parameters for the first task, and trains the model on a second different task, while securing performance in the first task. This is achieved by adjusting the model parameters to optimize the objective function incorporating penalty terms based on previously determined importance measures.
This specification highlights several technical advantages, including reduced storage requirements, reduced system complexity, and capabilities, including the ability to maintain performance across sequential tasks that address the well-known challenges of “catastrophic forgetting” in continuous learning systems.
Procedure history
The Board initially confirmed the rejection of all pending claims under 35 USC §103, introduced new grounds for rejection under §101, and found claims directed towards abstract ideas, particularly mathematical calculations. The applicant called for rehearsal, claiming it provided technical improvements to ML technology.
Legal Framework
ARP applied an established two-stage framework Alice Corp.v. CLSBank Furthermore, it is clarified in manual §2106 of the procedure manual. The analysis first considers whether the claim is directed towards a judicial exception (such as an abstract idea), and, if so, whether additional factors integrate the exception into a practical application.
Important findings
- ARP agreed that this claim recites an abstract idea in the form of mathematical calculations. However, the analysis did not end there.
- Once reviewed, ARP found that, when the claims are considered as a whole, it integrates abstract ideas into practical applications by improving the behavior of the ML model itself.
- The decision highlighted that the claims addressed technical challenges in continuous learning, such as maintaining knowledge from previous tasks and reducing storage needs that constitute improvements to computer technology.
- This decision highlights that claims directed towards ML methods may be eligible for patents when providing specific technical improvements rather than simply reciting abstract mathematical concepts or general computer implementations.
- The decision also highlights the continued relevance of §§102, 103, and 112 as the main statutory tools for assessing the scope and validity of claims, rather than relying on §101 to categorically exclude AI and ML innovations.
meaning
This decision addresses the legal framework for patent eligibility for AI inventions, but its broader meaning lies in a clear policy direction set by USPTO's leadership on patentability for AI and ML inventions. With this decision, USPTO demonstrates strong support for protecting AI innovation through its patented system.
This approach encourages inventors and companies to pursue patent protection for AI and ML advances, and encourages increased investment and growth in these areas. The decision reflects our commitment to ensuring that US patent policies maintain American leadership in AI by responding to technological development, providing robust incentives for innovation, and reducing uncertainty regarding eligibility for AI-related inventions. As a result, the USPTO could continue to increase AI patent applications and strengthen the state's position at the forefront of emerging technologies.
