Ex parte Desjardins

Ex parte Desjardins, decided September 26th, 2025. An Appeals Review Panel of the Patent Trial and Appeal Board, in an opinion by the Director of the Patent and Trademark Office, vacated a new ground of rejection the Board had entered under Section 101. The Panel held that a claim to training a machine-learning model, though it recites a mathematical calculation, is not directed to an abstract idea, because it integrates that idea into a practical application. It's a precedential decision. Here's the brief.

The application, assigned to DeepMind Technologies, claims a method of training a machine-learning model. The problem it targets is a familiar one in continual learning: a model retrained on a new task tends to overwrite what it learned before — what the applicant called catastrophic forgetting. The claimed method attacks that problem. After the model learns a first task, it measures how important each parameter is to that task, then trains the model on a second task while adjusting the parameters to protect performance on the first — an objective function with a penalty term keyed to those importance measures. The specification describes the payoff: learning new tasks in succession while protecting knowledge of earlier ones, using less storage, with reduced system complexity.

In March 2025, a Board panel affirmed the examiner's obviousness rejection of every claim under Section 103 — and, on its own, entered a new ground of rejection under Section 101. The Board reasoned that the claims' core — computing an approximation of a posterior distribution — was a mathematical concept, and that what remained was generic computer components. The applicant sought rehearing, directed at the Section 101 ground. The Board denied it.

The Director then convened this Appeals Review Panel to review the Board's work, with particular focus on the new Section 101 ground. The Panel took jurisdiction under Section 6(b) of the Patent Act — the provision that lets the Office review its own Board's decisions.

The eligibility question runs through the Alice framework, which the Office implements in the MPEP as Step 2A. The Panel confined its analysis to that step.

On Prong One, the Panel accepted the Board's finding that claim 1 recites a mathematical calculation — computing an approximation of a posterior distribution — and so recites an abstract idea. The applicant had not disputed that. So the Panel moved to Prong Two.

There, it parted from the Board. The claims, the Panel found, reflect a technical improvement to the model itself, not merely a mathematical result. It anchored the point in Enfish.

The Panel drew the line carefully. A specification's say-so is not enough on its own, it acknowledged — under the Office's own guidance and the Federal Circuit's Symantec decision, the claim itself has to reflect the disclosed improvement. Here, it found, the claim did. The specification identified the improvement — learning new tasks while protecting knowledge of prior ones — and the Panel located it in the claim, in the limitation adjusting parameters to optimize the second task while protecting performance on the first. That, the Panel said, is an improvement to how the model operates, not the abstract calculation. Contrast the cases the Board leaned toward, where a claim merely links an idea to a generic technological environment.

The Panel was pointed about the Board's approach — that examiners and panels should not evaluate claims at such a high level of generality, and should not brush past settled precedent like Enfish when acting on their own.

Desjardins is a precedential Office decision, so it binds examiners and the Board. Authored by the Director and issued through the Appeals Review Panel, it sets the frame for how the Office examines machine-learning claims going forward. Two limits keep it in perspective. The Panel vacated only the Section 101 ground; the claims still stand rejected under Section 103, so the applicant did not walk away with a patent. And the Panel located the real screens for patent scope elsewhere, describing Sections 102, 103, and 112 as the traditional and appropriate tools. What the decision does is direct the eligibility inquiry, for a machine-learning claim, to whether the claim reflects a concrete technical improvement.

Ex parte Desjardins, Appeal No. 2024-000567, decided September 26th, 2025. I'm John Goodhue. Thanks for watching.