A sparsely connected neural network is a class of mathematical functions until it is tied to a technical task

2022-11-07EPO

Decision T 702/20 (7 November 2022), Board 3.5.06, Sparsely connected neural network / Mitsubishi Electric Corporation, European patent application No. 14882049.1

Mitsubishi Electric Corporation's European application 14882049.1 claimed a hierarchical neural network apparatus implemented on a computer in which the connections between layers are made sparse according to the parity-check matrix of an error-correcting code, with a weight learning unit and a discriminating processor for classification or regression. Technical Board of Appeal 3.5.06 dismissed the appeal from the refusal for lack of inventive step under Article 56 of the European Patent Convention. The catchword states that "A neural network defines a class of mathematical functions which, as such, is excluded matter" and "can therefore only be considered for the assessment of inventive step when used to solve a technical problem." The computer implementation clears the first hurdle, but the claimed structure "only defines a class of mathematical functions," and the argument from Enlarged Board decision G 1/19 that reduced storage and computation inside the computer is a technical effect failed because the sparse network "does not do the same thing" as a fully connected one, so the comparison was incomplete, and the implementation "does not require any adaptation of the computer." The Board was not persuaded that neural networks in general function like a human brain, found that the claims specify no particular task, and remarked that neural networks can be useful technical tools but generally need to be sufficiently specified, particularly as to training data and technical task.

What it changesFor a European filing, an improvement to the architecture of a network, sparsity, an attention pattern, quantization or a new layer type, is not an invention by itself; it becomes one when the claim says what technical task the network performs and the specification shows the architecture does that task better, or when the implementation adapts the hardware in a way the Board can see. The drafting consequence is concrete: claim 1 for Europe should recite the input data, the technical output and its use, and the training data, and the specification should carry a comparison against the unmodified network on the same task, because the Board rejected a resource argument precisely for lack of a like-for-like comparison. A specification written for the United States that describes the architecture generally and lists applications at the end will not supply this. Decide before filing whether the architecture or the application is the invention, and if it is the architecture, be prepared to narrow to the application in Europe.

More on Eligibility of software and AI claims

← All recent developments · Search all developments

Burbage Law is the practice of Jon Burbage, a registered patent attorney. 342 United States patents have issued from applications he wrote; the full list is on the Patents page. More about Jon Burbage.

This is general information about a published decision, agency guidance or piece of legislation. It is not legal advice and does not create an attorney-client relationship. Items concerning offices outside the United States are reported for information only and are not counsel on the law of those countries.