A recent decision from the EPO's Boards of Appeal (T 0702/20) provides a practical example of where the boundary lies in terms of patent eligibility of AI algorithms at the EPO.
The decision relates to a patent application for a neural network with fewer node connections, to reduce the number of computations and avoid overfitting. As is often the case at the EPO, the question of patent eligibility was considered as part of the inventive step (obviousness) analysis. The Application was refused by the EPO's Examining Division as not solving a "technical" problem.
In the Appeal, the Applicant argued that the invention solved a technical problem by "providing effects within the computer related to the implementation of neural networks (storage requirements)". They also argued that "the proposed modification in the neural network structure, in comparison with standard fully connected networks, would reduce the amount of resources required, in particular storage, and that this should be recognized as a technical effect".
The Board of Appeal dismissed the Appeal, disagreeing for a number of reasons.
- The claimed neural network apparatus may have a new and non-obvious structure. However, the proposed network structure, only defines a class of mathematical functions, which without a technical implementation (a "further technical use") are always consider by the EPO as excluded matter.
- There is nothing hardware-specific about the invention that could support a technical effect, for example, no features that "emphasise the limited resources and therefore the relevance of a small network size" or "require any adaptation of the computer".
- While the storage and computational requirements are indeed reduced in comparison with the fully-connected network, the functionality will also be affected, so this does not in and by itself translate to a technical effect, but rather a circumvention of the resource constraint problem.
- The claimed learning and use of the network "to solve a classification problem or a regression problem" can use any data (i.e., including the possibility that the data is non-technical). Therefore, the outputs of the neural network do not have any implied "further technical use".
None of the Board's reasoning is surprising here, as it follows the well-publicised EPO Guidelines. Importantly though the Board also gave some advice on where the case went wrong, emphasizing that sufficiently specifying "the training data and the technical task addressed" is important for establishing patentability in Europe. Additionally, the Board suggested that to overcome technicality issues of this type in AI inventions, it may help if the Application demonstrates "which type of learning tasks the proposed structure may be of benefit, and to what extent".
This reinforces that inventions relating to machine learning methods and algorithms are patentable in Europe, but it is important that the patent application identifies at least one solution to a technical problem (rather than a generic AI method). This may allow the "further technical use" to be seen (whether explicit in the claims or implied) and avoid excluded subject matter problems. This settled approach at the EPO is welcome.
The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.