Issues concerning the grant of machine learning patents at the EPO
How much information should you include in your machine learning (ML) patent application to give it the best chance of being granted by the European Patent Office (EPO)?
As explained in our previous article (see link below), the answer to this question relies on a careful balance between disclosing enough information to meet EPO requirements, and not disclosing so much that competitors may gain commercially valuable information when the patent application is published.
"Practical considerations for patenting AI", 20 October 2020, Robbie Berryman.Read more
In this article, we discuss two issues to consider for obtaining ML patents.
Article 83 of the European Patent Convention (EPC) requires that: “The European patent application shall disclose the invention in a manner sufficiently clear and complete for it to be carried out by a person skilled in the art.” For ML patents, the requirements of Article 83 are particularly important when it comes to training data.
For example, in T 0161/18, the EPO Board of Appeal found that a patent application does not meet the requirements of Article 83 if: “… the application does not disclose which input data are suitable for training the artificial neural network of the invention, or at least one data set suitable for solving the present technical problem ...” (Reasons, 2.2).
Similarly, in T 1191/19, a patent application was deemed not to satisfy Article 83 because it did not disclose: “… any example set of training data … The application does not even disclose the minimum number of patients from which training data should be compiled to be able to give a meaningful prediction …” (Reasons, 4.1).
So, a description of the training data required to training a ML model is clearly important. In particular, the EPO appears to require disclosure of:
- which input data (for example, the minimum requirements of that data) would be suitable for training the ML model to solve the technical problem at hand; and/or
- at least one example set of training data suitable for training the ML model.
2. Inventive Step
In both of the above Board of Appeal decisions, the patent applications were also found to lack an inventive step. For a patent to be granted, the claimed invention needs to be both new and inventive – that is, “not obvious to a person skilled in the art” (Article 56 EPC). In applying this requirement, the two Boards of Appeal found that the mere application of machine learning to a known problem is not enough for an application to be considered inventive.
For example, in both T 0161/18 and T 1191/19, the Board of Appeal held that: “… the mere application of a known machine learning technique to problems in a particular field is a general trend in technology … and cannot be inventive as such …” (T 1191/19, Reasons, 3.2).
So, the EPO is unlikely to consider a ML patent to be inventive unless there is something about the model or how the model is used which is specifically adapted to the specific use case of the model.
Note also that the EPO requires the claimed invention to provide some technical improvement, so applying ML to a non-technical problem (for example, in relation to a business method or similar) is unlikely to lead to a granted European patent (T 0755/18, Reasons, 3.2).
We can see, therefore, that the Board of Appeal at the EPO does not consider the requirements of Article 83 EPC to be met unless a patent application identifies which input data (for example, the minimum requirements of that data) would be suitable for training the ML model, and does not consider a ML application to be inventive unless there is something about the model or how the model is used which is specifically adapted to the specific use case of the model and provides a technical improvement.
This means that a claim to the mere idea of applying machine learning to a known problem (with no further detail on how the model is used or implemented) is unlikely to lead to a granted European patent.