Training of the AI - EPO case law

The basis of any invention involving the application of machine learning and artificial intelligence is the training of the AI. Much discussed in this context are copyrights, see our blog post on ChatGPT, and data protection (GDPR).

With regard to patent law, further aspects become apparent. An invention can only be protected by a patent if it is technical. However, it is also essential that the technical teaching described in a patent application can be understood by a person skilled in the art and is sufficiently disclosed for this purpose.
This is a challenge in practice, as current case law shows.

EPO case-law on AI training

The European Patent Office has already ruled several times on inventions that use AI.

2020: Equivalent aortic pressure/ARC SEIBERSDORF

In the EPO decision of 2020 (T 0161/18, Equivalent Aortic Pressure/ARC SEIBERSDORF), a board of appeal of the EPO dealt with the training data used and the training of an artificial neural network according to the invention. The question under patent law was whether there was sufficient disclosure and inventive step.

Disclosure in particular is a challenge with AI-based inventions: How can the training according to the invention be reproduced in a self-learning AI?

Claim 1 of the patent in suit stated that the task (the precise determination of the cardiac output) was solved with the aid of an artificial neural network whose weighting values were determined by learning.

In the EPA's view, however, this was too general a description. The use of artificial neural networks was in line with a general trend in technology, including medical technology, the EPO explained.

The neural network was not adapted for the specific application claimed. Therefore, because of the lack of practicability, the EPO did not consider the patent invention to be sufficiently disclosed, nor did the use of the artificial neural network constitute an inventive step.

2022: Sparsely connected neural network/MITSUBISHI

Similarly, a board of appeal of the EPO ruled in November 2022 in T 0702/20 (Sparsely connected neural network/MITSUBISHI). A neural network is to be seen as a class of mathematical functions which as such were excluded from patent protection (under Article 52(2) EPC). At its core, a neural network was a mathematical approximation function, according to board of appeal.

It explained that, like other "non-technical" subject-matter, it could therefore only be taken into account in assessing inventive step if it was used to solve a technical problem, e.g. if it was trained with specific data for a particular technical task.

Mitsubishi did not specify a particular task for the neural network in the claims. However, it had to be shown that the trained neural network solves a technical problem, the EPO said. While the patent application mentions data-driven sparsity, the EPO said that given the content of the application, it was not possible to see for what kind of learning tasks the proposed structure could be of use and to what extent.

Although the operation of a neural network may not be predictable before training and the programmer may not understand the meaning of the individual parameters, the neural network nevertheless operates according to the programming of its structure and learning scheme, the EPO explained.
The patent application for this invention was therefore rejected for lack of inventive step.

2022: Neuronal plasticity - meta-learning scheme

In another case, also an EPO decision from 2022, the learning scheme for training the AI was mentioned in the patent application - it was a so-called meta-learning scheme.
Is meta-learning general knowledge in the field of AI? This question was the focus of this EPO decision (T 1191/19). Read more about this case in our blog post AI invention with meta-learning scheme.

The training for an invention with AI application must be well structured in the patent application. This requires both experience and strategic argumentation. Our patent law firm Köllner & Partner offers this with much expertise.

Please contact us for more information by phone at +49 69 69 59 60-0 or info@kollner.eu.