T 161/18: The sufficiency of description put to the test by artificial intelligence
In decision T 161/18 of 12 May 2020, Technical Board of Appeal 3.5.05 of the EPO upheld the refusal of an application relating to a neural network, considering the description to be insufficient in the absence of details on the data used, such that, moreover, the neural network did not achieve any technical effect contributing to the inventive activity. Although this position sheds further light on the interpretation of the sufficiency requirement as applied to artificial intelligence, its scope should nevertheless not be overstated..
The application specifically concerned a method for evaluating cardiac output from blood pressure based on an artificial neural network, the weights of which were determined through learning. The examination division rejected the application for lack of inventive step. The board of appeal upheld this rejection, though with slightly different reasoning. It first observed that the application provided few indications regarding the input data used to train the neural network, merely stating that said data should cover a wide range of patients. However, according to the board, without knowing what data was suitable for training the network, the person skilled in the art could not carry out the invention, meaning the disclosure was insufficient.Article 83 EPC. Furthermore, as the person skilled in the art would not have been able to implement this invention, the claimed neural network could not have produced a technical effect that contributes to inventive activity (Article 56 of the European Patent Convention).
A neural network, which is a form of artificial intelligence, allows a machine to evolve through a systematic process and to perform tasks for which it has not been programmed by learning from data. Two parameters are therefore fundamental in such a network: the data selected at the input and then the correlation algorithm. It is difficult to determine, in fact, the qualitative importance of each of these aspects given their complementarity within the learning process. It is, in part, this complementarity that leads the chamber in this case to consider that the mere disclosure of the algorithm is insufficient for a person skilled in the art to carry out the invention.
However, while the reasoning followed by the chamber elucidates the process with neural networks, the solution arrived at should not, in our opinion, require applicants to systematically provide input data. Let us recall that the description must provide intellectual access to the invention, without necessarily guaranteeing industrial access. This means that it is not a question of providing details of the invention's implementation, but only of allowing a person skilled in the art to understand the invention, such that some trial and error may be necessary to achieve execution and reproduction of the teaching. Regarding computer-implemented inventions, the category to which neural networks belong, the question of disclosure has always been sensitive. In the past, for instance, there has been debate about the necessity of providing source code lines. In order, However, the underlying problem remains the same whether providing source code or input data: determining whether the invention exists or if the applicant is encroaching upon the domain of science, which is too abstract and must be distinguished from the domain of technology reserved for patent law. Thus, as in this case, we return to the concern for the technical character of the invention: is the claimed invention sufficiently concrete to be considered to result in a technical effect, or is it a matter of claiming a right over a mathematical formula that falls more within the realm of science?
In this specific case, the chamber considered that the data were, in this precise instance, necessary to achieve the technical effect. In fact, in the presence of a machine learning system, inventive activity rests above all on the ability to train a machine to achieve a given result. It is therefore necessary to explain in the application how the neural network uses the data to arrive at the result. The mere provision of this data does not appear, in itself, indispensable, as long as it is sufficiently specified how the neural network is trained. This ultimately leads to the same conclusion as for source code: the description must explain to a person skilled in the art how to arrive at the invention, it must not provide the invention, such that a programmer can be asked to write code or train a machine without giving the code or data, or by only giving snippets. The neural network in question in this case was described in a very general way, without any precision on the characteristics allowing its interaction with the other claimed characteristics: no indication on the structure of the input data or on the output. It is therefore not surprising that this network was deemed unusable in view of the elements provided.
Therefore, one should not deduce from the decision T 161/18 that the provision of data constitutes a requirement sine qua nonto fulfil the condition of sufficiency of the statement. On the other hand, it underlines, once again, the importance of description in the field of computing, while providing details to the writers of applications relating to artificial intelligence.