In recent years, the European Patent Office (EPO) has made it eminently clear that European patents may be granted for machine learning inventions. Given the way in which such inventions are trained based on a training dataset, it can often be difficult to know exactly what algorithm the system will perform once trained up. This may, however, appear to conflict with one of the requirements for grant of a European patent application, namely that the application must “disclose the invention in a manner sufficiently clear and complete for it to be carried out by a person skilled in the art”. Recent decision T0161/18 has provided some useful insight into this requirement for machine learning patent applications.
Many applicants attempt to provide a sufficiently enabled disclosure by setting out details of how the machine learning algorithm is trained. However, in T0161/18 the EPO Board of Appeal concluded that the application’s lack of detail on the training dataset used in the invention resulted in a finding of insufficiency.
The application in question related to a system comprising an artificial neural network used to transform a peripheral blood pressure curve into an equivalent aortic pressure. The architecture of the artificial neural network was pre-defined and its weights were learned via a supervised training process using a backpropagation algorithm. The application emphasises the importance of training on datasets covering “the range of possible properties” to enable generalization of the trained network. Specifically, the application references datasets comprising “measurement data from patients of different ages, genders, constitutional types, health status and the like”. However, the Board of Appeal concluded that this description of the training data was not sufficiently clear and complete.
In reasons 2.2, the Board of Appeal stated “the application does not disclose which input data are suitable for training the artificial neural network according to the invention, or at least one data set suitable for solving the present technical problem. The training of the artificial neural network can therefore not be reworked by the person skilled in the art and the person skilled in the art therefore cannot carry out the invention” [English translation].
Thus, while the application as filed clearly identified the use of a machine learning algorithm and provided information on the input data used in training, an insufficiency conclusion was still reached due to a lack of detail about that input data.
The Board of Appeal also considered inventive step of the claimed invention in their decision. Their conclusions in this regard were closely related to their conclusions on insufficiency – i.e. that the lack of detail on the training of the artificial neural network meant that the neural network cannot be considered adapted to the specific, claimed application.
It therefore appears that the Board of Appeal in T0161/18 considered the lack of specific detail on training datasets to be a deficiency in the application for both sufficiency and inventive step, leading to a dismissal of the appeal against the refusal of the application.
To minimise the risk of such objections on applications employing machine learning, it is therefore advised that as complete a description as possible of the algorithm’s training process, and the data on which training is based, be included in the specification (even if what is actually claimed might be broader). Such a description should ideally include a full description of the specific form and function of any inputs and outputs of the algorithm, and preferably alongside clear and fully explained examples of training datasets.