6.1.1 Artificial neuronal network

In T 161/18 the application used an artificial neural network to transform the blood pressure curve measured on the periphery into the equivalent aortic pressure. As regards how the neural network according to the invention was trained, the application disclosed only that the input data should cover a wide range of patients differing in age, sex, constitution type, state of health, etc. to prevent the network from becoming specialised. However, it did not disclose what input data were suitable for training the network or even a suitable set of data for solving the technical problem in question. As a result, the skilled person could not reproduce the network's training and so could not carry out the invention. The invention, which was based on automated learning, in particular in connection with an artificial neural network, was thus insufficiently disclosed, because the training it involved could not be reproduced owing to a lack of disclosure in this regard. See also T 702/20 on the technical nature of a neural network implemented on a computer.

In T 1669/21 the appellant (patent proprietor) maintained that the invention was disclosed clearly and completely in the patent even though there was no specific embodiment example. The calculation model according to the claim was a machine-learning model. Owing to rapid progress made in the field, creating a suitable machine-learning calculation model was now part of the common general knowledge (by contrast with the situation in T 161/18), so no specific disclosure was required. The invention was aimed at developing a "comprehensive solution" for predicting the wear state of the refractory lining of a metallurgy vessel. In the board's view, the term "calculation model" was not limited to a machine-learning model, even in connection with an adaptation ("adapted"). Therefore, the model being "adapted" was not synonymous with the model being "adaptive", i.e. self-adapting or self-learning. Claim 1 thus was not limited to a machine-learning method. The patent did not provide an example of or pointers towards the relationships to be modelled for a calculation model of this kind. For this reason alone, the main request did not meet the requirements of Art. 83 EPC. There were many different options for configuring a calculation model. However, neither the patent nor the general technical knowledge presented contained any information on the specific requirements. As such, the skilled person was faced with a significant obstacle to reproducing the invention as early as when selecting a specific suitable machine-learning calculation model. And even if the skilled person were able to find a suitable model, they alone would have to select combinations from the multitude of possible input values. Each individual attempt by itself would represent a considerable burden. Absent any concrete workable embodiment example as a starting point in the patent and any specific guidance on which parameters were relevant, there was also no evidence that invention could essentially be reproduced using a calculation model trained in accordance with the claim. In this regard, the appellant maintained that this was unnecessary because it was precisely the essence of machine learning that the ability to predict the output variable without any knowledge of the causal relationships was acquired by self-learning through training; the impact of irrelevant input variables was filtered out automatically in the process. The board held that the patent specification did not contain any specific workable embodiment example or any pointers or criteria for selecting suitable concrete parameters within the claimed categories. As for the volume and quality of the training data (see also T 161/18), the calculation model had only been trained on a relatively small dataset. The contested patent did not mention the presented procedure for carrying out the invention or disclose how the training data were acquired. The appellant was unable to plausibly show that the calculation model could be successfully trained using such a limited set of training data. Thus, the disclosure of the patent was vague and incomplete even in relation to the decisive aspect for the success of the invention – the training data. The lack of detail in the disclosure of the patent was disproportionate to the breadth of the claimed invention and the resulting burden placed on the skilled person to fill in the gaps in order to carry out the invention over the whole scope. G 1/03 was invoked but unsuccessfully.

In ex parte case T 1539/20 (method of monitoring performance of an application system distributed across a plurality of network connected nodes), claim 1 was directed to a computer-implemented method comprising an automated step of "mapping" (concerning which term, Art. 84 EPC issues were raised) a given distributed application system to a hierarchical model, however the application as filed did not contain any information explaining how a skilled person could implement the "mapping" process in software. The appellant also submitted that the presence of non-working embodiments in the claim was of no harm, provided that the specification contained sufficient information on the relevant criteria to identify the working embodiments. However the board's objection to claim 1 was not that its scope encompassed specific non-working embodiments.

In ex parte case T 1191/19, concerning Art. 56 EPC, the board recalled that the mere application of a known machine learning technique to problems in a particular field was a general trend in technology (T 161/18, point 3.6 of the Reasons) and could not be inventive as such.  Concerning specifically Art. 83 EPC, the board noted that the application did not disclose any example set of training data and validation data, which the meta-learning scheme required as input. The application did not even disclose the minimum number of patients from which training data should be compiled to be able to give a meaningful prediction and a set of relevant parameters. In particular, the structure of the artificial neural networks used as classifiers, their topology, activation functions, end conditions or learning mechanism were not disclosed (citing T 161/18, point 2 of the Reasons). At the level of abstraction of the application available, disclosure was more like an invitation to a research programme. Under these circumstances, the skilled person could not reproduce without undue burden the application of the meta-learning scheme of AX1 (scientific publication) to solve the problem of predicting personalised interventions for a patient in processes, the substrate of which was neuronal plasticity.

In ex parte case T 1526/20 the application related to a computer-implemented "liveness testing method" for distinguishing between live faces and impersonations based on 2D images. The board found that the skilled person had to have had doubts that the model provided the information necessary for reliable liveness testing. The application provided no reasons as to why this model was correct, nor results to show that the proposed way of extracting object-related information correctly distinguished live from fake objects.

In ex parte case T 606/21 (method for evaluating predictions of trajectories by autonomous vehicles, including generating a Deep Neural Network, or DNN), the examining division considered that the application did not clearly disclose how the DNN model could provide a reliable output at the time of testing because it was fed with an incomplete input. The input at the time of testing only included the predicted trajectory, not the actual trajectory. The board considered the method comprised two main states, the first stage related to the generation of a deep neural network, DNN, and the training of this DNN by inputting pairs of trajectories in order to analyse the similarity of these trajectories and to improve the accuracy of this analysis. The board considered that there was a lack of disclosure regarding the second stage, the testing and evaluation stage.

In ex parte case T 509/18 the invention related to a driver alertness detection system configured to use a classification training process to register the driver's head position and eye vector at several pre-determined points within the vehicle, configured to save a matrix of inter-point metrics to be used for a look-up-table classification of the driver's attention state. In the board’s view, the application did not teach how to derive from the "matrix of inter-point metrics" a "look-up-table classification of the driver’s attention state", with such a "look-up-table classification" permitting to decide on the driver's attention state. The skilled person would not know how to construct a "look-up-table classification" and consequently how to decide on the driver's attention state based on the video camera's image of the actual position of driver’s head and eyes at a given instant. The step of comparing the video's camera image with this look-up table classification – in order to assess the driver’s attention state – required instructions and teaching concerning the kind of information to be extracted from a given video camera image and concerning the method and the criteria to be applied in order to compare; no such disclosure was to be found.