9.2.12 Assessment of features relating to mathematical algorithms
In T 1903/20 the invention related to neural network based machine translation. The board held that the translation of text from a source language to a target language was a matter of linguistics and not a technical effect. This was so even if the computer program included algorithmic aspects which were not directly based on linguistic concepts. It stated that merely finding a computer algorithm to implement an automated translation process did not render the resulting computer program technical (see also decisions T 598/14, T 2825/19 and T 2401/22).
In T 183/21 the board came to the conclusion that a technical effect was achieved by the subject-matter of a claim defining a method of automatically controlling the performance of a recommender system in a communications system, the communications system including a client device associated with a user to which the recommendations were provided, on average, over substantially the whole scope of the claim.
In T 702/20, the object of the invention was to reduce the number of connections between the nodes of a neural network (“loose coupling”). Claim 1 differed from the closest prior art in that the different layers of the neural network were connected in accordance with an error-code check matrix. The proposed network structure only defined a class of mathematical functions, which, as such, was excluded matter. Other "non-technical" matter, could only be considered for the assessment of inventive step when used to solve a technical problem. The appellant had argued that the claimed neural network solved a technical problem by providing effects within the computer related to the implementation of neural networks (storage requirements), and that neural networks generally solved technical problems by automating human tasks. The board remarked that a technical problem may also be solved if the outputs of the system have an implied further technical use (G 1/19), but held that the outputs of the neural network did not have any implied "further technical use"; they may e.g. be related to forecasting stock market evolution.
The appellant had argued that the proposed modification in the neural network structure, in comparison with standard fully-connected networks, would reduce the amount of resources required, in particular storage, and that this should be recognized as a technical effect, following G 1/19. The board however held that, while the storage and computational requirements were indeed reduced, in comparison with the fully-connected network, this meant that the modified network was different and would not learn in the same way. So it required less storage, but did not do the same thing.
The appellant had also argued that the neural networks were an automation tool, mimicking the human brain and that their behaviour could not be predicted or understood by their programmer. However, the board saw no evidence that neural networks functioned like a human brain. Neural networks are a mathematical approximation function, which can be simple and understandable if the network is small. It is only the sheer complexity of a larger neural network that makes it appear unpredictable. That a learning system is complex is not sufficient to conclude that it replicates the functioning of a brain.
The board stressed that there could be no reasonable doubt that neural networks could provide technical tools useful for automating human tasks or solving technical problems. In most cases, however, this required them to be sufficiently specified, in particular as regards the training data and the technical task addressed. What specificity was required would regularly depend on the problem being considered, as it had to be established that the trained neural network solved a technical problem in the claimed generality (see also T 2246/18 and T 161/18 where the boards found the use of neural networks obvious and T 748/19, as well as T 1191/19, where the claims were also too generalised). The board in T 1952/21 referred to T 702/20 and stated that it was in many ways similar. It reiterated that a trained machine learning model, namely a neural network, can only be considered for the assessment of inventive step when used to solve a technical problem, e.g. when trained with specific data for a specific technical task.
In T 1191/19 a so-called meta-learning scheme was the subject of the invention. The board held that this constituted the application of known techniques - such as meta-learning - to other areas, which was a general technological trend and could not be inventive as such. The board could not see in the method of claim 1 any non-obvious detail of the application of the meta-learning scheme known in the prior art to the problem at hand beyond a mere reiteration at an abstract level of the scheme disclosed in the prior art.
The board decided in a similar vein in T 1425/21, where approximating "cumbersome" machine learning models with "distilled" machine learning models which require less computation and/or memory was claimed. The board held that while the distilled model reduced the storage or computational requirements of a machine learning model, that was insufficient, by itself, to establish a technical effect, as one also had to consider the performance of the “reduced” learning model. The board found that it was not credible in general that any model with fewer parameters could be as accurate as the more complex one it is meant to replace.
In T 598/07 the invention concerned a heartbeat monitoring method, which was based on a neural network for the purpose of identifying irregular heartbeats. The board held that this made a technical contribution. The board also held, that since none of the method claims in question included the step relating to the diagnosis for curative purposes stricto sensu representing the deductive medical or veterinary decision phase, the method claims did not fall under the exclusion provisions of Art. 53(c) EPC.
In T 1286/09 the invention related generally to the field of digital image processing and, in particular, to a method for improving image classification by training a semantic classifier with a set of exemplar colour images, which represented "recomposed versions" of an exemplar image, in order to increase the diversity of training exemplars. The board found that it involved an inventive step.
In T 1510/10 the board decided that 'no inventive step can derive just from the use of machine learning. The Appeal was dismissed.
In T 1784/06 the automatic classifying of abstract data records was held to be non-technical since the data records were classified for the non-technical purpose of billing. A valuable mathematical property of the algorithm could imply technical benefits but only when used for a technical purpose.
In T 755/18 the board held that if neither the output of a machine-learning computer program nor the output's accuracy contributed to a technical effect, an improvement of the machine achieved automatically through supervised learning to generate a more accurate output was not in itself a technical effect.
In T 761/20, the claim defined a method of automated script grading using machine learning, which was effectively a computer-implemented process. Such processes may have technical effects - and thus be deemed to solve a technical problem - at their input or output, but also by way of their execution (see G 1/19, point 85 of the Reasons). A technical effect may also be acknowledged in view of their purpose, i.e. an (implied) technical use of their output (see G 1/19, point 137 of the Reasons). The claimed method contained steps for extracting numerical "linguistic" vectors from scripts (for all con‑sidered samples, training scripts and scripts to be graded), a step of training a perceptron, and a step of using the perceptron to grade the scripts. In principle, the claimed training procedure might constitute a technical contribution to the state of the art (see e.g. G 1/19, point 33 of the Reasons). Taken alone, however, this was a mathematical method, so this contribution was in the excluded field of mathematical methods (see T 702/20 and T 755/18, Catchwords) and was therefore not a patentable contribution. Assuming that the claimed invention served the purpose of supporting its users in evaluating linguistic competences, as the appellant had argued, the board could not see any other implied purposes. The board also did not agree with the appellant that the field of “educational technology” was a technical field.
In T 874/19 the application related to classifying search engine resources as a spam resource (or as belonging to the "spam" category) or not a spam resource (or as belonging to the "not spam" category) based on a “deep network”. This involved different layers of a neural network for automatically classifying search input data. The claimed classifier was capable of processing an alternative mathematical representation of input data to then calculate a so-called “category metric” for each category in a given set of categories. Each category metric also provided a probability value that the particular resource actually belonged to the corresponding category. The board could not see what a further technical effect of the distinguishing features could be, nor what objective technical problem the subject-matter of claim 1 would solve.