Neural networks are obvious? There are of course many applications of neural networks or machine learning in general where the way neural networks are used is far from obvious but what has become clear from recent EPO case law is the less surprising notion that merely referencing a neural network in a claim without any details of structure or how the network is trained to solve a technical problem is not going to provide a basis for convincing the EPO of an inventive step.
In the recent decision T 2246/18, the application relates to a gas turbine engine with a “controller trimmed in response to a neural network”. This was found to be the distinguishing feature over the closest prior art and so the Board considered whether introducing trimming in response to a neural network requires an inventive step. The board found that using a neural network to trim the controller is obvious in light of a general teaching in a second document that neural networks are candidates for the control of non-linear systems, confirming the known advantages of neural networks. While the Board acknowledged that the second document did provide a direct teaching to use a neural network for the specific claimed task, the second document contained general teachings with respect to “improved and adaptive control of a complex non-linear process” and that was enough to render the use of a neural network per se obvious. While the Board did have a document to hand in this case, the reference by the Board to “known advantages” indicates that the Board would have reached the same conclusion considering just the skilled person’s common general knowledge.
Two more cases relating to the general use of neural networks underline the point made in T 2246/18.
In T 0161/18 the invention used “a neural network with weights determined by learning” to provide a precise and computationally light method of estimating heart time volume from peripheral pressure wave forms. The Board did not buy it. The reference to learning did not extend beyond what the skilled person would in any case understand by a neural network and since there was nothing claimed about how the learning or the neural network is adapted to achieve the alleged effect, the Board found that this effect does not apply across the whole scope of the claims and therefore cannot be considered in the context of an improvement over the state of the art. Therefore, the invention was to provide an alternative to the closest prior art method. Merely using a neural network did not require an inventive step because this is a general technological trend (besides also being known in the specific technical field of the invention in this case).
In T1968/08, the invention concerned the combination of a neural network and a neuro-fuzzy controller. The neural network and neuro-fuzzy controller were disclosed in different documents but since no specific technical effect could be shown for the claimed combination, the skilled person was looking to find an alternative solution for the controller and in these circumstances did not need a hint to combine the two documents. It was enough to establish that the documents could be combined, rather than that the skilled person would actually do so, as would otherwise be the case.
In summary, merely reciting the use of a neural network or the learning of weights will not convince the EPO to acknowledge an inventive step. Neural networks are used in line with technological trends and their generally known advantages without the need for a hint to do so in the cited documents. Of course, none of these precludes an inventive step to be acknowledged for a specifically adapted neural network or learning method that can be shown to provide a technical effect and hence solves a technical problem (see here and here for the circumstances in which a technical effect can be acknowledged for machine learning features).