The EPO has published its Guidelines to provide guidance on various proceedings before the EPO, such as for the search and examination of European patent applications and for the post-grant opposition of European patents. Helpfully, the Guidelines have a section dedicated to the EPO’s approach to inventions that involve machine learning or artificial intelligence. So, what do the Guidelines say?
The Guidelines use the terms “machine learning” and “artificial intelligence” interchangeably, drawing no distinction between the two.
The main point made in the Guidelines on machine learning algorithms is that the EPO considers such algorithms, when viewed in isolation as abstract processes or as a core technology, to be purely mathematical methods. This is irrespective of the type of algorithm involved, such as neural networks, genetic algorithms, support vector machines, k-means, kernel regression and discriminant analysis. This helps explain why there can often be difficulties in obtaining protection at the EPO for inventions directed purely at machine learning algorithms (i.e. when divorced from any specific implementation, application, data types or use-case), since purely mathematical methods are excluded from patentability under the European Patent Convention. The section in the Guidelines that sets out the EPO’s approach to mathematical methods sheds more light on this, and can therefore be helpful when looking for further insight on how the EPO views machine learning algorithms.
Generally, inventions that involve machine learning have supporting context (e.g. being configured to operate on specific data to achieve certain goals), and this can often move the invention away from being considered as merely a mathematical method. With such inventions, making it clear that the claimed invention is targeted at achieving a technical effect can often be key to success at the EPO. Usefully, the Guidelines provide numerous examples of what may be viewed as a technical effect.Examples of this include the use of a neural network in a heart monitoring apparatus for the purpose of identifying an irregular heartbeat; the classification of digital images, videos, audio or speech signals based on low-level features (e.g. edges or pixel attributes for images); or optimising load distribution in a computer network. Based on the Guidelines, such inventions are, in principle, patentable. Likewise, the Guidelines provide examples of what is often viewed as non-technical, such as classifying text documents solely in respect of their textual content (which is considered by the EPO as a non-technical linguistic problem). Of course, there are often shades of grey within this, and being able to initially cast your invention in the best light, to ensure that the EPO views the purpose and context of the invention as achieving a technical effect and therefore as potentially patentable, can be extremely helpful – these sections of the Guidelines can provide a useful steer on this.
There are, of course, aspects of machine learning algorithms other than just how the algorithms are used. For example, for a classification algorithm, a set of training data and a set of test data may need to be created and a suitable procedure for the training phase may need to be devised (e.g. to make the whole process practical, efficient and effective). The Guidelines make it clear that the creation of training data and test data and the specific training process are aspects of the invention that may be separately protectable. In particular, the Guidelines explicitly state that, where a classification method serves a technical purpose, the steps of generating the training set and training the classifier may also contribute to the technical character of the invention if they support achieving that technical purpose.
The Guidelines also make it clear that, even if a machine learning algorithm is not targeted at a specific technical application, then patentability can lie in how the algorithm is implemented, in particular if the implementation/design is motivated by technical considerations of the internal functioning of the computer system or network on which the algorithm will run. A specific example given in the Guidelines is that of “assigning the execution of data-intensive training steps of a machine-learning algorithm to a graphical processing unit (GPU) and preparatory steps to a standard central processing unit (CPU) to take advantage of the parallel architecture of the computing platform”. Thus, the Guidelines make it clear that aspects of the invention that are tailored to specific hardware/network considerations may themselves be patentable, even in the absence of a technical purpose for the algorithm per se.
A final point of EPO practice that is often of particular interest for machine learning algorithms also discussed in the Guidelines is that an algorithm may be more efficient in comparison to previous algorithms (e.g. in terms of resource usage, execution speed, etc.). However, the Guidelines make it clear that this, in itself, is not necessarily enough to ensure that the new algorithm is patentable at the EPO.In particular, addressing a non-technical problem (such as the linguistic problem mentioned above) more efficiently still amounts to addressing a non-technical problem in the eyes of the EPO. This is therefore unlikely to be patentable. If the invention is viewed as achieving a technical effect (e.g. identifying an irregular heartbeat as mentioned above), however, then features of the algorithm that improve efficiency may then contribute to the assessment of patentability. For example,if the algorithm for identifying an irregular heartbeat has been designed to use less battery power, then this may make it more suitable for implanted devices.
Whilst the section in the Guidelines on machine learning algorithms is relatively short, the guidance and examples set out there are useful. When read in conjunction with the section on mathematical methods, the Guidelines give a clear overview of how the EPO approaches the various different aspects of machine learning algorithms, including training, implementation issues, deployment, core technology and real-world use cases.
