In its annual update of the “Guidelines for Examination”, the European Patent Office (EPO) has provided further guidance for its examiners in relation to the patentability of inventions relating to mathematical methods and computer programs. This updated guidance is of particular relevance to inventions relating to the fast-growing field of Artificial Intelligence (AI). In part 1 of this article, we provide a summary of the key points from the updated guidelines that are relevant to AI inventions. Part 2 will follow, in which we will provide an in-depth assessment of the impact of the new guidelines on the patentability of AI inventions.

By way of background, the patentability of computer implemented inventions at the EPO has long been governed by the general principle of requiring a non-obvious technical solution to a technical problem, as established by the EPO Boards of Appeal in T0641/00 (COMVIK).

In particular, features which are directed to one of the exclusions to patentability cited in Article 52(2) EPC can be considered to be non-technical if lacking an associated technical effect. One such exclusion is if a feature is considered to relate purely to a mathematical method. A close inspection of the technical workings of any AI application is likely to reveal some form of mathematical model at its heart. Therefore, it is possible that an AI invention may be considered to be excluded from patentability on the grounds of being directed to a mathematical method as such. In other words, an AI invention may be considered to be non-technical in the eyes of an EPO patent examiner and therefore unlikely to be allowed by the Examining Division of the EPO. Over the past few years, the EPO Examining Division has attempted to harmonize practice when it comes to determining what is and is not technical in nature by re-writing its guidance for examiners in relation to assessing the patentability of computer implemented inventions. Last year’s update to the “Guidelines for Examination” focused on Presentations of Information and User Interfaces. In this year’s update, the focus has been on providing further guidance on what the Examining Division of the EPO is likely to consider technical and non-technical with respect to mathematical methods and AI inventions.

The EPO operate a “two hurdle” approach. According to the first hurdle the claimed subject matter must have a technical character. However this is easy to overcome in practice by, for example, amending from a method [of performing a machine learning algorithm] to a claim computer-implemented method. The second hurdle is more difficult, and requires a technical contribution to inventive step. However claims may include a mix of technical and non-technical features and all features contributing to technical character should be taken into account – for example a mathematical method may contribute to inventive step if, say, it is applied to a specific technical problem such as image processing. In another example the necessary technical character may be provided by claiming specially adapted hardware to implement a procedure.

In more detail, the new guidance provides two related directions in which inventions comprising mathematical methods can be claimed in order to avoid potential exclusion from patentability. The first direction is to restrict the claims to a specific technical purpose or application. This direction is based upon the decision in T1227/05 (INFINEON) in which a mathematical method for generating random numbers according to a specific distribution was found to be technical if the claim was restricted to a “computer-implemented method for numerical simulation of a circuit subject to 1/f noise”. In this case, the Board of Appeal decided that “numerical simulation of a circuit subject to 1/f noise” was a specific technical purpose to which the mathematical method for generating random numbers according to 1/f noise contributed.

In relation to AI inventions, the guidelines provide the following as specific examples of technical applications:

  • use of a neural network in heart monitoring apparatus to detect irregular heartbeats
  • classification of images, videos, audio or speech signals based on low-level features (e.g. edges or pixed attributes for images) and the training of such classifiers, for example:
    • a. detecting persons in a digital image
    • b. seperation of sources in speech signals
    • c. speech recognition, e.g. mapping a speech input to a text output

It is particularly noted that classification of text documents solely in respect to their textual content is not regarded as a technical purpose but a linguistic one.

Other more general examples of technical purposes provided in the guidelines include:

  • control of a specific technical system/process
    • a. outside of a computer, e.g. X-ray apparatus or a steel cooling process
    • b. inside the computer, e.g. resource allocation in a distributed computing system
  • audio/image/video enhancement or analysis, e.g.
    • a. denoising of digitial images
    • b. estimation of quality of transmitted digital audio signals
  • data encoding or compression for transmission and/or storage
  • encryption/decryption or signing electronic communications
  • optimising load distribution in a computer network
  • providing a genotype estimage based on an analysis of DNA samples
  • providing a medical diagnosis by an automated system processing physiological measurements

and the somewhat vaguely defined:

  • simulating the behaviour of an adequately defined class of technical items or specific technical processes, under technically relevant conditions
  • technical design of a specific technical system/process, including determination of optimal technical parameters

Thus, an AI invention falling within one of the above example categories with claims restricted to that specific purpose is likely to be considered technical. The precise level of specificity required for restricting the claims is however not fully known. What is certain though is that a generic statement of “for the control of a technical system” will not be seen as a restriction to a specific technical purpose and will almost certainly be objected to.

In addition, the technical purpose is primarily determined by the direct technical relevance of the results provided by the mathematical method rather than the nature of the data input to the mathematic method.

The second direction requires claiming a specific technical implementation of a mathematical method that has been specifically adapted to run on a particular hardware configuration, or in the words of the EPO, “the design of the mathematical method is motivated by technical considerations of the internal functioning of the computer”.

The guidelines provide a specific example of the adaptation of a polynomial reduction algorithm to exploit word-size shifts matched to the word size of the computer hardware in order to provide an efficient hardware implementation of the algorithm.

Whilst there are few examples of what constitutes a specific adaptation, it seems likely that the specific adaptation of an AI algorithm to particular hardware, for example, the adaptations required for implementing a neural network on a GPU or other distributed/parallel processing hardware will be considered technical and therefore unlikely to fall under a non-patentable subject-matter exclusion in Europe.

In summary, a claim to an AI algorithm based upon a mathematical or computational model on its own is likely to be considered non-technical. A simple recitation of a type of artificial intelligence model being employed, such as, a neural network, a support vector machine, or reasoning engine alone in the claims is unlikely to overcome such objections. However, restriction of the claim to a specific technical purpose and/or a specific technical implementation may impart technical character onto the AI algorithm and thus the invention may be considered patentable by the EPO.

As AI becomes increasingly prevalent in our society, we are sure to see more patent applications directed to AI inventions. The updated “Guidelines for Examination” provides further guidance on what may be considered patentable by the EPO. In part 2 of this article, we will take a more in-depth look the new guidance and assess its impact for AI practitioners.

The EPO's full guidelines for examination, Index for Computer-Implemented Inventions, can be found here. A preview of the 2018 Guidelines can be found here.