With the Supreme Court set to hear the appeal in Emotional Perception AI’s patent application next week, the UK patent profession eagerly awaits the outcome of the appeal.

Regardless of the outcome of Emotional Perception AI’s patent application, we would like to see the Supreme Court take the opportunity to recognise that the training of an ANN (or other machine learning model) can – and often will – contribute technical character.

For example, an invention using an ANN may achieve technical character through the nature of the data that it processes, or the outputs that it provides, and this will typically be the case where the ANN is used for performing or controlling technical processes outside of the ANN implementation.

In many cases, however, and increasingly, the technical character of an invention using an ANN may reside (sometimes exclusively) in the training process. For example, the training process may provide a better ANN, that can provide more accurate predictions.

At least where the ANN serves a technical purpose, if the training process helps to achieve that purpose, the training process should also be considered to contribute to the invention’s technical character.  This is already acknowledged by the European Patent Office’s Guidelines for Examination (here), for example.

In Emotional Perception AI’s patent application, as previously discussed here, the ANN is used to find semantically similar data files to a target file using their objectively measurable properties (e.g. to suggest subjectively similar songs to a user by measuring objective properties of a target song). As noted by the Court of Appeal, therefore, in Emotional Perception AI’s patent application, the file recommendations are based on subjective (or ‘aesthetic’) qualities.

The “clever” bit of the application, then, appears to be how the ANN is trained to provide these recommendations. A database of reference media files (e.g. songs) is used, with a textual semantic description of each file provided by a human (e.g. describing a song as “frenetic” or “light”). The ANN is trained by taking pairs of files from this database and adjusting the parameters of the ANN to reproduce a semantic similarity distance between each pair of files (determined by natural language processing of the text descriptions) but using only the objective properties of each file. Once trained, the ANN can then be used on a new target file (without a text description) to find semantically similar data files from the database.

In our view, the Court of Appeal should not have dismissed the training process as being subsidiary in nature and irrelevant (see paragraph 74 of the judgment). That is, even if the system is used for non-technical purposes, if the training process is done in a technical way, this may still provide a technical contribution.

For instance, the Court of Appeal held that an ANN is a computer (i.e. in that it processes information), with the weights (and biases) then defining a computer program since these will control how the ANN will process that information.  We think care should be taken to distinguish between an ANN, as an abstract computational model, and its instantiation (e.g. stored as data, compiled in software, implemented in hardware), training, and use in execution.  However, even if the definitions used by the Court of Appeal are adopted, if there are “clever” aspects to the training process that potentially provide an advance on the prior art (and that take the system beyond the simple use of an “off the shelf” ANN to provide such file recommendations), this would seem to go beyond a computer program as such, and this contribution therefore needs to be assessed for the purpose of analysing the patentability of the claims.

This would apply similarly even if following an alternative view that an ANN is not a computer, but is rather a model that can then be executed as part of a computer program, for instance. Again, it should be recognised that in many cases it will be the training process that defines the technical character of an invention, and that depending on the facts, the existing “Aerotel” signposts used for assessing excluded subject-matter in the UK may accordingly point to patentability.

It may be helpful however for the Supreme Court to take this opportunity to explicitly recognise a new patentability “signpost” acknowledging that the training of an ANN may in such cases contribute to technical character.