Tomorrow, 14th May, the UK Court of Appeal will begin considering the patentability of an artificial neural network (ANN) in the eagerly awaited hearing of Comptroller-General of Patents, Designs and Trade Marks v Emotional Perception AI Limited. The findings of the Court of Appeal have potentially wide-reaching implications for applicants, offering new routes for patentability for trained AI inventions at the UKIPO. In this article we review the details of the case and the points that will be debated by the Court of Appeal.

Background

The case involves an invention by Emotional Perception AI which uses a trained ANN. Through the training process the ANN learns how to distinguish semantic similarity from the physical properties of a media file, allowing it to make recommendations to users. The patent claims consist of a method of training and the use of a trained ANN for prediction alongside a system comprising the same trained ANN.

In the first hearing before the UKIPO, the Hearing Officer found the ANN to be “no more than a computer program” with its contribution “no more than a computer programming activity“. It was therefore considered to be excluded subject-matter and not patentable. The Hearing Officer found that the purpose of the ANN (to provide semantically similar recommendations) did not demonstrate any technical effect beyond the running of a computer program and reasoned that if a trained ANN is decouplable from the software that supports it, the trained ANN would be abstracted far enough that it would constitute a mathematical method, again excluding it from patentability.

Emotional Perception AI appealed to the High Court (as discussed in our previous article) which overturned the decision finding that the implementation of the trained ANN did not constitute a program for a computer and therefore should not be excluded under UK patent law. The High Court found that the ANN learned through the training process how to operate and process the data, rather than simply implementing code which had been programmed by a human. Furthermore, the High Court found that the ANN did in fact make a technical contribution.

The appeal

The UKIPO Comptroller has appealed the High Court decision on four separate grounds:

1. The Judge erred in holding that the exclusion from patent protection for “a program for a computer … as such” was not engaged.
2. The Judge was wrong to rely on the Appellant’s ‘concession’ based on the division of hardware and software ANNs.
3. The Judge was wrong to exclude the consideration of the mathematical exclusion model.
4. The Judge was wrong to hold that the claimed invention involves a substantive technical contribution. According to the Aerotel test, if the claim is not solely excluded matter, it is required to demonstrate a technical contribution.

The grounds effectively seek to funnel the invention into an excluded subject-matter category, whether that is as a mathematical method or “a computer program… as such”, and demonstrate that the contribution made by the invention is non-technical, thereby rendering the invention unpatentable under the Aerotel test. While the assessment of the fourth step of the Aerotel test under the fourth ground is fact-sensitive, the consideration of exclusions to patentability under the first three grounds could have knock on effects across the field of machine learning inventions.

The judgement of the High Court viewed the trained ANN as operating at a (metaphorically) different level, between a computer program, since the act of training the ANN means that it learnt to complete a task itself, allegedly without programming, and a mathematical method. In essence, the avoidance of the exclusions to patentability presently relies upon the trained ANN not being considered a computer program by achieving “decoupling” from any underlying software, without abstracting the ANN to such a level that it could be considered a mathematical method. It remains to be seen how this approach will be justified on appeal.

The first and third grounds of appeal, therefore, may have particular significance for AI applicants. If it is upheld that the trained ANN is neither “a computer program… as such” nor a mathematical method, a carve out may be created for trained ANNs such that they effectively avoid both exclusions. Moreover, because the reasoning used thus far relies upon the ANN being trained, it appears as though this reasoning could be extended to most if not all trained AI.

Conclusion

In light of the Government’s ambition to make the UK a global AI superpower, the outcome of this appeal will be an important indicator of just how attractive the UK is as a jurisdiction for AI innovators. The Carpmaels team will be attending the hearing and sharing our thoughts afterwards on any key takeaways for ANN patentability in the UK. If you have any questions in the meantime, please do not hesitate to contact us.