AI Patents in the UK – Neural Networks Are Considered to Be Computer Programs

Last November, in an appeal from a decision of the UK IPO to refuse a patent application, the Patents Court held that artificial neural networks (ANNs) are not computer programs and so the computer program exclusion of Section 1(2) does not apply to neural networks. The Court of Appeal has now overturned that, holding that ANNs are computer programs and their patentability issues are no different than any other computer programs. The Court of Appeal’s judgment explicitly does not change the general position on patentability of computer-implemented inventions in the UK and, although some EPO case law is referred to, the UK approach remains different than the EPO’s “Comvik” approach to CIIS.

The invention at issue relates to a neural network trained to provide better recommendations, for example, better song recommendations. The neural network is trained to produce a vector representation of a file in an embedded space, with the distance between vectors being proportional to the distance between a semantic characterisation of the file, such as a textual description of the emotions induced by a song. Recommendations can then be issued based on the similarity of two vectors representing two songs. The application was first rejected for being a computer program as such, but an appeal to the High Court held that neural networks should not be considered as computer programs, meaning the invention did not fall within the exclusion.

The UK IPO appealed the Patents Court judgement on the grounds that a neural network should be considered as a computer program for the purposes of Section 1(2), and that the claimed invention did not involve a substantive technical contribution.

Whether a neural network should engage the computer program exclusion or not hinged on whether the weights of a neural network could be considered to be a program for a computer. The Court of Appeal held that a computer was defined as a machine which processes information (with this definition regarding both software and hardware implemented neural networks as computers), and that a program is a set of instructions that causes a computer to do something in a particular way. Emotional Perception attempted to further limit this definition, arguing that a program had to be written by a programmer to define exactly the steps to be followed by the computer. Emotional perception also argued that a programmer must understand how the problem must be solved in order to write a program, whilst a neural network can be trained without any knowledge of what the solution will look like. The Court of Appeal did not find these arguments persuasive, and held that the weights of a neural network are indeed a computer program, as they are instructions that cause the computer to act in a certain way. In other words, the weights as a whole instruct the machine regarding how to process information. The first instance’s emphasis on the fact that the instructions were created autonomously during training of a neural network was therefore rejected, with the exclusion of Section 1(2) being held to apply to a neural network.

The Court of Appeal then considered whether the claim made a technical contribution, thus avoiding falling within the exclusion of Section 1(2). The Court of Appeal held that the beneficial effect of the invention was subjective, as what makes a file worth recommending is subjective, with the semantic similarity between two files dependant on a user’s criteria and opinion. Although the analysis of the files is performed in a technical way, this is due to the use of a computer, which is excluded from patentability. The Hearing Officer’s initial refusal of the patent has therefore been reinstated.

The proceedings presented an opportunity to harmonise UK law with the approach taken to excluded subject-matter by the EPO. However, there was no suggestion during the proceedings that existing English case law should be departed from, despite the EPO using a consistent and longstanding approach to excluded subject-matter.

Although this decision may appear to be a setback for anyone looking to patent an invention involving some form of neural network, the overall result of this decision is only that neural network related inventions must meet the exact same criteria as any other invention involving a computer, as summed up by Birss LJ:

“The fact that the exclusion is engaged as a result of the first part of this appeal, simply means that ANN implemented inventions are in no better and no worse position than other computer implemented inventions.”

Thus, for an invention relating to, or employing, a neural network to be patentable in the UK, the invention’s contribution over the prior art must be technical, either improving the way in which the computer operates or better performing some technical function outside the computer.