A recent High Court appeal (Emotional Perception AI Ltd v Comptroller-General of Patents, Designs and Trade Marks  EWHC 2948 (Ch)) has up-ended the examination practice of ANNs in the UK. The decision reversed the patent office decision refusing Emotional Perception AI’s application for an Artificial Neural Network (ANN) recommender system as being excluded from patentability as a “program for a computer … as such”. The UK patent office has suspended its current guidelines for examination of AI inventions and instructed examiners not to object to claims involving ANN under the “program for a computer” exclusion. How did we get here and what does this mean for the future of ANN (and AI) inventions in the UK?
ANN Inventions Before Emotional Perception
Under section 1(2) of the UK Patents act, things like mathematical methods and programs for computers are excluded from patentability “as such”, meaning that they can only be patent if the invention goes beyond the excluded thing. According to the examination guidance issued by the UKIPO about a year ago (now suspended but archived here), ANNs, unless limited to being implemented in fixed hardware only, were considered to be as of themselves programs for a computer. This means ANN inventions could only be patented if there was something more, a technical effect inside or outside of the computer, to escape the clutches of the exclusions.
This was one of the challenges faced by Emotional Perception AI during examination of their patent application for an ANN based media recommender system.
Emotional Perception AI’s invention relates to a media recommender system that receives a media file, let’s say a music track, from a user, determines some measurable parameters of the track, such as tone, timbre, speed and loudness, and returns a semantically similar track to the user. The parameter are passed through an ANN to obtain a corresponding parameter embedding for the track in a parameter space and the embedding is used to find similar tracks (tracks nearby in embedding space) in a database.
The trick the invention uses to make the parameter embedding semantically useful is a dual embedding approach to training. The ANN is trained with pairs of tracks and corresponding semantic descriptions for each track. A semantic embedding is found for each of the semantic descriptions in an semantic space using natural language processing techniques. An ANN is set up to take measurable parameters of the track, and output the parameter embedding. The ANN is trained so that distances in the parameter space reflect distances in the semantic space. As a result, tracks that are close in the parameter space are also close in the semantic space and hence close in semantic meaning.
The patent office decision
The patent office decision identified a contribution of the invention as a simplified and more efficient recommender system as compared to prior art systems that were said to train multiple ANNs to make their recommendation. That much was undisputed. However, the contribution was considered to be a program for a computer and there was no technical real-world effect in selecting and sending the tracks, since any effect depended on the subjective perception of the user. Hence, the invention was excluded as a “program for a computer … as such”.
This was the only exclusion applied by the decision. While the concept of an ANN, abstracted from its implementation was said to be a mathematical method per se, its use in a recommender system was sufficient for the invention to escape the mathematical method exclusion.
The High Court Decision – not a “program for a computer … as such”
In considering the exclusion of a “program for a computer”, as such, the judge considered ANNs implemented in hardware or in software (the invention covered both). The judge asked two questions: Where is the computer? And where is the program?
Considering hardware ANNs first, while not a computer in the common use of the word it was an apparatus for processing data and therefore a computer for all intents and purposes. However, there was no sequence of steps or instructions that is executed by the ANN when it processed data and hence no program for a computer. As far as ANNs implemented in software, or “emulated”, there clearly is a computer running the emulation.
So there was a computer, but what about a program for a computer? As far operating a hardware ANN was concerned, be that once trained or the internal processes during training, it was agreed between the parties that there was no program involved, as there was no sequence of instructions specified by a programmer being executed. As far as the emulated (software) ANN was concerned, the judge considered the training stage and the stage when the computer operates the trained ANN. The former was found to involve a program that provided the training. The latter was found not to involve a program for a computer because there was no distinction of substance between a hardware and software implemented (“emulated”) ANN.
Having established that a computer program was involved in the training of the ANN, the judge went on to consider whether the exclusion applied, that is whether the invention was “a program for a computer…as such”. The judge noted that the “programming involves setting the training objectives in terms of the structure of the ANN (if in software) and the training objectives”. So there was a computer program but the claims went beyond that and claimed the idea of using pairs of files for training and setting the training objective and parameters accordingly. The actual program was considered “a subsidiary part of the claim and […] not what is claimed”. In the words of the judge:
The claims go beyond that. The idea of the parameters itself is not necessarily part of the program. On this footing as a matter of construction the claim is not to a computer program at all. The exclusion is not invoked.”
It followed that the application was not caught by the exclusion to patentability. The decision suggests that any invention that relates to “special” ways of training an ANN is not caught by the exclusions to patentability, because the contribution is not a computer program at all.
The High Court Decision – “technical contribution”
The first part of the decision suggests a radical departure from the current patent office practice in examining ANN related inventions (and possible all software inventions – I will come back to this at a later time). It is therefore not surprising that the judge considered the case in the event that he is wrong and that training the ANN does involve a computer program or that there are computer programs to be considered elsewhere, such as in the emulation of the ANN. In that case, it would be necessary to continue the enquiry as to whether the contribution fell solely within excluded subject matter and whether the contribution is technical in nature, known in the UK as steps 3 and 4 of the Aerotel questions.
The judge considered a range of prior decisions, from drill bits to television guides and parental alerts for computers. He arrived at the conclusion that the patent office decision again got it wrong and that, even if the invention related to a computer program, its contribution was technical in nature and did not fall solely within the exclusions. The reasoning is notable for the unstated assumptions that the operation of the ANN is technical per se and that the resulting sending of a file is enough to make a technical contribution, irrespective of the user. In the words of the judge:
The correct view of what happened, for these purposes, is that a file has been identified, and then moved, because it fulfilled certain criteria. True it is that those criteria are not technical criteria in the sense that they can be described in purely technical terms, but they are criteria nonetheless, and the ANN has certainly gone about its analysis and selection in a technical way. It is not just any old file; it is a file identified as being semantically similar by the application of technical criteria which the system has worked out for itself. So the output is of a file that would not otherwise be selected. That seems to me to be a technical effect outside the computer for these purposes, and when coupled with the purpose and method of selection it fulfils the requirement of technical effect in order to escape the exclusion. I do not see why the possible subjective effect within a user's own non-artificial neural network should disqualify it for these purposes. […]It would not matter if the user never listened to the file.”
The judge noted that this reasoning is based on the end result helping to take the case away from being a computer program as such. However, if one considered that the computer program in question is either the training program or the overall training activity, the trained ANN can be considered as a technical effect which prevents the exclusions applying. Again, in the words of the judge:
“[T]he trained hardware ANN is capable of being an external technical effect which prevents the exclusion applying to any prior computer program. There ought to be no difference between a hardware ANN and an emulated ANN for these purposes.”
This is even more remarkable. It suggests that any invention that involves training an ANN to do something is not excluded from patentability because, even if a computer program is involved in the training, there is in any event an external technical effect in the form of an ANN.
The decision was specific to the exclusion from patentability of “ program for a computer … as such”. Other exclusions, and in particular the exclusion of mathematical methods, were not part of the decision. Although patent office’s counsel tried to raise this point, it was not considered as it was raised too late in the proceedings and was not part of the decision under appeal (see above).
While it is unfortunate that the decision does not provide us with direct guidance on the mathematical method exclusion. Yet, this may be a moot point. The judge held that a trained ANN is an external technical effect which prevents the (computer program) exclusion from applying. The presence of a technical effect means that an invention is not an excluded thing “as such”, irrespective of which exclusion is evoked, so in principle, the High Court decision can be expected to also neutralise the mathematical method exclusion for ANNs.
The patent office has reacted swiftly, suspending its guidance on the patentability of AI invention and instructing examiners not to “object to inventions involving ANNs under the “program for a computer” exclusion”. For the time being, this leaves it open how examiners will deal with ANN inventions under the other exclusions, notably for mathematical methods, or with AI inventions that do not involve ANN. Hopefully these questions will be addressed in the updated guidance to be issued in due course.
The deadline for appealing the High Court decision has now passed earlier this month and we have just learned that the UKIPO has sought and been granted leave to appeal the decision, somewhat surprising given the swift reaction to change examination practice in line with the decision. We expect the appeal to take about 6-12 month and so it will be a little while before we know whether the current liberalisation is a blip or a permanent change. However, the UKIPO have confirmed that they will continue to apply the new guidance regardless of the pending appeal, so the decision reflects the current law regarding ANNs and the exclusions from patentability for the time being. While the details of how this will, in practice, change the patentability of ANN (and more generally AI) inventions in the UK remain open, it is clear that a significant liberalisation can be expected, at least for now. This may make the UK the European jurisdiction of choice for ANN patents, in particular if the decisions is upheld on appeal.