Mon 27th Nov 2023
Patenting AI and machine learning in the UK – all bets are off?
The UK High Court has issued a game changing judgement in Emotional Perception AI Ltd vs Comptroller-General of Patents, Designs and Trade Marks (2023).
Emotional Perception appealed the decision of the Hearing Officer (BL/O/542/22), which rejected the patent application as a “program for a computer…as such” excluded under Section 1(2)(c) of the Patent Act 1977. However, this has been overturned by the High Court on appeal.
The patent application relates to the use of artificial neural networks (ANNs) applied to file recommendation. The invention is concerned with training an ANN to perceive semantic similarity or dissimilarity between media files and using the trained ANN to recommend a file which is semantically similar to a given input. The files may be audio files, video files, static image files, or text files. According to the Appellant "[i]n these pairwise comparisons the distance in property space between the output (property) vectors of the ANN is converged to reflect the differences in semantic space between the semantic vectors of each pair of files. The result is that in the trained ANN, files clustered close together in property space will in fact have similar semantic characteristics, and those far apart in property space will have dissimilar semantic characteristics.”
Part of the judgement concern the application of existing UK case law on file recommendations. According to para. 76. of the judgment, “[t]he correct view…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.”
This conclusion, in itself, has potentially broader applicability to the field of machine learning recommendations, including applications that would generally not be regarded as technical per se, such as recommending items to add to a shopping cart or targeted advertising.
However, the conclusions of para. 58 are wider reaching still. According to the judgement, a software-implemented neural network “is not a program for a computer” as it operates “at a different level (albeit metaphorically) from the underlying software on the computer” (para. 56). The software-implemented neural network is characterised as “computer emulation” of a “hardware ANN”. On the face of it, any invention which involves the use of a neural network is now patent eligible on the same reasoning.
Paras. 77-78 consider technical effects arising from neural network training. Recognising the existence of a training computer program, the judgment concludes that a trained neural network is per se a technical effect which prevents the computer program exclusions from applying to the training program. This characterisation of neural network training as a technical activity contrasts with the current practice of the European Patent Office, which requires the trained model to serve a “technical purpose” . The characterisation of a software neural network as a hardware emulation appears also in this part of the judgement.
Para. 61 provides an alternative line of reasoning, characterising the ANN parameters learned in training as “not necessarily part of the program”, and concluding that as a matter of construction, the training program is merely a “subsidiary part of the claim” that does not invoke the exclusion.
An emphasis is placed on the learning of semantic representations, whereby semantic relationships are captured as geometric relationships in semantic embedding space. This is a widely-used dimensionality reduction technique in machine learning, with perhaps one of the most famous modern examples being the “Word2Vec” algorithm published in 2013 .
It is interesting that the UKIPO apparently conceded that a hardware ANN would not be excluded (para. 43). This appears to have been influential in the line of reasoning concerning ‘hardware emulations’. It remains to be seen how this will be applied to other machine learning models or computer programs, or whether “artificial neural networks” have been afforded a special status. It is true that neural nets are sometimes visualised using a computational graph of nodes and edges representing calculations performed on the computer. However, any computer program can be visualised in the same manner as a computational graph of nodes and edges and, in principle, could be implemented as a ‘hardware’ version of this graph. Presumably it is not the case that any abstract computer program now can be characterised as a “software emulation” of the non-excluded and hypothetical hardware implementation of its computational graph, although it is hard to see what would set neural nets apart from other computer models or computer programs in this respect.
Curiously absent is any detailed discussion of the “mathematical methods” exclusion, which is the predominant exclusion under which machine learning inventions are examined at the EPO . This appears to have arisen from a procedural omission on the part of the UKIPO, with para. 81 noting that “Hearing Officer did not make mathematical method an alternative basis of his decision and Ms Edwards-Stuart did not file a respondent’s notice to resurrect it”. The implications of this remain to be seen.
For more information on patents relating to AI and machine learning, please contact Tom Woodhouse.
This briefing is for general information purposes only and should not be used as a substitute for legal advice relating to your particular circumstances. We can discuss specific issues and facts on an individual basis. Please note that the law may have changed since the day this was first published in November 2023.
 Tomas Mikolov, Kai Chen, Greg Corrado, Jeffrey Dean “Efficient Estimation of Word Representations in Vector Space” (2013) https://arxiv.org/abs/1301.3781
 EPO Guidelines, G-II 3.3.1 Artificial intelligence and machine learning