In this update, Caroline Day delves into a landmark UK legal case that could redefine the patentability of artificial neural networks (ANNs).
Emotional Perception AI recently defended their invention in court, challenging the UK Intellectual Property Office's (UK IPO) decision to exclude their ANN from patent protection. This case raises fundamental questions about what constitutes a 'computer' and a 'program,' and explores whether ANNs can be considered technical contributions under UK patent law.
In one of the most hotly anticipated cases in UK patent law of recent years, Emotional Perception AI had their day (and a half) in court last week, defending the patentability of their invention against the UK IPO. The UK IPO had appealed a judgement of the High Court in November last year, in which it was determined that Emotional Perception AI's patent was not excluded from patentability in the UK.
As a quick refresher, this case concerns an ANN which recommends content such as music tracks, images or text files. The case has focused on music files, but the claims are broader than that. The ANN is trained using pairs of data files, the first describing a semantic similarity between previously generated text descriptions of two pieces of content, and the second describing a property similarity between previously determined physical characteristics of the same two pieces of content. The ANN learns, through back propagation, how to map the differences in property space to the distances in semantic space.
Now, using the trained ANN, content which is semantically similar to a target content can be identified based on the physical properties. This content can be recommended to a user, and the idea is that it is more likely to appeal to that user.
The case turns on whether the claims are excluded subject matter because they relate to a 'computer program as such'. If so, the invention is not patentable.1
What is a computer and what is a program?
These questions are at the heart of this dispute and there was no shortage of suggestions during the hearing as to what could be either.
Let's take the first part first: what is a computer? Here, the UK IPO's case was clear: they submit that a hardware ANN is a computer and a software ANN is run on a computer. This was also the conclusion of the judge in the appealed decision. Emotional Perception AI's case was less straight forward. Was a sextant a computer? Was a slide rule? Did it need some form of central intelligence (e.g. a CPU)? However, their ultimate position was that an ANN wasn't one.
If you were to ask me (and nobody has), I'd likely conclude that a hardware ANN computes and is therefore a computer. Whether or not a sextant or slide rule fulfils the sematic definition matters not at all in my view as they do not execute programs.
But what is a program? Again, the UK IPO was fairly adamant: it's the weights and biases which result from the training. Emotional Perception AI favoured a program being a set of logical instructions. The judge at first instance didn't see it in either way and called the weights and biases parameters, which appears (to me at least) to be a more accurate description. While a program would run to cause the ANN to carry out its method, this would also run in a perfectly satisfactory manner with different parameters. The results it outputs would however be unlikely to be as useful. Thus, it is the systematic tuning of the parameters which makes the difference here.
Is it technical?
In the first instance proceedings, the judge considered the question of whether or not the claim made a technical contribution and decided yes: the recommended file was not 'any old file', it was a file identified as being semantically similar. Moreover, something was actually output.
This question also arose in the Appeal hearing. The UK IPO advanced that what was output was indeed 'any old file', as the recommendation was to a piece of content in an unaltered form. By contrast, Emotional Perception AI said that the claimed system performed a technical analysis based on the physical characteristics of the content. What is more, the system has gone about its analysis and selection in a technical way. The result was a better system as a whole, which produced a better recommendation.
What happens next?
Currently, the UK IPO is not making the objection that claims to artificial neural networks are excluded from patentability on the basis that they are computer programs (although other objections may be made). In practical terms, this means that ANNs are getting a slightly easier ride at the UK IPO than many other computer implemented inventions. This will likely continue at least until a decision has issued. The decision may issue at any time now and, based on recent track records, we can expect it to be issued fairly soon, likely within a month or so.
At that point, we will know whether trained ANNs will maintain their somewhat special status at the UK IPO. If they do, this author at least lives in hope that some of that magic dust may fall on other computer implemented inventions which may similarly provide technical advances based on carefully tuned parameters.
Footnote
1. It is also important for patentability that the claims are not deemed to relate to "a mathematical method as such", but this was a lesser part of the discussion and it remains to be seen whether or to what extend the judgement might address this particular issue.
The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.