G4/19: Double Patenting Decision Solves Only Half the Problem
View more
The UK approach to patent-eligibility for software-implemented inventions appears to be shifting, in a more pro-patentee direction, based on the reasoning from the judgement in Emotional Perception AI Ltd v Comptroller-General of Patents, Designs and Trade Marks [2023] EWHC 2948 (Ch), released on 21 November 2023. The original refusal of the UKIPO was based on a determination that the patent application under consideration consisted of "a program for a computer ... as such", subject matter which is not patentable under UK patent law.
There is a lot to unpack in the decision, though a little background to the case is helpful first.
The technology
The case related to the training of an Artificial Neural Network (ANN), a form of machine-learning AI model, and then the use of the trained ANN for selection of media (such as a song) that was semantically similar to a reference media file. Importantly, the training methodology for the ANN was novel. In a high-level sense, an ANN model takes one or more inputs and determines one or more outputs from the input(s). The specific output(s) that is given for particular input(s) can be altered by altering internal “weights” within the ANN model. Training is the process by which the weights to be used are determined. Once the ANN model is trained, the weights would not usually be expected to change further.
During training, pairs of files sharing a content modality (e.g. both files were music tracks) are supplied to the ANN being trained. Each file has a semantic description associated with the content (at a very simple level, “Happy” or “Sad” or similar). A number of measurable properties are also determined for each file (such as tone, timbre, speed, loudness). When processed by the ANN being trained, a measurable property separation distance is calculated between the pair of content files, using the measurable properties and the weights of the ANN. At the same time, a semantic property separation distance is calculated between the pair of content files, using the semantic descriptions of the files.
After each training iteration, the weights are altered to reduce the measurable property separation distance for pairs of content files with low semantic property separation distances, and to increase the measurable property separation distance for pairs of content files with higher semantic property separation distances.
In other words, the ANN is being trained to return semantically similar content when considering only the measurable properties of the content.
The independent claims being pursued included claim 1 directed to a system, and claim 4 directed to a method. Both claims included the general steps of:
1) Training the ANN;
2) Using the trained ANN with a target data file to output an ordered list of reference data files which are semantically similar to the target data; and
3) Sending the list of reference data files to a user device and outputting the content of the reference data files at the user device.
Conclusions from the Case
There are several points which can be inferred from the decision, and which are certain to influence the IP strategy for AI-related innovations. They are worth drawing out here:
1) Subjective vs Objective is not relevant. One common principle applied to assessing whether an innovation results in a technical benefit (and therefore relates to patent-eligible subject matter) is to consider whether the alleged benefit of the invention is objective or subjective. An objective benefit is determinative and applies to all persons, sometimes all persons having a particular characteristic. A subjective benefit is something which may only apply for some people and is not determinative (i.e. it is not possible to determine to which people the benefit will apply). In paragraph 76 of this decision, the judge states that (emphasis of final sentence added):
“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.”
This seems to open the door to patent protection for innovations that provide subjective benefits, providing the underlying working of the innovation uses technical criteria.
2) An ANN is not a program for a computer. It was the case that an Artificial Neural Network can be implemented either in dedicated hardware, or through an emulation on a computer. In written submissions made after the oral hearing of the present case (in response to an invitation from the judge for further written submissions), the UKIPO’s representative conceded that an ANN implemented in hardware does not have a relevant computer program and would not fall foul of the exclusion to patentability for programs for a computer. The judge has taken this conclusion and used it to reason that if a hardware ANN is not a program for a computer, then neither is an ANN emulated on a computer, as found in paragraph 56.
This seems to provide helpful case law when seeking to argue in favour of patentability of inventions implemented on computers that are emulating hardware which might otherwise be used to implement the invention. Often, such specialised hardware can be found in machine learning applications, with trained models having fixed weights.
3) The trained ANN is capable of being an external technical effect. In paragraph 78 of the decision, it is said that:
“I therefore consider that, insofar as necessary, the 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 paragraph was made considering the situation where the ANN was considered a program for a computer, and concludes that even where the ANN was considered a program for a computer, the exclusion still would not apply because the trained ANN itself is an external technical effect. No in-depth further discussion is made of this point, but it seems to raise the interesting possibility that a claim to the trained ANN alone (either implemented in hardware or emulated using a computer) would also not fall foul of the exclusion to a program for a computer. Were this to be borne out by future practice at the UKIPO, it would certainly make the UK a much more friendly jurisdiction for machine-learning and other AI-related inventions.
4) The UKIPO has set an overly high bar for excluded subject-matter when considering programs for a computer. Several of the authorities cited in the decision (Protecting Kids all over the World (PTTWO) Ltd’s Application [2012] RPC 13 and Halliburton Energy Services Inc’s Patent Application [2012] RPC 12 and Symbian v Comptroller-General of Patents [2009] RPC 1 and AT&T Knowledge Venture v Comptroller of Patents [2009] FSR 19 are cases appealing against a refusal of the office. Indeed, in the first three of these, the UKIPO’s initial refusals were overturned, and the present case is another. This seems to indicate a general reluctance of the UKIPO to exercise a pro-patentee slant. Noting that the UKIPO is only bound by decisions of the courts, and that the courts generally have a remit to be able to consider the evidence of expert witnesses in a way that a Hearing at the UKIPO is less likely to do, it makes it much more difficult for the courts to influence practice in this area because they rarely get the chance. Applicants find that their cases are refused by the office, without awarding them any benefit of the doubt. Were the UKIPO to lower their bar slightly, more applications would proceed to grant. This would allow third parties the opportunity to challenge these through the courts, leading to a more rapid development of the case law in this area, which is sorely needed. It is notable that the proportion of patent decisions at the High Court which are appealing against a refusal by the UKIPO compared with patent decisions at the High Court in which validity is challenged is significantly higher for software-related inventions than for other fields.
As with all court decisions, it remains to be seen whether the legal principles and developments here are cemented in future decisions, or are overturned or otherwise further developed.
We look forward to future developments in this area with interest. If you would like to discuss patentability of your machine-learning innovations in the UK or Europe, please contact Chris Cottingham.