Emotional Perception and the limits of technical character at the UKIPO

Members of Marks & Clerk’s AI team attended the Emotional Perception hearing at the UKIPO yesterday as observers. Unfortunately no decision is due yet – the UKIPO will issue their decision sometime in the next 12 weeks. Having said this, it raised some interesting questions which may become central to patentability for AI inventions in the UK: 

When does an AI training objective contribute technical character, and when is it merely an abstract goal? 

The applicant's position is deceptively simple. The invention is not merely a neural network. The technical contribution lies in the learning objective itself: training the network so that distances in a semantic space converge with distances derived from measurable properties of data files. The argument was not that all machine learning is technical. The argument was that this particular training objective is technical. 

One interesting point to note was that the arguments focussed primarily on the training objective, and did not place as much weight on the fact that the claims are limited to data with shared "content modalities” of video, audio or image. The network therefore operates with properties of data that patent offices have traditionally accepted as technical in nature. The applicant did argue that the convergence is not based on subjective concepts alone. Instead, it seeks to map relationships derived from measurable characteristics of media signals onto semantic relationships perceived by users. Viewed in this way, it could be argued that the invention resembles familiar technical fields of signal processing, image analysis, or speech recognition. 

But what happens if the data is not rooted in physical modalities? This is where Emotional Perception starts to become genuinely complex.  

In the hearing, banking software was used as an example, where an automated tool could replace a loan officer assessing data and criteria in deciding whether to award a loan. Suppose the properties of the training data are a number of County Court Judgements to a person's name, details of their previous financial applications etc. The UKIPO queried whether the application of a similar neural network architecture to such a system would be considered technical.  

During the detailed discussion of the specific functioning of the invention, the Hearing Officer attempted to focus the arguments on one clear question – what is the specific technical contribution of the invention.  The applicant and the Hearing Officer spent some time debating the limitations and boundaries of what makes the training objective technical in Emotional Perception.  

This tension over boundaries was further amplified by the applicant’s attempt to argue with reference to G1/19 that the trained ANN reflects a “very specific technical implementation” and that it improves the internal functioning of the computer.  

However, this argument treads on precarious ground given the Supreme Court’s definitive ruling that an "ANN is an abstract entity, and not a physical object” (e.g., a computer). Ultimately, whether the Hearing Officer accepts that this training objective genuinely alters how the computer functions, or simply views it as an attempt to artificially reframe the algorithm as physical hardware, remains a critical open question.  

Is the technical nature of a training objective related to the nature of the data? With this view, images, audio and video support technicality whereas business records do not. However, this seems too simplistic and does not consider the difficult middle-ground or nuance. The boundary is much messier than a simple categorization of data types: text processing can clearly be technical in some contexts, with translation systems routinely processing textual information. Image processing can sometimes be considered borderline non-technical, for example when image recognition strays into the cognitive regime of assessing emotions induced by an image.  This difficulty was amplified by repeated references to G1/19 where it is cautioned that technicality cannot be confined to a closed list of qualifying data types, observing that “it cannot be a necessary condition, if only because the notion of technicality needs to remain open”. 

Is it related to the technical purpose? The EPO tends to focus on technical purpose, with technical application being the ‘get out of jail free card’ for many software-related inventions. This approach is attractive but uncomfortable:  technicality starts to look less like an inherent property of the invention and more like a value judgment of the application domain. If technicality rests primarily on the application domain, then inventions with identical implementations may receive different treatment simply because one is directed to medicine and the other to finance. 

If instead we look at the implementation, is it related to the computer being a better computer? It was repeatedly highlighted that the Emotional Perception invention performs a better search because of the convergence of distances between the semantic space and the measurable property space. By training on convergences, you end up with a better computer implemented search engine. However, the hearing repeatedly returned to an issue that remains unresolved: would training based on converging distances still be technical if applied in the field of banking? If so, perhaps the technical contribution lies in the learning objective itself and how it makes the computer function. If not, then the apparent technicality may depend on features external to the learning objective, such as the nature of the data being processed or the field in which the model is deployed. 

The danger for the applicant during the hearing was that framing the invention around its ability to identify semantically similar music or make better recommendations risked walking straight into a trap. For example, if the Hearing Officer considered improving the process of recommending media to a user to be non-technical, this framing might inadvertently invite the UKIPO to filter out the features related to the core training mechanism as purely non-technical features because their downstream purpose is just that. While this specific vulnerability was not dissected in detail in the hearing, the argument might be framed more safely if it were argued that the training mechanism results in more accurate representations of the physical properties that are used downstream to make a given prediction of an image, a video, or audio, rather than referring specifically to the problem of reducing the “semantic gap”, which the Hearing Officer might consider subjective and non-technical.  

As you can see, the question quickly becomes circular and we do not propose to guess the conclusion the UKIPO will reach. However, the answer may ultimately determine the limits of Emotional Perception.