Removing emotions from the equation: the Court of Appeal clarifies the patentability requirements for artificial neural networks

On 19 July 2024, the Court of Appeal overturned the High Court in its decision in Comptroller-General of Patents, Designs and Trade Marks v Emotional Perception AI Ltd, finding that the patentability of an artificial neural network (ANN) should be considered in the same way as any other computer implemented invention. The leading judgment, which was given by Birss LJ with Arnold LJ and Davies LJ concurring, reversed the High Court’s finding that the output of the ANN was technical in nature and affirms the original decision of the UK IPO.

In doing so, the Court of Appeal has not been ‘blinded’ by AI-related buzzwords. Instead, it has conducted a thorough and logical analysis of the technology in question. The judgment confirms that the technical output of the invention should be considered carefully, and has reined in the conclusions made by the High Court (which risked expanding patentability for computer programs far beyond the scope established in case law).

Background

Emotional Perception AI’s (Emotional Perception) patent application concerned a method and system which provides media file recommendations based on human perception or emotion rather than objective categories given to the file. These recommendations are created using ANNs and are typically in the form of music tracks.

As the Court of Appeal highlighted, ANNs are nothing new. Early (albeit simpler) examples of ANNs date back to the 1950s. Modern ANNs consist of neurons arranged in layers. All neurons receive inputs and transmit outputs. The first neuron in these layers receives inputs from ‘outside’ the system, and the last neuron in the layers produces the ‘output’ from the system. Each neuron will receive a number of inputs and apply weights to those inputs. The weighted values are added together, a bias is given to the result, and a function is then applied to produce an overall output.

Emotional Perception essentially utilises two ANN systems:

1. The NLP system: The ‘input’ is a word or words used to describe the emotional qualities of a piece of music, and the difference between semantic classes is reflected by the distance between vectors in semantic space (each vector representing a piece of music).
2. The ELP system: The ‘input’ is a piece of music, and the difference between the physical properties of this music is reflected in the distance between vectors in ‘property space’.

The weights and biases used in the ANN to produce a ‘correct’ output are generated during a training process in which the distances between vectors in ‘property space’ are adjusted by reference to differences between vectors in ‘sematic space’. For example, vectors in property space will be moved closer together if there is a high degree of semantic similarity.

Importantly, every ANN does not need to go through this training process in order to make appropriate recommendations. The weights and biases in a ‘trained’ ANN can be transferred to a ‘naive’ ANN. Birss LJ held that the process of training an ANN should be characterised as creating a computer program.

Emotional Perception claims that its tool delivers more sophisticated suggestions by using semantic and physical descriptions of music which mimic human perception and emotion. However, is such a tool patentable?

The Comptroller initially rejected Emotional Perception’s application on the basis that the invention was excluded subject matter. However, on appeal, Sir Anthony Mann concluded that s.1(2) of the Patents Act 1977 was not engaged, and even if it were, the subject matter of the claim was not excluded.

The computer program exclusion

Under s.1(2), “a program for a computer… as such” is excluded subject matter. Birss LJ therefore began his analysis by going back to first principles. Having reviewed previous rulings, Birss LJ formulated two definitions: a computer is "a machine which processes information" and a computer program is "a set of instructions for a computer to do something".

With these broad definitions in mind, the Court of Appeal then asked – do ANNs have any unique features which would mean they fall outside the definition of a computer program?

Birss LJ rejected attempts to distinguish the ANN using characteristics related to AI-inventions. Neither the involvement of human programmers, the training process, the permanence of ‘instructions’ nor the complex nature of the problem being solved were considered relevant to construction. Rather, the Court of Appeal took a straightforward approach and considered that an ANN is “clearly a computer” and the weights and biases of an ANN are merely “a set of instructions for a computer to do something”. Thus, an ANN should be treated just like any other computer program and s.1(2) should be engaged.

The technical contribution test

In its assessment of whether the claimed invention related to a computer program “as such”, the Court of Appeal dealt with stages 1 and 2 of the Aerotel v Telco 4-stage test in short order. It was agreed that the actual contribution of the invention was the provision of improved file recommendations. Birss LJ therefore analysed (i) whether this contribution fell solely within the exclusion and (ii) whether the contribution was actually technical in nature.

In its analysis of technical contribution, the Court of Appeal restricted its review to the step of sending the recommendation to the user as it considered that every other part of the contribution consisted of a computer program. However, Birss LJ also noted that sending an improved recommendation would fall within another category of subject matter excluded under s. 1(2), namely the presentation of information. Birss LJ considered the technical contribution of the technology by reference to its function: the invention was distinguished from standard file transmission by improving the recommendation to the user. However, this improvement is based on semantic or aesthetic qualities of music since the similarity or differences between files which form part of the recommendation are subjective, not technical, in nature. The recommendation is therefore not a technical effect, and the fact that there is an external transfer of data could not assist Emotional Perception for this reason.

Comment

In this judgment, the Court of Appeal has refocussed the analysis of patentability on the technical contribution of an invention. Mann J’s judgment that Emotional Perception’s invention reached its recommendation by “going about its analysis and selection in a technical way” risked a massive expansion of the scope of ‘technical contribution’. As Birss LJ pointed out, the fact that a computer uses measurable properties to make a semantic recommendation should make no difference to an Aerotel analysis: the very fact that computer systems are technical should not feature in this assessment. Any other approach would mean that previously excluded subject matter (such as the financial trading system in Merrill Lynch) could be considered technical.

Further, this decision does not mean that all ANN-based inventions will be strictly unpatentable, only that patentees must disclose a technical contribution in order to obtain protection. Unfortunately for Emotional Perception, the benefit of its technology (i.e. the ability to produce outputs categorised by aesthetic qualities) was arguably what ‘killed it’. If the Court of Appeal had affirmed the approach of the High Court, it risked a large expansion of patentability that arguably had no proper basis in law.

This judgment is not a death knell for patenting AI-related inventions, it simply re-confirms that the output of these inventions needs to be carefully considered. Crucially, one should not fall into the trap of seeing the term ‘AI’ or ‘ANN’ and assuming that it must produce a technical output.

In its review of the terms ‘computer’ and ‘computer program’, the Court of Appeal lamented the lack of guidance for interpreting these terms, particularly in light of the significant developments to these concepts in recent years. Accordingly, the UKIPO has now issued new guidance reflecting the fact that ANN-implemented inventions should be treated in the same way as any other computer implemented invention.