In December 2023, the UK IPO announced an immediate change to the examination of patent applications involving artificial neural networks (ANN). Patent examiners should no longer object to inventions involving an ANN under the “program for a computer” exclusion in the Patents Act.
This prompt response by the IPO follows the High Court ruling in Emotional Perception AI Ltd v Comptroller-General of Patents, Designs and Trade Marks [2023] EWHC 2948 (Ch) which overturned the hearing officer’s decision. The judge concluded that a patent for an ANN which makes an emotional analysis of pieces of music is not to be considered a patent for a computer program as such.
In this article we take a brief look at this decision and the IPO’s new guidelines.
Emotional Perception AI’s application claims an ANN-based process for associating musical tracks together based on both their emotional and musical similarity, with songs of the same genre and mood being considered closely related. This is then to be used to select a new track that is sufficiently similar to any given input track.
This similarity assessment is done by the ANN first analysing human-written descriptions of the mood of each piece, using a natural language processing algorithm to determine which descriptions are most similar and quantifying this similarity into a “distance” between them. The AI then analyses the tracks automatically, on a number of (human-specified) musical qualities such as tone and speed, to create a second distance. Finally, the mood-based distance is then modified to either be longer or shorter based on the musical distance value, to create an overall similarity score.
A brief summary of how ANNs work is included at the end of this article. The key feature is that they function through having a set of “weights” developed through training: these weights are adapted autonomously by the ANN until their analysis of training data provides the same outcome as the human-determined “right” outcome for that data, and are then immutable in further operation of the ANN.
The Patents Act 1977 excludes from patent protection "a program for a computer ... as such". (section 1(2)(c)).
The IPO hearing officer had rejected Emotional Perception’s application on the basis of the computer program exclusion. Emotional Perception had appealed to the High Court.
The judge, Sir Anthony Mann, considered separately whether (1) this ANN would be a computer program, and then (2) whether (if it were a computer program) it would nonetheless be patentable for creating a technical effect beyond just being a simple program. In both cases, Sir Anthony Mann appears to have adopted a broad and permissive approach to applying the principles developed in the case law under s.1(2)(c). He allowed the appeal. In brief his conclusions were:
In light of these broad conclusions, the IPO has (for now) chosen to render this as effectively an absolute rule - the computer program exemption is not to be applied to any ANN patent. It will be updating its Manual of Patent Practice, and guidelines on AI inventions to reflect the judgment.
The IPO’s announcement does not mean that AI inventions are suddenly to be afforded a carte blanche even when the ordinary requirements for a patent have been met. For example, whether the ANN was a mathematical method as such (and therefore fell within that exclusion) was a further point of dispute in this case not addressed in the appeal. Future applications may find this exclusion to be more closely scrutinised. In addition, this broad approach may be refined by further judgments, as the courts develop their approach towards AI, not least because it has been reported that the IPO has appealed this decision.
Watch this space!
Neural networks consist of layers of operators (often anthropomorphised as “neurons”), each of which can perform simple operations on their input data based on a set of variable “weights”, before passing their output on to the neurons of the next layer. The input value gets passed to the initial layer of these neurons, leading to a cascade of information flowing through the network, providing a simplified mimic of the neural pathways of a brain.
To ensure the ANN reaches an appropriate conclusion, a “training stage” is run first, where the ANN is first fed information which a human has already analysed and determined a defined expected conclusion. The outcome of the analysis by the ANN is then compared to the human-determined expected conclusion, with the neurons then altering their weights (in accordance with a pre-determined algorithm) to try to improve this match. This analysis, comparison, and improvement cycle continues until the outcome of the analysis agrees with the expected outcome to a sufficient degree and level of consistency.
When this consistency has been reached, the ANN is deemed “trained” and the operator removes the ability for the weights to be changed. The ANN can then be used to analyse new information, to come to a conclusion that is likely to be consistent with the analysis the human made in the training data.
Written by Fred Cascarini.
