In light of the UK and EU decisions in the Emotional Perception AI Ltd and Mitsubishi cases, respectively, regarding the patentability of Artificial Neural Networks (ANNs), Bharathwaj Ramakrishnan examines the situation in India and analyzes whether ANNs would fall under the ambit of Section 3(k) interpretations by the courts. Part I of his two part guest post deals with the above decisions of the UK Court of Appeal and European Patent Office’s Board of Appeal, and Part II deals with how Indian courts may interpret ANNs in light of Section 3(k) and the recent rulings like the ones in the Open TV and the Lava-Ericsson cases. Bharathwaj is a student at the Rajiv Gandhi School of Intellectual Property Law, IIT Kharagpur and loves reading books and IP law. His previous posts can be accessed here and here.
Part I-Artificial Neural Networks: Are They Mathematical Methods or Computer Programmes, or Does it Even Matter?
By Bharathwaj Ramakrishnan
Recently, the UK Court of Appeal in Comptroller-General of Patents, Designs and Trade Marks v Emotional Perception AI Ltd [2024] EWCA Civ 825 (here, for a summary of the judgment, see here) dealt with the question of whether an Artificial Neural Network (ANN) is a computer program and then concluded that it is a computer program as provided in Section 1(2)(c) of the UK Patent Act. This ruling is significant and must be understood along with a Board of Appeal judgment from European Patent Office T0702/20 (Mitsubishi) (see here), which also dealt with ANN and classified ANN as a set of mathematical operations implemented on a computer. These two judgments have set the ball rolling as to the questions of the ANN and their patentability. These judgements are relevant in the Indian context for the following reasons: judgements arising out of the UK and EPO have significant pervasive value before the Indian courts (See, for instance, Ferid Allani, where the Delhi High Court has an express statement on the similarity between Article 52 and Section 3(k) or see Ericsson vs Lava wherein the test of further technical effect was stated which itself was propounded in T 1173/97. Likewise, most cases interpreting Section 3(k) tend to mention the UK case laws on computer programmes and technical effect) hence, these differing attempts to classify neural networks as a mathematical method or a computer programme gain significance in the Indian context. Therefore, in Part I, I will first define what ANNs are and explain how they function. Secondly, I will discuss the two foreign rulings and show how these judgments dealt with ANN. In part II, I will explain the possible paths of interpretation available to the Indian courts in the context of ANN and that classification of ANN as a mathematical method or a computer programme might seem significant to the extent that classifying ANNs as a mathematical method would mean it would not be patentable as there is an absolute bar. But I wish to argue that the recent court ruling in Ericsson vs Lava might prevent this conclusion, thus paving the way for extending the per se exception to all judicial classes in section 3(k), including ANNs.
Artificial Neural Networks Defined
The Board of Appeal and the UK Court of Appeal understood ANNs as “artificial neurons are akin to the neurons in the brain”. In a sense it is accurate that they were designed to mimic the functioning of neurons in our brain.
But for a more technical definition suitable for our analysis, one can refer to the succinct explanation provided by Ronald T. Kneusel in How AI Works, who defines the functioning of ANN as follows:
- Multiply every input value, x0, x1, and x2, by its associated weights
- Add all the products from Step 1 together along with the bias value, b. This produces a single number.
- Give the single number to h, the activation function, to produce the output, also a single number.
That’s all a neuron does: it multiplies its inputs by the weights, sums the products, adds the bias value, and passes that total to the activation function to produce the output.
Now, when these nodes are combined and connected with each other in multiple layers (with an input, output and possibly multiple hidden layers), as provided in the diagram below, these nodes would create your average neural network. These machine learning systems, which are based on neural networks, can be employed for a variety of purposes, from recommending songs to identifying irregular heartbeats and classifying images, etc.
Mitsubishi and Treating ANNs as a Mathematical Method Implemented on a Computer
With the technical definition done away with, one can move on to the case laws. The Board of Appeal in T0702/20 was deciding an appeal from the examination division, which had refused the grant of the patent in question under Article 56 of EPC as lacking an inventive step (the technical character analysis of an invention in the EPO happens in a twofold manner “Two hurdle approach” unlike the case in India wherein the technicality analysis is restricted and ends with Section 3(k)). The Board summarised the invention as a new neural network design that sought to address a common issue called overfitting in training machine learning systems by reducing the number of connections between the nodes and a neural network design that performs fewer computations and provide increased performance.
The examining division has rejected the patent for lack of technical purpose and specific technical implementation. In the appeal, the applicant had counter-argued by stating that the invention had a technical purpose and laid emphasis on the new network design with less connection between the nodes, which sought to address the issue of overfitting and improve performance, and it was also argued that the invention had a specific technical implementation. The Board of Appeal rejected both arguments and stated that the independent claim is too abstract and does not have a limiting technical purpose other than increasing efficiency or computational performance.
But the most important observation for our purpose is the observation by the Board of Appeal that “the claim as a whole specifies abstract computer-implemented mathematical operations on unspecified data, namely that of defining a class of approximating functions (the network with its structure), solving a (complex) system of (non-linear) equations to obtain the parameters of the functions (the learning of the weights), and using it to compute outputs for new inputs. Its subject matter cannot be said to solve any technical problem, and thus it does not go beyond a mathematical method, in the sense of Article 52(2) EPC, implemented on a computer.”
Emotional Perception and Weights as Computer Programmes
In Comptroller-General of Patents, Designs, and Trade Marks v Emotional Perception AI, the invention involved “a system for providing media file recommendations to a user”. The invention used two separate machine learning systems (1 and 2) with ANN, which generate vectors in a semantic and property-embedded space. In other words, ML system 1 trains on data provided for every song by the user, describing them as happy or sad (semantic space). Meanwhile, the ML system 2 trains on inputs relating to the physical properties of the songs, such as tone, timbre, speed, and loudness (physical space). As explained in the judgment, the trick of the invention involved “EPL ANN is then trained to make the distances between pairs of the property co-ordinates converge or diverge in alignment with the distancing between that pair in the semantic space.” The purpose is to make the system learn “how to discern semantic (dis)similarity from physical properties.” The entire purpose being correcting physical dissimilarities and distances in physical space in songs with reference to the user classification of the song (semantic space).
The UK Patent Office rejected the patent as lacking technical character, which the High Court overturned, and the issue went up to the Court of Appeal. There were four grounds for appeal, and the central question was how to tackle ANNs in the context of section 1(2)(c) of the UK Patent Act.
The Court first started with the definition of a computer. It concluded that a computer is a device that processes information, and a computer program can be defined as a “sequence of instructions.” Then the Court applied this definition to Artificial neural networks, stating, “Turning to an ANN, the first point to make is that however it is implemented, such a machine is clearly a computer – it is a machine for processing information. Focussing on the weights of an ANN, in my judgment, irrespective of the manner in which an ANN is implemented (hardware or software), the Comptroller is right that these weights are a computer program.” The Court also rejected the idea that a computer program should look in the form of a series of if-then statements and that ANN is a computer program, albeit a different kind. Later, the Court ruled that the invention had any form of technical character due to the nature of the data it processed (semantic data, not technical) and its subsequent contribution.
Thus, to summarise, the following points from both judgments emerge as to the nature of ANN in the legal context:
- Mitsubishi concluded that ANNs are a set of mathematical operations implemented in a computer.
- Emotional Perception concluded that ANNs are computer program but albeit a different kind, and the weights and biases act as program instructions and that it is not necessary that a computer program must be structured in an ‘if-then’ type statements.
Thus, both the courts have arrived at slightly different conclusions about the nature of ANNs. In Part II of this post, I will explain how these issues might play out in the Indian context.