Examination of AI-related patent applications in Singapore

Following a slew of updates to examination guidelines across various jurisdictions over the last year, the Intellectual Property Office of Singapore (IPOS) has recently issued a Supplemental Guidance for the Examination of Artificial Intelligence (AI)-related Patent Applications in October 2024. Particularly, the Supplemental Guidance provides guidelines in relation to assessing subject matter patentability of AI-related patent applications in Singapore.

Differing from other jurisdictions such as Europe and China, there is no definition of excluded patentable subject matter in the present Singapore Patents Act (SPA). Examination of subject matter patentability of Singapore patent applications therefore relies on examination guidelines maintained by the IPOS. Notably, however, the examination guidelines have maintained that subject matter relating to (i) discoveries, (ii) scientific theories and mathematical methods, (iii) aesthetic creations, (iv) schemes, rules or methods for performing a mental act, playing a game or doing business and (v) presentation of information are not considered to be an invention, in order to adhere to international patent norms and practices in major jurisdictions. Therefore, despite the removal of the definition of non-patentable subject matter in the SPA, subject matter patentability is still relevant in prosecuting Singapore patent applications. The present Supplemental Guidance sets out the test for subject matter patentability for AI-related inventions.  

The test for subject matter patentability for AI-related inventions is set out as follows:

1. Properly construing the claim;
2. Identifying the actual contribution; and
3. Asking if the actual contribution falls solely within subject matter that is not patentable.

If the actual contribution is found to fall solely within non-patentable subject matter, the claim is not considered to define an invention.

Step (i) involves construing the claim. Singapore adopts a purposive approach and claims should be construed in order to determine what the person skilled in the art would have understood the patentee to mean by using the language of the claims. 

Step (ii) involves identifying the actual contribution by considering the problem to be solved, how the invention works, what its advantages are. It is the substance of the claim and not the form of the claim which should be considered. Therefore, a computer programmed to perform a task, or the computer program itself, which makes a contribution to the art which is technical in nature, is patentable and may be claimed as such. 

Step (iii) involves asking whether the actual contribution falls solely within non-patentable subject matter. If the actual contribution of an AI-related patent claim solves a specific (as opposed to a generic) problem, then it is likely not solely a mathematical method. Conversely, if the actual contribution of an AI-related patent claim is, for example, an AI algorithm which does not appear to solve a specific problem, it is likely to be solely a mathematical method and not an invention. The guidelines further emphasise that the mere fact that the mathematical method may solve a specific problem is unlikely to be sufficient. The claimed method should be functionally limited to solve the problem by establishing a sufficient link between the problem and the steps of the mathematical method by clearly specifying how the input and the output of the sequence of mathematical steps relate to the problem so that the mathematical method is causally linked to solve said problem.

Further, the mere determination that a specific, rather than a generic, problem is being solved is likely not enough since the specific problem could itself be, for example, a pure business method. It is therefore necessary to consider if the actual contribution falls solely within non-patentable subject matter in determining if the AI-related patent claim relates to patentable subject matter. For example, if an actual contribution of a patent claim is determined to be an AI algorithm for dynamically determining the level of compensation for members occupying each tier of a multi-level marketing scheme, while the AI algorithm is arguably directed to solving a specific problem, the patent claim does not define an invention because the specific problem being solved falls into the category of a pure business method. 

Selected examples provided in the Supplemental Guidance

Example (A): Optimising the size of a trained neural network with respect to a processor

Background

Trained neural networks may not be optimised for the processors they run on resulting in longer than expected computation times. This may be a disadvantage in scenarios where timing is critical. For example, an autonomous vehicle using image recognition to detect potential road hazards must be able to do so in a timely manner so that an appropriate action, such as bringing the vehicle to a stop, may be taken to prevent an accident. The applicant has devised a method to speed up execution times by reducing the size of a trained neural network. This is done by selectively removing nodes from the neural network that are less heavily weighted so that the total computation time for the task may be minimised.

Claim

A server that reduces the number of arithmetic operations of a neural network performing a computation task on a processor, the server comprising:

• a determination unit that determines the number of the arithmetic calculators in the processor;
• a setting unit that sets the number of arithmetic operations required to be equal to the number of arithmetic calculators in the processor; and
• a reduction unit that reduces the number of nodes of the neural network; and
• wherein the number of arithmetic operations in the reduced neural network is equal to the number of arithmetic operations set by the setting unit and the neural network nodes are removed in order of their weight; and
• executing the reduced neural network to perform the computational task on the processor.

Analysis 

i) Construe the claim 

The claim defines a server comprising several types of units with the specified functions. The claim does not further define if these units are in fact physical hardware or computer programs and so the server may simply be executing a computer program or a set of computer programs. A processor is also defined in the claim. However, the claim does not limit this processor to being part of the server i.e. the processor could be part of a different device or computer. The computational task to be executed by the reduced neural network on the processor is not further defined. There appears to be no other issues in construing the claim. 

ii) Identify the actual contribution 

The actual contribution is removing nodes of a neural network in order of their weight so that the number of arithmetic operations to be performed by the reduced neural network is equal to the number of arithmetic calculators in a processor. 

iii) Does the actual contribution fall solely within non-patentable subject matter? 

The actual contribution comprises an algorithm for removing nodes from a neural network to perform a computational task on a processor which is a specific problem. This specific problem is functionally limited to the steps of the method as seen at least in the determination of the number of arithmetic calculators in the processor and the execution of the reduced neural network on the processor. The actual contribution is thus not solely a mathematical method and therefore the claim defines an invention.

Example (B): Method of extending a neural network

Background

As the size of a neural network increases, its performance may also increase, however, this is usually also accompanied by increasing complexity when training the neural network. The applicant has devised a method that progressively extends a trained neural network by adding nodes to hidden layers of the network and then trains the structure-extended neural network. The method allows the neural network to be progressively trained while carrying over the learning obtained before the neural network is extended.

Claim

A computer-implemented method of extending a trained neural network, the method comprising:

• storing the trained neural network in a memory;
• selecting, with a processor, a node in a hidden layer of the neural network;
• adding a new node in the hidden layer that includes the selected node; and
• connecting the new node to nodes included in a layer preceding and subsequent to the hidden layer;
• setting weights of the new node; and
• adjusting the weights of the new node by training the neural network extended to include the new node.

Analysis 

i) Construe the claim 

The claim defines physical components i.e. a memory and a processor that would be associated with an ordinary computer running a computer program. There appears to be no issues in construing the claim. 

ii) Identify the actual contribution

The actual contribution is an algorithm to extend a trained neural network by adding a node and its weights to a hidden layer of the neural network. 

iii) Does the actual contribution fall solely within non-patentable subject matter? 

The actual contribution is an algorithm for extending a trained neural network and does not appear to be directed towards solving a specific problem. The claimed method may be applied to any field of technology that uses a neural network and thus the problem appears to be a generic one. Functionally, the tasks the computer in the claim performs, i.e. storing in memory and running a computer program appear to be generic in nature and so the computer in the claim is not considered to interact with the method to a material extent and adds nothing more to the actual contribution. The actual contribution is thus solely a mathematical method and therefore the claim does not define an invention.

Concluding remarks

As illustrated in the foregoing, in assessing whether an AI-related patent claim falls within subject matter that is not patentable, it is important to determine if the actual contribution by the patent claim is directed to solving a specific problem and whether the specific problem itself fall solely into an excluded category such as a mathematical method or a business method. For example, if the patent claim relates to a mathematical method, it is critical that the claimed method should be functionally limited to solve a problem by establishing a sufficient causal link between the problem and the steps of the mathematical method.

In contrast with the practice of IPOS, the exclusions from patentability under the practice of the European Patent Office (EPO) play a role in assessing both patent eligibility and inventive step in a “two-hurdle approach”. In the first patent eligibility hurdle, it requires that the claimed subject-matter as a whole must not fall under the "non-inventions" as defined in Art. 52(2) and (3) “as such”. This has a low bar: if the claimed subject-matter is directed to or uses technical means, for example a computer, it is an invention within the meaning of Art. 52(1). The second hurdle is where inventive step is assessed. Inventive step of claims comprising a mix of technical and non-technical features is assessed using a problem-solution approach that involves establishing which features of the invention contribute to its technical character (i.e. contribute to the technical solution of a technical problem by providing a technical effect). A feature may support the presence of an inventive step if and to the extent that it contributes to the technical character of the invention. For AI and machine-learning (ML) related inventions, the technical character of the invention may be related to whether these inventions serve a “specific technical purpose” (e.g., classification of digital audio, image, or video data) or are directed to a “specific technical implementation” (e.g., a new arrangement of computing hardware). It is noted that the China National Intellectual Property Administration (CNIPA) has a similar approach to examining AI-related inventions as the EPO.

In our opinion, it appears that IPOS may have a more lenient approach to AI-related or ML-related inventions as compared to the EPO or CNIPA. This is because when assessing a patentability of an AI-related or ML-related invention, the Singapore approach is to construe the claim and to identify the actual contribution, then ask if the actual contribution solves a specific problem which does not fall solely within non-patentable subject matter. The Singapore approach does not disregard features which may otherwise be considered non-technical, but involves considering features of the claim as a whole in identifying the actual contribution and whether the actual contribution solves a specific problem. On the other hand, the EPO’s problem-solution approach to inventive step involves first identifying the technical problem by considering whether specific features of the claim are technical in nature, where non-technical features do not contribute to the technical character of the claim. A scenario may therefore arise, e.g. in Main-Line Corporate Holdings Ltd v DBS Bank Ltd [2012] SGHC 147, where a claim relating to determining an operating currency in a payment transaction was considered patentable in Singapore but deemed to lack an inventive step before the EPO because features which are directed to addressing a non-technical problem are disregarded by the EPO when considering inventive step.