The UK Intellectual Property Office (UKIPO) has published guidance for examining patent applications relating to artificial intelligence (AI) inventions. UKIPO has confirmed that patents can be granted for AI inventions, as they make a technical contribution to the state of the art.
After a consultation period that ran from September 7 to November 30, 2020, the new guidance note details the requirements for AI technologies to meet patentability criteria. Since computer programs are specifically excluded from patentability criteria, the new guidance note and accompanying scenarios clarify attempts to patent AI-based technologies.
UKIPO defines AI as:
“Technologies capable of performing tasks that would otherwise require human intelligence, such as visual perception, speech recognition, and language translation.”
In summary, the new guidance note states that:
- AI patents are available for all areas of technology.
- While mathematical methods or computer programs are excluded from patent protection, where the task or process performed by an AI invention contains a technical contribution, it is not excluded.
- An AI invention is likely to make a technical contribution if it:
- performs or controls an existing technical process outside the computer;
- contributes to the solution of a technical problem, external to the computer;
- solves a technical problem in the computer itself; Where
- defines a new way of technically operating a computer.
- AI inventions are not excluded if they are claimed in material form only (i.e. they are not based on program instructions or a programmable device).
- AI inventions are likely to be excluded from patentability if they relate to an excluded subject matter, relate only to data processing, or are a general improvement to a conventional program or computer.
When considering the patentability of AI technology, the focus is therefore on the technical contribution that the invention makes to the state of the art.
Examples of scenarios
UKIPO has also published a series of scenarios regarding AI or machine learning (ML) technologies and their compliance with patentability criteria. These scenarios focus on the question of excluded subject matter and cover a range of fields and technologies with concrete examples of why each invention is or is not excluded from patentability.
Particularly interesting scenarios include the training of a neural network, where the end result of the process and its intended use may be the deciding factor in whether or not to exclude patentability.
For example, training a neural network classification system to detect cavitation in a pump system is allowed. One such method involves correlating data pairs with class values to produce a training data set (where each class value indicates an extent of cavitation in the pump system) and then training the classification system neural network, using the training dataset and backpropagation.
The fact that this method is based on a computer program does not exclude it from patentability, since it makes a contribution which uses physical data to train a classifier for a technical purpose, namely the detection of cavitation in a pumping. The end result of this training, and its contribution, is therefore technical in nature.
On the other hand, active training of a neural network is not allowed. One such process involves determining areas of weakness in the neural network by comparing the confidence levels to a threshold and then augmenting the training data with data related to the area of weakness. For example, a neural network used to detect animals in images may have trouble identifying cats, so data from specimens may be supplemented with additional images of cats. This is more efficient than just expanding the dataset to all elements.
While this method may result in a more efficient learning process for a neural network, it does not itself produce a neural network that itself operates more effectively or efficiently. The mere identification of specific additional training data cannot be considered to be related to a technical problem. Thus, no technical problem has been solved within the neural network, and no technical effect is produced. A claim directed to this subject would therefore be excluded as a program for a computer as such.