brAIn blog: patenting machine learning innovation

What, why, how, where, when?

One of the most persistent myths that I hear from entrepreneurs, innovators and investors is that "you can't patent software". I've spent years trying to dispel this myth in talks and articles, as have other patent attorneys specialising in software, but it remains! So, in this blog post, I discuss what you can patent when it comes to artificial intelligence and machine learning, and why you might want to patent these software-based innovations. This blog post is aimed at:

• innovators, scientists and academics in AI/ML - this post will hopefully empower you to determine whether your AI and ML innovations are patentable too;
• start-ups, spin-outs and scale-ups in AI/ML - this post will help you to strengthen your IP portfolio, build your intangible assets, and attract investors; and
• service providers (accountants, lawyers, solicitors) - this post will help you to point your clients who are working in the AI/ML space to patent attorneys like me who can advise on whether their innovation is patentable.

What AI/ML research can you patent?

The reason why the myth mentioned above exists is because many countries' patent laws specify a list of things that you cannot, generally speaking, protect via patents. This list usually specifies, among other things, that "mathematical methods" and "programs for computers" cannot be patented as such. It is worth bearing in mind that these laws were written a long time ago, before anyone really envisaged software becoming so ubiquitous.

Mathematical methods

Maths underpins many AI and ML ideas. Therefore, it might seem that you cannot patent AI and ML but, the situation is more complex in practice! The patent laws of most jurisdictions say that if the claims of a patent (i.e. the part of a patent which defines the scope of protection you are seeking for your invention) are directed to a purely abstract mathematical method, then the claim is not technical. Inventions are considered to be things which have "technical character", i.e. they produce a technical effect that serves a technical purpose or solves a technical problem. A painting, for example, is not considered to have technical character. So, if we tried to claim a method for performing a Fast Fourier Transform on some data, that would be considered non-technical and not patentable. Similarly, if a mathematical method is algorithmically more efficient than existing mathematical methods, that is not considered enough to establish a technical effect. However, if we instead claimed the application of some new maths to a particular field of technology, that would be considered to have technical character. Examples of this include mathematical methods for performing audio, image or video enhancement or analysis (e.g. detecting people in a digital image or removing noise), and mathematical methods for performing speech recognition or encrypting/decrypting signals.

The same is true for AI and ML innovation, which are based on computational models and algorithms that are typically mathematical in nature, even if they have been trained on real data. But AI and ML is used in a wide variety of fields, and the uses of AI/ML may make the innovation "technical". Examples of this include the use of neural networks for identifying tumors in medical images, for classifying digital images, and for performing speech analysis.

Computer programs

With respect to the computer program exclusion, you might think that the fact your AI or ML model runs on a computer would make it technical. However, computer-implemented methods must produce a "further technical effect", which goes beyond the normal physical interactions between the software and the hardware. There are lots of ways we can show that an AI/ML invention has a further technical effect. Examples of computer-implemented ML models that have a further technical effect are those used to control an anti-lock braking system in a car, to compress videos, to restore distorted digital images, to encrypt electronic communications, to perform processor load balancing or memory allocation, and so on.

You can hopefully see now that AI and ML innovations which can be tied to a real-world application are more likely to have a technical effect. Similarly, new maths, even if it results in faster processing or more accurate processing, is not on its own enough to generate a technical effect.

What if your AI/ML innovation has no technical effect?

In this case, while the patent route may be closed to you, you can still protect your innovation using other forms of intellectual property rights. For example, you could protect your innovation using trade secrets or you could rely on copyright to protect your code. You can find some more information on this here and here. You can also get in touch!

Why patent your AI/ML research?

If you are an entrepreneur or company, there are lots of reasons why you might want to patent your AI/ML research. Patenting an invention can help you to secure market position and financial return on commercially successful innovations, obtain greater negotiating power when collaborating with other companies, attract investment partners, and support business expansion. The decision about whether to patent any research or innovation, including software, is a commercial one. A patent gives a patent holder a monopoly right to stop other companies from exploiting their invention without permission. However, in exchange for this right, you have to reveal the details of your invention to the world, as patents are published and available for anyone to read. Patent law around the world specifies the level of detail you need to include in a patent application in exchange for the monopoly right, and for AI/ML inventions, there is a move in some countries towards including more and more detail. You need to decide whether you are happy to reveal the details of your invention, bearing in mind that when you launch your AI-based product or service, some parts of your invention may be disclosed to the public anyway.

You can patent both the training phase and the inference phase of a ML model. You can patent new architectures and ways to augment training data. It is important to understand that while a patent gives you the right to stop others from copying your invention without your permission, before you can enforce your patent you need to be able to determine if someone has copied the invention. Many mechanical, electrical and software products can be reverse-engineered or inspected without too much difficulty and this can enable you to determine if someone has copied your patented invention. However, it is more difficult with AI/ML. The inference phase of an ML model may happen on-device, such as in an end-user product like a smartphone, and it may happen many times. So, it may be possible to detect whether the inference phase of a third party's ML model is the same as your patented inference method. However, sometimes, the training phase is the clever part of the invention, and while having a patent for this might help you secure investment and build your IP porfolio, it may not be very useful when it comes to catching infringers (i.e. people who are using your invention without your permission). This is because it is difficult to determine whether someone else's model has been trained in the same way, when you can only inspect or see the trained model that is used for inference. Of course, if your training performs on-device rather than in the cloud, then it may be possible to detect if someone is infringing your patent.

Therefore, there are lots of reasons for patenting your AL/ML invention, and which aspects of your invention we patent will depend on these reasons.

How do you patent your AI/ML research?

For all AI/ML patent applications, the patent application needs to sufficiently describe the invention such that someone else working in the field could pick up your patent application and make your invention themselves. You don't need to include your code, but you do need to explain how the model is trained, and what training data is used for the training. It is not sufficient to say something like "suitable datasets are used for training the ML model". Often, the training is performed using data taken from publicly available datasets like ImageNet and in this case, it is straightforward to explain how the training is performed. However, in some cases, the training is performed using private data or data that cannot be shared for privacy reasons, such as medical data. In this case, we would need to consider whether a patent application for the invention would satisfy the sufficiency of disclosure requirement.

The five largest intellectual property offices, which includes the USPTO and the European Patent Office, issued a statement in 2018 that states that algorithms and training data need to be described in sufficient detail, but there are other areas in AI/ML inventions which may need to be described in detail, such as the details of the model architecture itself (i.e. the specific layers of the neural network), the training process, the inference process, any intermediate processing, how input data is labelled, selected or classified, etc. Generally speaking, we recommend including as much detail as possible in the patent application, particularly as this detail cannot be added into the patent application at a later stage.

Where and when should you patent your AI/ML invention?

Finally, I want to briefly touch on patent filing strategy. As with all inventions, it is imperative to file a patent application for your AI/ML invention before it is disclosed to the public. That means, before launching a product and before submitting a paper on your research to a journal or placing it on the arXiv.

In my view, the EPO and USPTO assess whether AI/ML inventions are patentable (i.e. whether they are directed to patentable subject matter or are excluded from patentability for the reasons discussed above) in very similar ways. US and European patent office examiners both look for whether an AI/ML invention has a technical effect or has technical character. One way these jurisdictions differ is in the sufficiency requirement. Currently at least, the EPO has a higher sufficiency requirement and therefore requires much more information on training, architecture, and so on, as explained above. Therefore, whether or not you file patent applications in Europe and the US will depend partly on how many details of the invention you are happy to reveal. This is something we can explore with you.

Summary

Hopefully, this has helped you to understand what sorts of AI and ML inventions are patentable, and what sort of information is needed to prepare a patent application. If you have developed new AI or ML technology and would like help patenting it, get in touch with me: [email protected]. At Appleyard Lees, we have a dedicated team of patent attorneys who specialise in protecting software and AI inventions, so give us a shout!