Drafting AI patent applications for success at the EPO – drafting the full specification
In this co-published article, Haseltine Lake Kempner’s Kimberley Bayliss takes a practical look at drafting the specifications of different types of machine learning inventions in view of the EPO’s patent eligibility requirements
As discussed in our earlier article, there are, broadly speaking, three types of machine learning (ML) inventions: applied-AI, core-AI and hardware inventions. In this article we look at drafting considerations for each in turn.
Applied-AI: inventions that are patentable by virtue of their “application to a field of technology”
These inventions are where AI is applied to a specific problem, such as: robotic surgery, self-driving vehicles or medical image diagnosis. The independent claims usually make reference to the field; for example, a claim might be directed to “A method in an automated vehicle”. If the field of technology is a field deemed technical by the EPO, then the EPO is likely to consider the invention to constitute patentable subject matter.
Tip 1: Describe models in terms of parameters that are easy to detect for infringement purposes
The model used in the method is likely to appear in the independent claims; for example, the independent claims might make reference to “a model trained using a machine learning process”, or a “machine learning model”, depending on how the description is set up.
When making reference to a model in such claims, consider what is readily detectable; for example, you may want to describe the model in terms of its inputs and outputs, the type of model used, or the properties of the training data that would have been used to train it. Generally, this is preferable to making reference to how the model might be making its predictions or classifications.
Not only is the internal functioning of the model difficult to detect from an infringement perspective, but how can you be sure that your model is actually functioning in the manner that you think? You may think that a model is classifying images as containing people based on anatomical features, but given the black-box nature of many AI models, can you be sure?
Tip 2: Consider adding experimental data to your applications
From a sufficiency perspective, the EPO may currently be willing to give you the benefit of the doubt in terms of sufficiency if there is a clear causal connection between the inputs and outputs of the model that you are describing.
That said, we are seeing increased interest and conversation around whether experimental data should be submittable (or even a requirement) for AI applications.
The UK Intellectual Property Office recently held a roundtable discussion on the merits of data repositories for AI inventions and we have even seen at least one EPO examiner remark in an examination report that: “It needs to be demonstrated that the prediction of the model is improved by adding such potential parameters or calculation steps.”
Thus, there is increasing interest in more detailed examples and supporting evidence to demonstrate that AI inventions produce the effects that they purport.
Until the case-law in this area settles, it is therefore advisable to provide as much detail as possible, including at least one (preferably more) detailed example in the specification.
For inventions using supervised learning, the application could contain an example, including:
- model details such as model architecture with hyper-parameters;
- pseudo code showing how the model might be trained or used;
- examples of model inputs and outputs; and
- a few lines of example training data.
For reinforcement learning, a detailed description of possible state information, actions and reward schemes might be provided, as well as pseudocode detailing the training procedure.
Furthermore, consider adding experimental data to your detailed examples, if available. Ideally, this should be included on filing, but experimental data can also be submitted in prosecution to support an advantage or effect described in the description.
Therefore, it is good practice to ensure the specification makes reference to the causal connections between proposed inputs and outputs, and indicates advantages and effects associated with each embodiment described.
Tip 3: Beware of generating lists
Inventors will often say they want to cover “any type of model” and provide a wide range of example input and output data types - and this potentially creates several lists. Are the proposed inputs and outputs compatible with every proposed model type? Be sure to add some specific example combinations.
Hardware: Inventions that are patentable by virtue of “being adapted to a specific technical implementation”
Turning to hardware inventions, these are inventions where the AI is adapted to a specific technical implementation. Good examples of this type of invention are algorithms in the emerging field of TinyML where the AI algorithms are adapted for IoT devices with extremely limited processing power, low-memory and often low connectivity. The EPO tends to consider these inventions as constituting patentable subject matter.
Many of the same considerations apply when drafting hardware inventions as for the applied-AI inventions above.
Tip 4: Frame the technical problem in terms of the hardware limitations that the invention solves
Clearly describe the hardware that the algorithm is designed for – what are the hardware limitations that the AI invention improves upon? Consider setting up the technical problem in terms of why prior art processes either can’t be orchestrated on specific hardware or suffer from disadvantages that make them commercially or technically non-viable (eg, too slow for real time deployment).
Tip 5: Link method steps to specific hardware
Note that (perhaps surprisingly) case law in this area indicates that the hardware doesn’t have to be specified in the independent claims. Nonetheless, in dependent claims or the description, consider linking the specific steps of the method to the hardware elements that execute them, or for which they are designed (eg, this step on a GPU, this step on a CPU, or this step by a constrained node, this step by a centralised function).
Again, experimental data can be provided, if possible, that demonstrates an improvement over what has gone before.
Experimental data can be particularly useful in terms of demonstrating an improvement in speed or accuracy over the prior art. For an invention that enables performance of a task on completely new hardware, or hardware configurations, experimental data may provide evidence that the claimed task performance is possible.
Core-AI: Improved AI algorithms and processes
This group of inventions represent the more fundamental AI inventions, including improved models, improved methods of training and improved methods of data processing.
Core-AI inventions tend to be applicable to more than one field and more than one problem. As such, the EPO tends to consider them as mathematical methods and is more likely to raise unpatentable subject matter objections. Core-AI applications, therefore, require more careful consideration at the drafting stage to ensure that the specification contains details that can be used to obtain commercially useful protection.
Tip 6: Add use cases that describe the application of the Core-AI invention to the applicant’s most commercially important AI models
Patent applications in this area should start with broad claims that essentially amount to a mathematical method, particularly if the application will be prosecuted internationally in jurisdictions that may have different requirements to the EPO. The description should then be carefully graded with support for claims of different breadth and different levels of “technicality” according to the EPO’s requirements.
Don’t forget to add in the description (if not in the claims themselves) that the methods are “computer implemented”. Describing an invention in this way ensures that the invention meets the EPO’s first criteria for patentability and ensures that claims aren’t rejected out of hand as being a mere mental act or abstract mathematical method.
One way of circumnavigating a mathematical method objection is to add a field restriction that limits the mathematical method to its application to a particular technical problem. This effectively converts the core-AI invention to an applied-AI invention as described above. Thus, the description of a core-AI invention should contain different use cases that demonstrate how the invention can be applied to different practical problems.
As the applicant may be required to limit their claims to one of the use cases in order to confer patentability, the use case cases should be chosen with care and should describe how the core-AI invention can be applied to the applicant’s most commercially important models. In this sense, particularly where the applicant is a large corporate entity, the use cases should be selected by the business (eg, rather than the individual inventors of the core-AI invention), as it is the business that is best placed to select use cases in an appropriately strategic manner.
From a patenting-perspective, businesses may therefore wish to create standard text describing their most commercially important models that can easily be adapted and added into all core-AI applications. In this way, if a field restriction is required during prosecution, the amended claims will at least protect the applicant’s key products.
To summarise, the points above can be used to assist in drafting ML inventions suitable for submission to the EPO and can be used to provide the applicant with the best possible chance of obtaining a commercially useful patent, whatever the invention type.
This is the third co-published article by Haseltine Lake Kempner in the series on Patenting AI before the EPO. The first two can be viewed here: