Brevity versus giving away the crown jewels: getting the balance right for protecting machine learning technology with patents

In a recent decision of the European Patent Office technical board of appeal the board considered whether there was enough detail about a machine learning invention in the patent application documents. Patent application documents are published on the internet and applicants have to be aware of the “patent bargain” whereby in return for disclosing details of innovation an applicant can seek potential monopoly rights over the innovation. 

In the case of machine learning technology it is particularly difficult for applicants to know how much detail needs to be given since there is little case law to use as guidance.  This recent decision is therefore very helpful. 

The technology involved was artificial intelligence technology for estimating waste composition from images, such as images of refuse pits of incinerators in waste processing facilities.

The patent application was about generating training data from waste‑pit images and training a machine learning model. The trained model takes an image of a waste pit and outputs a “value representing composition” of the waste depicted in the image.

The board focussed on four features in the definition of the scope of monopoly:

1. What type of ‘data of’ a captured image is used;
2. The ‘value representing composition’ that is output;
3. The accuracy of the output; and
4. The type, architecture and training of the machine learning model.

All of these were generally undefined and the board interpreted these to cover a wide range of possibilities.

A party attacking the patent referred to two previous decisions regarding sufficiency of machine learning inventions, arguing that they provided guidance for assessing sufficiency of disclosure for machine learning inventions. However, in both these cases the board held that no general criteria relating to sufficiency of inventions in the field of machine learning could be derived. They considered that there is no reason to treat machine learning inventions differently from those of other types of technology, and that the normal requirements apply as for all European patent applications.

One point raised by the board was that the patent did not contain a specific example. To counter this, the patent owner submitted Annex 1, disclosing a very specific example. However, the Board said this was not convincing for two reasons. Firstly, the example didn’t appear in the patent and it was not common general knowledge.

Secondly, requirements for patentability include that the invention can be carried out over the whole claimed breadth without undue burden, so even if there had been ‘one way’ disclosed of carrying out the invention, it would still not be sufficient considering the breadth of monopoly in the patent.

In this case, the board noted that the disclosure is mostly limited to a ‘result to be achieved’ and that the skilled person has the task of designing a model for all possible input and output types. Due to the wide variety of possibilities of each of the four features as listed above, there are a huge number of combinations that they could choose from, and each would need to be tested.

Importantly though, the board held that ‘the undue burden also does not stem from the fact that each evaluation alone requires some effort’ but that the patent doesn’t provide guidance on where to start. Nor does the undue burden stem solely from the breadth of the monopoly, it is that the breadth is disproportionate to the disclosure in the patent.

In this case, the board found that the patent did not sufficiently disclose the invention, and as a result, the patent was revoked.

Conclusion

The board emphasised that, as with the previous decisions referred to by the appellant, this decision does not set out general criteria for sufficiency of inventions in the field of machine learning. The board again stresses that machine learning inventions are not a special category, and the normal principles apply.

We note that the patent document did not have a description of the model architecture. The training was mentioned to be supervised training but there were no detailed training algorithms described. There were a few examples of training data images but these were not explained as being raw images or average properties.

The decision should not be read as insisting that every machine learning‑related patent must describe all conceivable variables. Providing clarity on just one of the relevant aspects, such as input data or output requirements, may be sufficient to guide the skilled person. As the board observed, ‘clearly defined input and output data might allow the skilled person to deduce the necessary properties of a suitably trained machine learning system.’

The decision also makes clear that a patent does not need to spell out every possible variable in a machine‑learning system. What matters is that the disclosure provides enough guidance for the skilled person. In some cases, defining a single, concrete combination of input data, output data and model may be sufficient to enable the invention across the full claim breadth.