IP-Managers in large international companies often find that the grant rates of their patents are significantly higher in their own jurisdiction than in other jurisdictions or that the obtained scope of protection is comparatively narrow in the other jurisdictions. One of the reasons for this effect is that the patent applications are drafted taking into account the requirements of only one jurisdiction.

Drafting patent applications is a challenging task anyway, but it becomes even more difficult if the requirements of different jurisdictions should be considered. A drafting style working for one jurisdiction may cause major problems in other jurisdictions. Many problems could be avoided if they were considered at the drafting stage.

I will show periodically examples of problems caused by drafting or prosecution weaknesses resulting in rejections or narrow claims. The examples are taken from real cases, with the original wording changed to focus only on the key issue.

Today´s example relates to: Mixing Application and Training of AI

Example claim:

“A computer-implemented method for estimating vegetation coverage …:

receiving an RGB image of a real-world scenery…;

obtaining an estimate of a near-infrared channel by applying a first neural network trained to estimate a near-infrared channel from the RGB image,…;

providing a multi-channel image comprising a channel of the RGB image and the estimated near-infrared channel as test input to a second neural network trained with a training data set comprising multi-channel images of the test input type,…; and

segmenting the test input using the second neural network… resulting in a vegetation coverage map.”

Problem:

The provided example claim tries to claim the application of neural networks to estimate vegetation coverage. On the other hand, the claim defines aspects of the training.

For the first neural network, it is defined that it is trained to estimate a near-infrared channel from the RGB image. Although this definition is unnecessary, it would still be possible to read this feature on an infringing neural network as the infringing neural network would estimate a near-infrared channel and would have been necessarily trained to do so.

In contrast, the claim defines that the second neural network is trained with a training data set comprising multi-channel images of the test input type. This definition leads to the problem that one usually does not know how an AI algorithm was trained when looking at the application of the AI algorithm. Consequently, it would be difficult to prove that a competitor infringes this claim.

So the question is: How can such a situation be avoided?

Solution:

For the first neural network, it would be sufficient to define that the first neural network estimates the near-infrared channel. It is redundant to additionally define that the neural network has been trained for this purpose.

More important, it would be sufficient to define that the second neural network segments a test input comprising a channel of the RGB image and the estimated near-infrared channel resulting in a vegetation coverage map… or even better: “determining a vegetation coverage map by segmenting a test input by a second neural network, wherein the test input comprises a channel of the RGB image and the estimated near-infrared channel”.

If the inventive aspect (or one of the inventive aspects) relates actually to the training of the second neural network with a training data set comprising multi-channel images of the test input type, a separate claim should be directed to the training of the AI algorithm.

Summary:

Make sure that you clearly identify where the invention takes place. Distinguish between features relating to the AI/ML algorithm itself, to the training of the AI/ML algorithm and to the application of the AI/ML algorithm. Do not mix features of different categories in one claim.

In order to achieve high grant rates and broad scope of protection in all relevant jurisdictions, IP-Managers should take care that the requirements of all these jurisdictions are considered by their drafting attorneys at the drafting stage.

Drafting guidelines are a good way of ensuring high-quality applications that are suitable for all jurisdictions relevant to a company.

About the blogpost author:

Stefan Müller is partner at 2SPL PartG mbB since 2012 and an expert for the protection of digital solutions in Europe. Before becoming a patent attorney, he studied physics and gained practical experience in the semiconductor industry.

Feel free to connect with him on LinkedIn.