Reading about an app using machine learning to recognise lemurs led me to imagine the conversation I might have with a client using machine learning to recognise some animal, perhaps cats. Whilst there are no specific provision of patent law directed at machine learning (ML), the particular characteristics of ML do present specific issues when trying to obtain intellectual property rights to protect developments in this field. The following strings together elements of many similar conversations I have had with clients in this field.
Inventor: We’ve got this great system based on machine learning for recognising cats – can we get a patent?
Me: That sounds fascinating. Given that ML is well known to be used to recognise all sorts of things, we couldn’t get a patent on the broad concept of using ML to recognise cats. Frankly it’s hard to sell “applying ML to X” as inventive at all. However, there may well be something we can protect in the way your system achieves good results – tell me more.
Inventor: Well for a start we’ve got a great set of training data.
Me: [sotto voice] Doesn’t everyone?
Me: Nothing. It wouldn’t be possible to patent the data itself, but you may have copyright or database right in it. Unfortunately that doesn’t prevent a competitor collecting their own comparable set of data. However, it may be that there is something protectable in the process of collecting or treating the data. How was that done?
Inventor: Well we needed images from all angles so we developed a technique of taking video and using a laser pointer to make the cats look in the right directions. Then we selected the best frames.
Me: Sounds like fun but I think there’s lots of prior art on YouTube. Was the selection of frames automatic? If so there might be a patentable invention there.
Inventor: No we had to do it manually. But then we applied a special mathematical filter to emphasise the distinctive features of the cats.
Me: That sounds promising. Although we can’t patent the mathematical formula of the filter itself we can protect its use in processing images and more specifically its use in processing the input to a ML algorithm.
Inventor: But I thought that mathematical methods aren’t patentable?
Me: Mathematical methods in the abstract aren’t patentable but they are if you limit to a specific technical application of the mathematical method, such as processing images. The EPO (European Patent Office) has been happy with this since a case in 1986 (Vicom) about processing satellite images. It’s a little less clear in the US since the Supreme Court case Alice v CLS Bank – now you have to make sure you are not trying to monopolise all uses of the method and claim “substantially more” than just “do it on a computer”.
Inventor: Can we patent the trained ML system itself?
Me: To get a patent we have to set out in a claim what is different from the prior art and that defines the scope of the protection you’d get. If we try and define the ML system by the data that’s been used to train it or by the coefficients that result this would provide only uselessly narrow protection. We need to identify some general feature of the ML system that is new, inventive and might be useful to others. How about the structure of the ML algorithm?
Inventor: It’s based on a convolutional neural network with max pooling but we have some unusual structures of memory cells and back propagation.
Me: CNN with pooling is fairly conventional for image recognition isn’t it? But I don’t know about memory cells and back propagation. Perhaps we should do some searching to see if it’s new?
Inventor: I have a grad student who can do a literature search but maybe you can cover patent documents better?
Me: Certainly, I can arrange that. What about the output stage and the commercial applications of the application?
Inventor: The output is standard: a closest match or a ranked list of candidates. We have several possible applications:
- a facial recognition cat flap that only lets your own cat in and not other pesky neighbourhood mogs
- a selective cat feeder that only allows access for a particular cat, enabling different diets to be fed in a multi-cat household (fat cat/thin cat, young cat/old cat)
- integration with existing pet passport schemes (allowing international travel without quarantine) to give a kitty biometric-type pet passport.
Me: Those sound promising. It’s good to have some potential products to show the invention can provide some useful output, not just an abstract decision. Why don’t you send me whatever documentation you’ve got and we’ll set up another meeting to go through it in detail before we start work on a draft patent.
Inventor: I’ve got a draft paper I can send you and a demo version you can play with.
Written by John Leeming, Partner, J A Kemp
John is a Partner and Patent Attorney at J A Kemp with specialisms in semiconductors, software, optics and designs. With over 25 years of professional experience, John’s clients range from large multinationals to start-ups, including firms that have successfully progressed beyond the early stages to develop extensive patent filing programmes. He is the primary contact at J A Kemp for UK and Community Registered Design matters.
Additional research by Elizabeth Slack
(The machine learning app used to recognise lemurs as mentioned above.)
If you are interested in writing for us, please email email@example.com for more details.