DeepMind, a leading artificial intelligence (AI) research company, has filed a series of international patent applications, which have now been published for the first time. The applications relate to a number of the fundamental aspects of modern day machine learning, and are therefore of potential significance to anyone operating in the commercial AI sector.
Background to DeepMind
DeepMind is a London based artificial intelligence (AI) research company, widely recognized as being at the forefront of the field. DeepMind was founded in 2010 and acquired by Google in 2014 for £400m. In 2017, DeepMind famously developed AI capable of defeating a world-champion at Go (Silver et al. Nature).
DeepMind’s approach to AI is explained here by founder Dr Demis Hassabis (the article also serves as a useful introduction for the uninitiated to AI/Ml in general). Dr Hassabis is a popular guy, described as “a superhero of artificial intelligence” and “a genius” by the Guardian (here), a “modern polymath” by the FT and “London’s megamind” by The Evening Standard (here). He has even been on the BBC’s desert island discs (here).
DeepMind’s Patent Applications
Deepmind’s published patent applications so far include:
- WO 2018/048934, “Generating Audio using neural networks”, Priority date: 6 Sep 2016
- WO 2018/048945, “Processing sequences using convolutional neural networks”, Priority date: 6 Sep 2016
- WO 2018064591, “Generating video frames using neural networks”, Priority date: 6 Sep 2016
- WO 2018071392, “Neural networks for selecting actions to be performed by a robotic agent”, Priority date: 10 Oct 2016
- WO 2018/081089, “Processing text sequences using neural networks”, Priority date: 26 Oct 2016
- WO 2018/083532, “Training action selection using neural networks”, Priority date: 3 Nov 2016
- WO 2018/083667, “Reinforcement learning systems”, Priority date: 4 Nov 2016
- WO 2018/083668, “Scene understanding and generation using neural networks”, Priority date: 4 Nov 2016
- WO 2018/083669, “Recurrent neural networks”, Priority date: 4 Nov 2016
- WO 2018083670, “Sequence transduction neural networks”, Priority date: 4 Nov 2016
- WO 2018083671, “Reinforcement learning with auxiliary tasks”, Priority date: 4 Nov 2016
- WO 2018/083672, “Environment navigation using reinforcement learning”, Priority date: 4 Nov 2016.
These applications represent a filing rate of almost 1 application per week in the period of September 2016 – December 2016. Patent applications are published 18 months from the priority date, so any applications filed after Dec 2016 will not yet have been published.
A quick perusal of the applications reveals that the claims generally fulfill promised broadness of the titles, and whilst not being directed to fundamental algorithms, claim ML platforms for solving general problems, without being restricted to a particular application. Claim 1 of WO 2018/048945, for example, relates to the application of convolutional networks (the majority of modern neural networks of significance are convolutional networks) to the processing of any sequences, i.e. any temporal data such as audio, and text. The claims of WO 2018/081089 apply this to text sequences for the purpose of translation. Whilst there are other ways of achieving similar results using alternative AI approaches, the application covers a fundamental class of approaches.
DeepMind is, of course, not the first company to file applications in the AI field. As has been widely reported, AI is a large and rapidly expanding field of patent filings (for more on patenting AI see IPKat posts here and here). However, DeepMind’s applications are significant in view of their position as a (if not the) leading AI research company, and for the broadness of the claims. These are not applications that reveal a clear commercial product or aim, but appear directed at covering broad swathes of AI technology.
Applications vs Patent
Contrary to popular belief on anti-patent blogs, a patent application is not equivalent to a granted, enforceable patent. The DeepMind applications are still at a very early stage of prosecution, and will be examined for patentability in each of the jurisdictions in which DeepMind seeks grant. As might be expected in a relatively new field, the International Search Reports (ISRs) (published for a number of the applications) do not cite reams of prior art. In fact, many of the documents cited for novelty and inventive step appear to be DeepMind’s own arXiv papers. However, DeepMind will have to overcome Examiner objections based on these documents through argument or by limitation of the claims. Furthermore, patenting machine learning systems, whilst possible, is not always a straightforward matter in many jurisdictions (not least in Europe and the US).
“Don’t be evil”
DeepMind researchers publish extensively in academic journals, and the company has the stated aim of supporting and accelerating research in the wider AI community. DeepMind founder Dr Hasabiss believes AI has the potential to “become a kind of meta-solution for scientists to deploy, enhancing our daily lives and allowing us to all work more quickly and effectively. If we can deploy these tools broadly and fairly, fostering an environment in which everyone can participate in and benefit from them, we have the opportunity to enrich and advance humanity as a whole” (FT article). At the PR level, DeepMind thus heavily emphasizes their collaborative attitude and openness. This approach is understandable, given the animosity of the AI academic community towards anything other than open source, and DeepMind’s ongoing need to attract nothing less than the best researchers in the field.
It will be interesting to see how news of DeepMind’s patent filings will go down in the AI community and the public at large; more commercially focused companies than DeepMind have encountered difficulties with public perception in this area. Google is of course well-known as a prolific patent filer, including within the field of AI. By contrast, the raison d’etre of DeepMind has hither-to been presented as the development of AI research for the good of the research community and humanity as a whole. Achieving such an aim may of course be facilitated by acquisition of the appropriate IP protection. But how does DeepMind’s newly revealed patent filings fit with the mission statement that “AI should ultimately belong to the world” (quote source here)? Perhaps a topic of conversation for Dr Hassabis’s next interview.
This article appeared on IPKat on Thursday 7 June 2018.
This article is for general information only. Its content is not a statement of the law on any subject and does not constitute advice. Please contact Reddie & Grose LLP for advice before taking any action in reliance on it.