Patent filing trends for artificial intelligence (AI) and machine learning (ML) innovations
‘By far the greatest danger of artificial intelligence is that people conclude too early that they understand it.’ ― Eliezer Yudkowsky
With the continued explosion of advances in artificial intelligence (in particular its technological cousin machine learning) and the evermore pervasive use of these approaches in consumer products, it is of little surprise that the rate of patent filing in this field is increasing so rapidly. ClearViewIP has worked with a multitude of clients at the very forefront of new technology. Recently this has included a large number of companies specialising in AI and ML, helping us increase our understanding of the development and investment in this field.
In particular, companies in this field are increasingly interested in the limitations and requirements for patenting their technology. This is reflected across the industry, from large technology giants to startups making their first investments in protecting their AI inventions. As a result, both the European Patent Office (EPO) and the US Patent and Trademark Office (USPTO) have recently issued guidelines that seek to clarify what can and cannot be considered patentable inventions in cases where artificial intelligence is involved. This article aims to discuss these recent guidelines, and continue the examination of filing trends as introduced in our previous article from February 2017.
Most algorithms as they exist now, and so most programs and software, are functional: inputs are supplied, and outputs are computed according to instructions a human has written. One could introduce (pseudo) random elements, but fundamentally a human has ‘told’ the machine what to do. In comparison, artificial intelligence (AI) aims to let a computer perform tasks that we might call ‘reasoning’ or ‘adapting’. In principle, one could conceive of an enormously complex program, nonetheless written by hand, that mimics this behaviour to arbitrary precision with an enormous tree of actions for every conceivable situation a machine would encounter: in practice, this is totally impractical.
Instead, most AI relies on ‘machine learning’ (ML), which is a subset of the whole field of AI that describes a process where the behaviour of the system is learned rather than manually coded. Some implementations of ML are extremely basic, some are very complex: like the whole AI field, it represents a broad spectrum but remains a useful term, especially in patents because it encompasses so many different types of algorithm. For a more in-depth introduction to the world of ML and AI, this article from Medium.com provides a comprehensive overview.
To dispel a popular misconception: it is not the case that you cannot patent software at all. However, it is also not true that you automatically can patent software. Although it varies depending on jurisdiction, and more detail will be given below, the general stipulation is that software or algorithms can form part of a patentable invention. It is not usually possible to patent software if it is not in the context of a specific use to solve a specific problem. Although ostensibly an extension of situations already encountered in software patenting, AI, in particular, is garnering considerable attention because it often forms a core part of inventions in its own right; perhaps more so than has ever been true with algorithms in general. This is part of the reason patent offices are currently devoting time to clarifying guidelines on patentability.
EPO Requirements for AI Patents
General Software Patents
Article 52 of the European Patent Convention describes patentable inventions, and (in paragraph 2) specific exclusions:
- discoveries, scientific theories, and mathematical methods;
- aesthetic creations;
- schemes, rules, and methods for performing mental acts, playing games or doing business, and programs for computers;
- presentations of information
The third of these exclusions would appear to preclude software from patentability entirely (as does the first exclusion, since nearly all algorithms can be said to be ‘mathematical methods’). However, the Act clarifies that these exclusions only apply to inventions that comprise the exceptions listed above ‘as such’: that is, without additional features. Therefore, while not eligible for patent protection in isolation it appears software can form part of a patentable invention.
Perhaps the most important terminology to consider with regards to EPO patentability of software is the ‘technical character’ of an invention, referenced repeatedly in EPO cases and guidelines. This has been the focus of a recent clarification that seeks to present examples where mathematical methods or software contribute to the ‘technical means’ by which a problem is being solved, and are thus eligible for patentability. These examples include:
- controlling a specific technical system or process, e.g. an X-ray apparatus or a steel cooling process;
- determining from measurements a required number of passes of a compaction machine to achieve a desired material density;
- separation of sources in speech signals; speech recognition, e.g. mapping a speech input to a text output;
- encoding data for reliable and/or efficient transmission or storage (and corresponding decoding), e.g. error-correction coding of data for transmission over a noisy channel, compression of audio, image, video or sensor data;
These examples seem to show that, although unlikely to be eligible for patent protection on its own, an algorithm may be eligible if it is used as part of an invention that does something that is ‘technical’.
From these examples, it can be inferred that the EPO is more likely to grant patents to AI/ML inventions that interact with videos, images, machines, (to an extent) text, or other external factors, as opposed to algorithms that simply sort or tag data. Again, the question of the problem solved being ‘technical’ must also be considered: they have specifically ruled against an ML system to generate customer bills in the past. A further example of the EPO rejecting an invention related to an ML document classifier; in this case, the implementation was regarded by the appeal board as being too ‘obvious’ a use of ML to warrant protection.
As well as the examples highlighted above, the EPO recently held a conference on the problem of patenting AI inventions (see the summary here). At this event, it was reaffirmed that the EPO asks of AI/ML inventions: ‘Do(es) the AI and ML method (steps) contribute to the technical character of the invention?’. Elaborating on the recurring use of the term ‘technical effect’, it was recommended that applicants include in the specification of their invention ‘as much information about the technical effect as possible’ to increase the likelihood of acceptance.
|EP3219564B1||IMRA Europe SAS||2016-03-14||Active||Driving prediction with a deep neural network|
|EP3203330B1||Vaillant GmbH||2016-02-05||Active||Soft sensor for the identification and regulation or control of a ventilation system|
|EP2646911B1||Cisco Technology Inc||2010-12-01||Active||Detecting malicious software through contextual convictions, generic signatures, and machine learning techniques|
These are a few granted EP filings that involve AI/ML. The first is typical of grants where the ML architecture (in this case a neural network) is tangential to the problem being solved. The method is framed as a way of helping drive a vehicle, and the neural network is simply a means to an end (the same way any other software might be). The second is remarkably specific in its claims, which apply exclusively to ventilation and air conditioning systems. It includes specific reference to neural networks and represents another example of a very specific technical problem where ML is just one component of the broader process being protected.
The third example takes the approach of not mentioning ML explicitly. There is some subtext in the patent claims, for instance it can be inferred that the ‘aggressive detection capability’ is meant to allude to ML applications. This method of so-called ‘black boxing’ the ML portion of an invention is fairly typical, although it can lead to pitfalls with regards to prior-art if the claims are overtly broad.
US Requirements for AI Patents
The USPTO has been decidedly tight-lipped on explicit guidance for examining ML/AI patents, and unlike the EPO and World IPO has not held any major events or conferences on it. Nonetheless, some guidance can be attained from first looking at the two criteria to determine whether the subject matter considered is eligible for a patent. This is not quite the same as the ‘patentability’ test (of ‘novelty’ and ‘non-obviousness’); instead it focusses on the form of the invention and some exceptions that are excluded from patent protection:
- The first test is that the invention should fall into one of four categories:
- Methods of manufacture
- Compositions of matter
- The second is that the invention be ‘useful’, and does not fall into any of the explicitly listed exceptions:
- Laws of nature
- Physical phenomena
- Abstract ideas
For the purposes of ML and AI, the third exception (‘abstract ideas’) is the most relevant: an example of this in practice is highlighted in this article where a pure-data application of ML was deemed to be too abstract by a California district court to qualify for patent protection. The exact wording found in the judgement was that the invention makes ‘use of computers only as tools, rather than provide a specific improvement on a computer-related technology’. The claims of the invention in question essentially comprise a three-step process of learning some functions, evaluating them, and selecting the best one. Since in this case (as in many others) the question of software patentability revolves heavily on the ‘abstract ideas’ exception above, the USPTO released guidelines in January 2019 designed to clarify this exception specifically in the context of software inventions.
Recent guidelines on ‘abstract ideas’
It should be acknowledged that the 2019 guidelines are not specifically focussed on AI or ML inventions; in this respect the USPTO still lags behind the EPO. Instead, they provide extra guidance on abstract ideas in light of the famous 2014 ‘Alice’ case, dividing them into three categories:
- Mathematical concepts
- Methods of organising human activity
- Mental processes
The USPTO has since given some specific examples of some fictitious inventions where (all else being equal) the question of patentability rests on these criteria. They especially focus on the ‘mental processes’ exception. One example given includes a process of ranking things, and notes that there is no reason a computer is required to do the ranking, a human could conceivably do it mentally. Thus, in this case, the invention is deemed ineligible for patent protection.
However, and here we return to the general concept of inventions having practical applications, if the invention provides some kind of output that requires manipulating images or creating a sound file or anything else that means the process could not be replicated ‘mentally’, then the subject matter is eligible for patenting.
It is important to underscore that the examples demonstrate that the ‘mental process’ need not be something a normal human could do in any practical length of time or without pen and paper, it is more a statement on whether the process is abstract enough that it does not interact with anything.
Even if the answer to this question is ‘yes’ (that is, there is a ‘mental process’ involved) that does not automatically preclude patentability. In these cases, the examiners are being directed to consider if the abstract mental process is ‘Integrated into a Practical Application’ (quoting directly). The way this is phrased in one of the examples is ‘The claim as a whole integrates the mental process into a practical application’. An example given for where this does not hold is where an invention comprises a process for solving a problem (which could be framed purely as a ‘mental process’), but it happens to be the case that the problem is solved by a computer. In other words, it is not enough to claim that a mental process is part of a practical application by saying ‘it is performed by a computer’ with no further qualifiers.
Expanding on this, and underscoring the importance of carefully defining the practical application by being descriptive of the problem being solved, IP Watchdog provides this quote: ‘In a nutshell, if you are going to write a patent application in such a way that the reader will be left wondering what the innovation is, what the problem being solved is, or the technical particulars on how the innovation actually solves the problem and achieves the specified functionality, you should not expect a patent’.
|US8880445B2||Certica Solutions Inc||2012-04-20||Active||Method and apparatus for performing dynamic textual complexity analysis using machine learning artificial intelligence|
|US10140553B1||Capital One Services LLC||2015-12-22||Active||Machine learning artificial intelligence system for identifying vehicles|
|US20180108025A1||SAP SE||2016-10-19||Pending||Systems and methods for sentiment insight in electronic media postings|
The first example satisfies the previously discussed criterion of being a very concrete application of machine learning. Although it could be argued that analysing text is a process that could be performed purely ‘mentally’, the claims do describe supplying predictions as an output, which may be enough to meet the conditions of the new USPTO guidelines.
The second example is perhaps a more ‘classic’ case of a defined ‘practical’ problem (in this case, classifying images) having a solution implementing some form of machine learning. In the claims, some steps are nebulously described as ‘calculating’ values or ‘extracting features’ from the images: the implication being that these may be performed by ML algorithms. As with the third example of EPO patents, this is not made explicit in the claims themselves.
The third example evades explicit mention of machine learning even further, providing only one indication (in the detailed description) that ML may be employed to perform sentiment analysis as part of the method. Once again, the claims focus instead on the stages of the invention ‘analysing’ and ‘applying rules’ only, effectively ‘black boxing’ the algorithms used to perform these steps.
AI and ML Patent Filing Data
Worldwide, applications for AI and ML patents continue to grow approximately exponentially, with a compound annual growth rate of around 20%. Funding remains strong, with an estimated $9.3bn of VC investment in AI focussed startups last year (an increase of 72% on 2017). In the graphs below, the shaded area represents incomplete data from recent filings.
Worldwide AI and ML Patent Applications and Grants (20 Years)
‘Big Tech’ companies dominate in AI patents, with Microsoft showing a commanding lead. In recent years, Chinese companies such as Baidu and the Alibaba Group have shown increased filing, in line with increased filing from China as a whole.
Terminology used in the patent texts indicates a focus on image processing and sensing, but also a large number refer to databases (although this may only be in relation to storage of processed data, rather than as the main focus of inventions). Vehicles, mobile devices, and robots were mentioned most frequently with regards to hardware. This reflects trends in the development of autonomous vehicles and robotics, as well as an increase of AI deployments on mobile phones.
Count of Key Terminology in AI and ML Patents
The subject of patenting AI/ML is an active and ongoing issue in most patent offices worldwide. The EPO’s recent conference highlighted several future challenges, and the USPTO has indicated they may make additional changes or clarifications. As such, it is important to keep abreast of updates or new information from these authorities as they are released. SMEs can expect ongoing interest from ‘big tech’ companies, and should be mindful of their freedom to operate within AI and ML in light of the large scale of some of the portfolios of these large corporations especially as they get nearer to a significant funding round, trade sale or IPO.
Prior art searches remain a vital component of qualifying potential patent filings, especially as ML and AI are not always mentioned explicitly in patent claims, but may fall under general umbrella terms such as method steps involving “analysing” or “extracting information” or “classifying”. In addition, landscapes of the current patent environment can provide invaluable information for two major reasons. Firstly, they can serve to highlight as-yet untapped areas of applications for AI and ML that may be exploited with future filings. Secondly, they can be a valuable source of knowledge of which patents are currently being granted: examining successful patents may help with the construction and scope of future applications. The findings may also serve to highlight important distinctions in different jurisdictions that could prove the difference between a narrow, low-value patent and a high-value asset that helps protect a company’s core technology.