Artificial intelligence patents. Law practice.
The topic of artificial intelligence is still not present in Bulgarian law as something normative and practical, which is on the legislative agenda in the 21st century, not only for technological and legal reasons, which undoubtedly determines the future of the innovative sector and through it the public, business and personal life of every modern person. The latest research on the subject in many of the world's advanced technology countries has shown that it is not far off the time when besides the afternoon chess game with some electronic device, artificial intelligence will be issued with instant visas, will be approved faster and secure (personal data - finger, face recognition) bank credits, national and cross-border (eg European) elections will be held and health services will be provided. Other studies have shown that artificial intelligence will replace many professions - lawyers, notaries, bailiffs, judges, revolutionize medical precision and monitoring, robotize our industry, services and lifestyle, thus putting new intellectual, philosophical and psychological challenges to everyday life and perhaps to the relationships between us - human beings. Realizing the inevitability of all this, many companies operating in different spheres of social and business life began to develop dynamic and focused inventions based entirely on artificial intelligence. Taking this into account, I want to pay close attention to this statement of these patents, driven by my belief that today's inventions in the area of Artificial Intelligence (AI) are the basis of our more interesting tomorrow.
1.Historical development. Artificial Intelligence (AI) appeared in the 1950s, with the first mention of the term coming from a summer 1956 research project of Dartmouth College, New Hampshire, USA. A year earlier, in 1955, John McCarthy, a young assistant professor of mathematics at Dartmouth College, decided to organize a group for exploring and developing digital thinking machines. McCarthy selects the name "Artificial Intelligence" as a "new field" of scientific search. It presumes mostly neutral neutrality in order to avoid focusing on the narrow theory of automation and cybernetics, as already known achievements of analog technology. In early 1955, Mr. McCarthy turned to Robert Morrison, director of biological and medical research at the Rockefeller Foundation, to request funding for the Dartmouth summer seminar for about 10 mathematicians. On 2 September 1955, the project was officially presented to the board members under the notion of "artificial intellect".
McCarthy's team [1] proposes to conduct a two-month study of artificial intelligence by 10 scientists (in fact, 11 mathematicians are involved in the final project) in the summer of 1956 at Dartmouth College in Hanover, New Hampshire on several topics. First of all, the study concerns the assumption that every aspect of learning or any other intelligence characteristic can in principle be so accurately described as to make a machine to simulate it. The experiment also involves attempting machines to use language, to form abstractions and concepts, to solve problems that have so far been a priority of people's mental activity, and to improve the described results. Last but not least, the proposal discusses the development of computers, natural language processing, neural networks, calculus theory, abstraction and creativity - areas in artificial intelligence that are still relevant today. Since the 1950s, innovators and researchers have published over 1.6 million scientific articles related to artificial intelligence, with around 340,000 applications for inventions related to the AI industry being filed. The demand for interest is so great that only in 2013, half of the patent applications on the subject of artificial intelligence have been made.
The McCarthy team and its students created it in the 50 programs that the press describes as "astonishing": computers solve algebraic problems, proving logical theorems, speaking English, and so on. By the mid-1960s, US research was heavily funded by the Ministry of Defence, with artificial intelligence laboratories being established around the world. Progress in the field slowed down in 1974, in response to Sir James Lighthill's criticism [2], and continued pressure from the US Congress to fund more productive projects, and therefore the US and British governments were interrupting exploratory research at the AI sector. The next few years are called "Winter Periods" for Artificial Intelligence, as receiving funding for such projects is extremely difficult.
In the early 1980s, artificial intelligence research was resumed, stimulated by the commercial success of so-called "expert systems" [3], as a form of artificial intelligence program that simulates the knowledge and analytical skills of human experts. Until 1985 the market for artificial intelligence programs reaches over one billion dollars. At the same time, Japan's computing project for fifth-generation computing technology has led US and British governments to rebuild funding for academic research in the AI area. However, starting with the 1987 Lisp market crash, the AI industry once again fell into an unfamiliar market situation, viewed as the second, long-lasting "frost" in research on the subject.
In the late 1990s and early 21st century, AI innovations began to materialize in logistics, data mining, medical diagnostics, and other areas of science and technology. Their tremendous success reaches its peak due to the increasing computational power of Artificial Intelligence computers (see Moore's Law [5]), with emphasis being placed on addressing practical practical problems from business and the economy, new connections between AI and others areas (such as statistics, economics and mathematics, medicine [6]) and the commitment of researchers to new mathematical methods and scientific standards. All in all, he finds his logical technological expression in the first chess system based on artificial intelligence "Deep Blue", which defeated world chess champion Garry Kasparov on May 11, 1997.
A little later, in the beginning of 2015, Jack Clark of Bloomberg notes that the year in question was remarkable for artificial intelligence, with the number of software AI projects on which different types of Google searches are based increased from "sporadic use" in 2012. to more than 2700 projects. Clark also presented factual data showing that the percentage of image processing errors in "Google Images" (artificial intelligence software algorithm) has dropped significantly since 2011. He explains this phenomenon with the increase of available neural computer networks, due to the cloud computing infrastructure being upgraded and the storage of research tools and massive databases in them. Another example in the commented direction is the development by Microsoft of a "Skype" software system [7] which, based on artificial intelligence, can automatically translate from one language to another and "Facebook" software that can describe some images of blind people. In a US study on the subject in 2017, it was found that one in five companies in the country "has included AI technologies in some proposals or processes". In 2016 China also significantly accelerated its state funding for artificial intelligence, taking into account this large database supply and the rapid increase in research results, some observers believe China is about to become the "super power of AI innovation ".
Machine learning is the dominant AI topic covered by multiple patents for artificial intelligence and included in more than a third of all identified inventions on the topic (134,777 patent documents). Patents related to machine training increase annually by 28% annually, and in 2016, 20 195 patent applications were filed (compared to 9,567 in 2013).
Some of the most popular machine learning techniques that revolutionize AI technologies are so-called "deep learning and neural networks." They are also the fastest growing achievements at the technological level reflected in patent applications: so-called "deep learning" shows an impressive average annual growth rate of 175% from 2013 to 2016. , reaching 2,399 patent applications in 2016. Neural networks grew at a rate of 46% over the same period, with 6,506 patent applications in 2016.
Among the AI functional applications subject to patents, the so-called "computer vision" (8), which includes image recognition, is most popular. Computer vision is mentioned in 49% of all patents related to AI technologies (167 038 patent documents), increasing annually by 24% on average (21 011 patent applications filed in 2016). Other AI functional applications with the highest growth rates in patent applications in 2013 - 2016 are the artificial intelligence for robotics and control methods that grow on average by 55% per year.
2.Challenges in front of European Patents for Artificial Intelligence. The European Patent Office or abbreviated as "EPO" is increasingly receiving patent applications that include the term "programmed computer" as a key part of the invention described. Moreover, this increase in document submission is observed in technical areas that are traditionally not considered computer-centric. For example, according to EPO statistics, 40% of new patent applications filed in healthcare have an AI or machine learning aspect.
In recognition of the growing importance of artificial intelligence and machine learning for patent applications in all fields, the EPO has devoted time and place to the updated "Expertise Guidelines 2018" [9] to focus specifically on the patentability of inventions that have AI elements and / or affect aspects of machine learning. This shows a particularly detailed view of the EPO on the most up-to-date trends in software patents, which will actually assist applicants and patent attorneys across Europe.
The new "Expertise Guidelines" mentioned clearly show that EPO intends to treat AI technologies and machine learning as a form of mathematical method. Mathematical methods are part of the objects listed in the list of non-patentable inventions defined in Art. 52, para 2 of the European Patent Convention (EPC), and in this line of thought are essentially non-registrable "as such". However, the mathematical method that is related to the control of a technical system or process may be technical in nature, thus overcoming its "exclusion" as a non-patentable invention.
This has always been the constant position of EPO when commenting on exceptions to patentability, and so it is not surprising that the Artificial Intelligence and Machine Learning section of the new "Guidelines for Expertise" is largely based on the practice of the Office. It was therefore accepted that inventions involving artificial intelligence and machine training would be patentable as far as they are described and stated in the context of working in a technical system or controlling a technical process. Careful preparation of the application in this context will be sufficient to ensure that this requirement is met - describing and claiming artificial intelligence or machine learning component in the context of the technical system in which they work or maintain technically, such as any abstraction in the opposite context is excluded. Only such an approach according to the EPO would lead to the issuance of a European patent. Artificial intelligence or machine learning algorithms that work in the context of non-technical systems, such as business processes and models, are unlikely to be accepted as patentable.
The shared motives are reflected in EPO Board of Appeal Decision T1510 / 10, issued in December 2013, which shares the view that the use of machine learning (as well as AI technologies) is not in itself sufficient to make an invention is patentable. This means that the conventional use of machine learning or artificial intelligence to solve a problem that may or may not be solved by that means does not, a priori, mean that a technical effect has been achieved even if the problem it decides technically on its merits.
In Section "G - Patentability", Chapter II, "Inventions", section 3.3.1 "Artificial Intelligence and Machine Training", the new "Guidelines” of the EPO from 2018, lead to the following conclusions:
Artificial intelligence and machine learning based on computational models and algorithms for classification, clustering, regression and reduced dimensionality such as neural networks, genetic algorithms, support vector machines, k-means regression and kernel discriminant analysis. Such computational models and algorithms have per se an abstract mathematical nature, whether they can be "trained" on the basis of certain existing bases. Consequently, the guidelines presented in G-II, 3.3 [10] are generally applicable to such computational models and algorithms.
In considering whether the patentability of the invention is technical in its entirety (Article 52 (1) (2) and (3) of the EPC), it expressed as a "support vectors machine", a "reasoning engine" or "neural network" should be interpreted and interpreted very precisely and carefully by the expertise because they usually refer to abstract models of a technical nature.
On the other hand, it should be noted correctly that artificial intelligence and machine learning are applied in various fields of technology. For example, using a neural network in a heart rate monitor to identify irregular heartbeats is a process of technical input. Classification of digital images, video, audio or speech signals based on features at a low level (eg edges or attributes of pixels for images) are other typical technical applications of mathematics, computer AI algorithms for classification. However, the classification of textual documents only with regard to their textual content is not considered a technical objective in itself but as a linguistic one [11]. The classification of abstract data records or even "data records for telecommunications networks", without specifying the technical use of the resulting classification is not a technical purpose, even if a classification algorithm can be considered to have valuable mathematical properties such as such as sustainability [12]. When the classified method (the subject of the patent) maintains a specific technical purpose, the steps for its generation can also contribute to the technical nature of the invention if they support the same objective.
The detailed analysis carried out by the EPO in order to understand the depth of discussions theoretical problem and the correct interpretation of the patentability of artificial intelligence, could lead to significant legal development of the practice, as it will open the door to the possibility of being received European patent protection on training methodologies algorithms, AI innovations or machine learning as well as mechanisms for generating sets of data that are used for the intended purpose.
In my opinion, because of the said European patent, a method of training an artificial intelligence or machine learning algorithm or a method of generating training data for this purpose would be provided if it is possible to make a reliable justification of the proven, and a repeatable technical effect. For example, a training method that makes the neural network "converge" faster with technology or uses a smaller set of data can be credited as serving to solve a technical problem and thus meet the legal requirements for a European patent protection.
The mentioned patentability analysis of AI innovations introduces a well-known aspect of patent law that is commonly found only in the pharmaceutical and biotechnology spheres - "credibility". For example, it may be proven that a particular untrained software AI model has "converged" faster when "trained" using a specific method and a specific set of databases for its "learning", but only such evidence will not be sufficient to make a plausible claim. As it has already become clear, "credibility" itself as a criterion is also conditioned by whether an AI patent will result in a real and guaranteed technical result or not. Any abstractness of the claim would result in a lack of sufficient plausibility, which in turn will end with an opinion of the lack of patentability.
3.AI Artificial Intelligence in the US - Trends and Legal Framework. Several recent reports from America show that patentability of research objects in AI technologies has been extremely active over the past few years. In December 2016, Google and Elon Musk [13] opened their AI platforms publicly, Uber launched the Uber AI Lab69 project, and Apple announced that it would be publishing its research in the field of artificial intelligence for the first time.
Significant interest in the US also exists in future applications of artificial intelligence with other "intertwined" parallel technologies such as robotics, virtual reality, autonomous vehicles, block, 3-D printing and IoT.62 [14]. There is fierce competition for leadership in the AI sector among several leading companies, which helps to stimulate artificial intelligence-based innovation, along with accelerating progress in current and future applications. Technology giants such as Google, IBM, Microsoft, Intel, Facebook, Amazon, Baidu, Samsung and Apple have patented hundreds of AI patents and some of the industrial multinational companies like Boeing have and to acquiring start-up AI companies. The deep learning sub-sector [15] is currently also promoting innovation pioneering as an investment activity.
In the United States, most artificial intelligence technologies can be protected by a patent. But some inventions of artificial intelligence are faced with the increased legal control of the US Patent and Trademark Office (USPTO) expertise, especially with regard to whether the invention is included in the criterion of patentability". The USPTO follows a two-step analysis to determine whether a patent claim is permissible for a patent. First, the USPTO determines whether the patent application is focused on a concept that meets the patent requirements. Some areas are not patent admissible: abstract ideas, mathematical models, natural laws, and natural phenomena. If the patent application addresses one of these areas, the USPTO examines in its essence whether the claim under investigation as a whole amounts to "more than" the above-mentioned concepts, in which case it is decisive whether the formulated claims reach a real technical result and not just an abstraction. Inventions based on AI technologies related to autonomous vehicles or robots that aim to control, move or manipulate a tangible object (for example, a vehicle or a postal package) are generally subject to a comparatively minimal analysis of the expertise. These technologies are generally considered to be eligible for patenting, as it is assumed that they lead to a tangible technical result and are therefore not abstract.
By contrast, an artificial intellect-based invention that is not aimed at controlling material objects - such as a software algorithm - may face increased control in the expertise of whether it is focused on an abstract idea or not. However, many technical aspects of even an invention based on AI innovations can overcome stringent substantive patent requirements in the United States. To the extent that the USPTO as an institution has not provided explicitly a legal definition of the term "abstract idea", this loophole in US case law has been filled by a number of judgments that have been sufficiently illustrative in this respect. For example, the Federal Court of Justice has held that patent claims relating to a particular data arrangement are permissible for patents, stating in particular that the "self-refering table" set forth in the claims is a specific type of data structure. In this context, the compilation of specific data structures, specific rules, specific combinations of technical steps, or specific hardware configurations that improve computer performance are accepted as eligible claims, while most cases of usable use of a general-purpose computer are often considered as non-putative.
On the basis of these criteria, the US Patent Office (USPTO) identifies patentable AI objects, and in this line of thought, the AI-based technology developers are required to go beyond expected user scenarios (for example, patenting method for the use of conventional technologies to solve a common problem), the latter attempting to identify the unique technical features of its AI patent application by improving the performance of computer. These technical features may include the following components:
- pre-processing of training data (so-called "taxonomy");
- the learning process itself (e.g., neural network topology, configuration of parameters, termination conditions, etc.);
-use of trained classifiers or solutions (eg, sequence using classifiers, modeling of a space for solution of a genetic algorithm);
- "end-to-end" workflow (eg user interfaces);
- hardware (integration of artificial intelligence algorithms into hardware components, hardware acceleration);
Section 35 of the Civil Code of the United States, Section 101 (hereafter "35 USC § 101") limits the patentable objects to "new, useful technical processes, machinery, production or composition of matter, or any new and useful improvement in her ". As I have explained several times, patent claims that are aimed at abstract ideas (eg mathematical algorithms), natural phenomena or laws of nature are not eligible for patent protection in the United States. The Supreme Court of the United States motivates this constant view with the fact that these objects "are the main tools for scientific and technological work", and the granting of monopolies on these instruments through patent rights can seriously hamper innovation.
An example of the requirement of US case law, and in particular the hypothesis of 35 U.S.C. § 101 that a patentable AI invention should not be "aimed at an abstract idea" or should include an "inventive concept" that goes beyond an "abstract idea" – it has emblematic case law of Alice v. SEEels Bank Alice v. CLS Bank "[16] just on the topic. In 2016 however, the US Federal Court has examined another case - Enfish v. Microsoft (17), which significantly disproved the motives of the Alice v. SEEels Bank Decision. At the turn of the century, the “Enfish” firm registered U.S. Patent No. 6,151,604 and 6,163,775 which claimed a logical model for a computer database. The logical model is a computer database system that explains how the various elements of the information in the database are linked together. Unlike conventional logical models, “Enfish” includes all data in a single table, with column definitions provided by rows in the same table. Patents describe this as a "stand-alone" property of the database. In a standard, conventional relational database, each unit (ie, each type of thing) that is modeled is provided in a separate table. For example, a corporate file replication model may include the following tables: document table, face table, and company table. The document table may contain information about the documents stored, the persons table may contain information about the authors of the documents, and the company table may contain information about the companies that employ the persons. “Enfish” patents describe a table structure that allows information normally appearing in several different tables to be stored in one. Columns are defined by rows in the same table. Initially, the “Enfish” case against Microsoft was misinterpreted under the hypothesis of the "abstract idea" of the precedent “Alice”. In fact, the Federal Court of Justice ruled in its reasoning that “Enfish” patent claims are aimed at a specific improvement in the way computers work, embodied in the claimed "self-referencing table" for a database that the prior art does not contain. An interesting fact is that this case is often used as an example of one of the first cases concerning key details in understanding the patentability of AI technologies.
4.Conclusion. At the end of this presentation, I would like to point out that, in my opinion, the implementation of highly specialized legal norms regulating artificial intelligence in patent law at national and cross-border level as well as the development of legal regulation of AI technologies would have a profound impact on innovation, the economy and society. Given the global, explosive development of the AI sector, it is of paramount importance that relevant stakeholders - patent specialists and businesses actively participate in further research and discussions among themselves, as well as in a more comprehensible presentation of this particularly complex topic of the public to find the most appropriate ways for artificial intelligence to promote innovation while minimizing any potentially negative social, business, legal and ethical implications.
Author: Mr.Atanas Kostov – patent attorney
[1] See. Russell, Stuart, J .; Norvig, Peter (2003), Artificial Intelligence: A Modern Approach (2nd ed.), Upper Saddle River, New Jersey: Prentice Hall, ISBN 0-13-790395-2., Page 17, over the next 20 years, artificial intelligence will be dominated by these people and their students. "
[2] British scientist, mathematician, pioneer in the field of aeroacoustics;
[3] For them again Russell, Stuart J .; Norvig, Peter (2003), Artificial Intelligence: A Modern Approach (2nd ed.), Upper Saddle River, New Jersey: Prentice Hall, ISBN 0-13-790395-2., Pages 22-24;
[4] "Lisp" machines are general purpose computers designed for effective performance through “Lisp” as the primary software and programming language, usually through hardware support. They are an example of hi-tech computing architecture and, in a sense, the first unified workstations. “Lisp” machines are commercial pioneers in many of the most common technologies, including laser printing, computer mice, high resolution raster graphics, and so on.
[5] Moore's law is expressed in the observation that the number of transistors in a dense integrated circuit is doubled every two years. The observation is named after Gordon Moore, co-founder of Fairchild Semiconductor and CEO of Intel, whose 1965 study. describes doubling the number of components for an integrated circuit each year by predicting that this growth rate will continue.
[6] See the Google DeepMind Health project, which was launched in 2016. The project has successfully started working with the National Healthcare System (NHS) in England.
[7] The product is called "Microsoft translator API" and is protected in 2010. with United States Patent US20110307244A1 registered. The owner of the patent is Microsoft Corporation, but the interesting moment is that one of the two inventors is Bulgarian Christina Nikolova Tautanova, graduated from Stanford University. Ms.Tautanova as an inventor and computer scientist is at the heart of several other patents in the field of artificial intelligence with Microsoft Corporation.
[8] See PCT / EP2014 / 071032 "Method for determining a property belonging to at least part of a real environment" Applicant Apple Inc.
[9] A special scientific conference - "Patenting Artificial Intelligence" - was held on 30 May 2018 at the headquarters of EPO, Munich, Germany.
[10] This point G-II, 3.3 commented on the lack of patentability of mathematical methods.
[11] Argument of decision of the EPO Board of Appeal T 1358/09;
[12] This is the decision of the EPO Board of Appeal T 1784/06.
[13] The creator of the PayPal electronic payment system and the car giant for electric cars - "Tesla".
[14] The "LoT" abbreviation comes from the "The Internet of things". This is an Internet connection extension to physical devices and everyday objects.
[15] "Deep learning" is a form of machine learning in which neural networks provide computer-based information for decision-making, training, and process correction based on what the computer has learned within certain parameters.
[16] Alice Corp. v. CLS Bank International, 573 U.S.A. 208, 134 S. Ct. 2347 (2014) is a decision since 2014. of the United States Supreme Court on patent patentability. The case concerns the question whether certain claims for a computer-implemented electronic escrow service to facilitate financial transactions embrace abstract ideas that do not qualify for patent protection. The patent has been declared invalid, as the claims are formulated as an abstract idea, and their application to the functionality of a computer is not enough to make this idea a patentable object;
[17] See Enfish, LLC v. Microsoft Corp., 822 F.3d 1327 (Fed Cir., 2016).