By Hidemichi Fujii and Shunsuke Managi*
Patent applications are a good indicator of the nature of technological progress. This column compares trends in applications for artificial intelligence patents in Japan and the US. One finding is that the Japanese market appears to be less attractive for artificial intelligence technology application, perhaps due to its stricter regulations on the collection and use of data.
The number of patents granted has rapidly increased. Figure 1 shows the number of artificial intelligence (AI) patents granted by application country and technology type and reveals that it has increased more than threefold (from 708 items in 2012 to 2,888 items in 2016). In particular, AI patents granted in the US increased by 1,628 items during this period (Figure 1a), accounting for approximately 75% of the increase worldwide.
Figure 1 Trend of AI patents granted, 2000 to 2016 (number of items)
a) Number of AI patents granted by country
b) Number of AI patents granted by technology
Source: Author estimates using IPC code in Appendix 1 and PATSTAT database.
Note: USPTO: United States Patent and Trademark Office; SIPO: State Intellectual Property Office of The People’s Republic of China; JPO: Japan Patent Office; PCT: Patent Cooperation Treaty; EPO: European Patent Office
As shown in Figure 1b, the patent share of each AI technology type changed from 2012 to 2016. In 2012, biological and knowledge-based models were the leaders in patented AI technologies. However, from 2012 to 2016, the number of patents granted for specific mathematical models and other AI technologies rapidly increased, doubling from 2015 to 2016. These two figures show the short-term trend of AI patenting based on country and technology.
The time and geographical specifics of AI technology patents
Table 1 represents the change of AI patents granted by type of technology at each patent office. The table shows that the composition of patent-granted shares differs among countries. The knowledge-based model represents more than half of the total number of AI patents granted by the US Patent and Trademark Office (USPTO), whereas the biological model is the major technology type granted by the State Intellectual Property Office (SIPO) and the Japan Patent Office (JPO). Another finding is that the share of the specific mathematical model is only 1.7% in the JPO, which is extremely low compared with that of other patent offices. This outcome occurs because Japanese AI researchers primarily focus on android technology-based research and development (R&D) – and not mathematical elements – which represents the core AI technology (The Japan News, 2017). The Patent Cooperation Treaty (PCT), the European Patent Office (EPO), and the patent offices of other countries exhibit similar trends with respect to the technology share pattern of AI patent publications.
Table 1 Data descriptions of AI technology patents granted (item)
Source: Author estimates using IPC code in Appendix 1 and PATSTAT database.
Note: USPTO: United States Patent and Trademark Office; SIPO: State Intellectual Property Office of The People’s Republic of China; JPO: Japan Patent Office; PCT: Patent Cooperation Treaty; EPO: European Patent Office.
Next, we consider the numerical change of AI patents granted. As shown in Table 1, all of the patent offices except the JPO published the highest number of AI patents from 2015 to 2016. Notably, the number of patents granted more than doubled at the USPTO, SIPO, and PCT. However, the average number of patents granted per year at the JPO was highest from 2005 to 2009 for the biological model, and from 2010 to 2014 for the knowledge-based model.
One interpretation of this result is that the Japanese market is less attractive for AI technology application. Most AI technology services are strongly related to big data collection through the internet (such as social network systems, credit card payments, and sensors). Because of concerns among its residents, Japan is strict regarding the use of private information for business (Kawasaki 2015). The business barrier regarding big data collection and use minimises the incentive to obtain AI patents in Japan. In the US, the government has established rules and regulations regarding the use of private information as Big Data (Hardy and Maurushat 2017, Manyika et al. 2011). Additionally, there are large governmental R&D expenditures for AI technology innovation in the US, which is another strong incentive for AI technology development (National Science and Technology Council 2016).
Attributing AI technology patents correctly
Figure 2 lists the top 30 applicants for AI patents granted worldwide. The bottom rows represent the number of patents granted to universities in the US, China, and Japan. As shown in Figure 2, IBM Corporation is the world’s leading recipient for AI patents granted. Additionally, of the top 30 grantees for AI patents granted, 18 applicants are US companies, 8 applicants are Japanese companies, and 4 applicants are companies from other countries. Notably, Chinese companies and universities are not listed among the top 30 countries evaluated for the 2000-2016 period, which implies that AI patents granted in China are obtained by many applicants.
Figure 2 Number of AI patents granted and technology portfolios, 2000 to 2016
Next, we discuss the composition of the AI technology patent share for each applicant. Figure 2 indicates that the patent portfolio of AI technology varies among applicants. Qualcomm and BRAIN Corporation garnered the highest share for the biological model. However, SAP and Cognitive Scale had the largest share for the knowledge-based model. D-Wave Systems obtained 92% of other AI patents, an outcome that represents a completely different trend from other companies. Notably, the companies listed in the top half of the list obtained patents in a wide range of AI technology areas.
According to Figure 2, a large proportion of the AI patents granted to Chinese and Japanese universities were for technology based on the biological model. This trend differs from that found for US universities. In addition, US universities have obtained a large proportion of patents for AI technology that uses a knowledge-based model. This trend resembles that found for the composition of patents granted by the USPTO (see Table 1).
Who invented a patented AI technology and where?
As shown in Figure 3, most US companies have a large share of AI patent invention according to the USPTO data. By contrast, the share of patented inventions of US companies from the JPO and SIPO is small. With the exception of NTT Corporation, non-US companies have more than a 16% patent share at the USPTO. Specifically, Samsung Group has 61% of all AI patents issued by the USPTO. Surprisingly, four of the eight Japanese companies were granted more AI patents by the USPTO than by the JPO. These results imply that Japanese companies have strong incentives to obtain AI patents from the USPTO, while there is less incentive for US companies to obtain AI patents from the JPO. This result is consistent with the interpretation that big data use creates an advantage for the US market with respect to AI technology application.
Figure 3 Distribution of country or organisation of AI patents granted from 2000 to 2016
Based on Figure 3, universities clearly tend to apply for AI patents at domestic patent offices. In particular, 98% of the AI patents obtained by Chinese universities were granted by the SIPO, with a low number granted by other patent offices. By contrast, US and Japanese universities apply for AI patents at the PCT in addition to their domestic patent offices. Notably, approximately 45% of the AI technology patents granted in China were obtained by Chinese universities. This outcome is unique. In other countries, private companies are the primary patent applicants.
Priority shift of AI technology invention
Figure 4 shows the results of a decomposition analysis for four specific AI technology patents granted at all of the patent offices listed in the patent scope database. Because the AI patent trend changes beginning in 2012 (Figure 1), we divided the decomposition analysis results into two periods (the first period runs from 2000 to 2011, and the second period from 2012 to 2016). The plotted point in red indicates the change in the number of specific patents granted, and the bar chart shows the effects of each decomposed factor on the number of patents granted related to specific AI technologies. The sum of the bars is equivalent to the value of the plotted point. The figure shows the differences in the driving factors for patents granted based on the type of AI technology.
Figure 4 Results of patent decomposition analysis (number of items)
Note: The vertical axis is standardised by setting the number of changes in patents granted in 2000 and 2012 to zero.
Figure 4 shows that the number of patents granted for technology based on the biological and knowledge-based models increased during the first period. However, the priority of specific AI technology affects these two technology types differently. As shown in Figure 4, during the first period, the relative priority of the biological model was negative, whereas that of the knowledge-based model was positive. This result implies that the priority of AI technology patent invention shifted from the biological model to the knowledge-based model over the first period. The number of patents granted for the other two technology types did not change significantly during the first period, which indicates that these two technologies were treated as less important than technologies based on the biological and knowledge-based models during that period.
Based on the results for the second period, the number of patents granted substantially increased for all four AI technologies. In addition, the priority of specific technologies shifted from the biological and knowledge-based models to the specific mathematical model and other AI model during the second period. Specifically, the number of patents granted for other AI technology was 624 items during the second period, which is more than that for the biological model (565 items) and close to that for the knowledge-based model (693 items) (see red points in Figure 4).
No comments:
Post a Comment