Stanford University released the famous “Artificial Intelligence Index Report 2021” (hereinafter referred to as “Report”), which quickly attracted circulation in the field of artificial intelligence.
Since 2017, Stanford University AI Index has released a report every year. Due to the advantages of detailed data, clear sources of investigation, and abundant opinions, it has a relatively authoritative voice in summarizing the development of AI in the past year, and has always been a concern for readers. This year, led by the HAI lab led by Feifei Li, is the fifth edition.
In addition to the original technology trend analysis, this year’s report also adds a survey of global robots, as well as legislative activities at the AI level in 25 countries and regions (excluding China, so I won’t discuss it here).
What’s more noteworthy is that this year’s Stanford AI Index report for the first time discussed the implementation of artificial intelligence in real life, including the role of AI in economic development, and the potential ethical issues caused by the promotion of AI. .
The report points out that, on the one hand, the proportion of external investment in the field of AI is on the rise: compared with 2020 ($46 billion), the funds invested in AI in 2021 will increase by 103% ($96.5 billion). At the same time, AI has the characteristics of “more beautiful and cheaper”. For example, since 2018, the cost of training an image classification system has been reduced by 63.6%, while the training efficiency has been reduced by 94.4%.
On the other hand, the increasing penetration of AI in real life has also exacerbated the crisis of social ethics. Therefore, next, we may see that more and more discussions on AI are no longer limited to the innovation of models and algorithms, or whether connectionism and symbolism are better, but: before the era of artificial intelligence fully arrives , what preparations should we do at the “infrastructure level”?
Since the content of the report is more than 200 pages, AI Technology Review only organizes the report from the perspective of comparison between China and the United States:
According to the report, from the perspective of research institutions, universities worldwide have the highest contribution to AI research publications, at 59.58%, while companies only account for 5.21%:
On this level alone, the proportion of colleges and universities in AI publications in the United States continues to decline, accounting for 57.63% in 2021, and enterprises accounting for 9.76%:
In China’s AI publications, the proportion of colleges and universities is increasing, accounting for 60.24%, and although the proportion of enterprises is also increasing, it will only account for 3.93% in 2021:
Despite the tension between China and the United States, the report found that between 2010 and 2021, the number of AI papers co-authored by China and the United States ranked first in the number of cross-border collaborative papers, 2.7 times that of the second place (China-UK cooperation):
The report also compares the proportion of AI publications in China, the EU & the UK, and the US in the past 12 years. Among them, China has always maintained the number one, accounting for 31.04%, followed by the EU and the UK (19.05%), and the US is 13.67% %:
In terms of the number of citations of AI journal papers, China’s proportion has gradually increased, ranking first in 2021 (27.84%), while the number of citations in the United States is 17.45%:
At the AI conference, China also ranked first in the world in the number of papers published, accounting for 27.6%, while the United States ranked third with 16.9%:
Although China has an advantage in the number of publications, the report found that the United States has the highest number of citations in AI conferences, accounting for 29.52%, while China has only 15.32%:
In AI repository (such as arXiv) publications, the United States has maintained its lead since 2011, accounting for 32.52% in 2021, but China is not far behind, and the proportion has been rising, accounting for 16.6% in 2021:
In 2021, the U.S. also ranks first in the number of citations to AI repository publications, with a citation rate of 38.6%, compared with 16.4% for China:
In the application of artificial intelligence patents, in 2021, China will apply for more than half of the world’s artificial intelligence patents (51.69%), compared with 16.92% in the United States:
First of all, in terms of talent recruitment, the report shows that Hong Kong, China ranks second in the world in the growth rate of AI talent recruitment, an increase of 1.56 times compared with 2016:
By calculating how often LinkedIn users self-added skills in a given field between 2015 and 2021, the report found that in terms of AI skills penetration, India has the highest average penetration rate, followed by the US (2.24) and China at 1.56 , ranked fourth in the world:
In addition, in terms of investment in the AI industry, US AI companies ranked first in the world in terms of overall private investment, with about $52.9 billion, while China ranked second with $17.2 billion, and the United States was three times that of China:
In terms of total private investment from 2013 to 2021, US investment totaled $149 billion and Chinese investment totaled $61.9 billion:
Notably, from 2013 to 2021, private investment in AI companies in the US was more than double that of China, which itself was around six times the total investment in the UK over the same period. By geographic region, as shown in Figure 4.2.6, investment in the United States, China and the European Union increased from 2020 to 2021, with the United States leading China and the European Union by 3.1 times and 8.2 times, respectively:
In terms of the number of AI companies, in 2021, the United States will lead with 299 companies, followed by China with 119:
In terms of AI adoption rate, the highest adoption rate in 2021 is product and/or service development in high-tech/telecom (45%), followed by service operations in financial services (40%), and high-tech/telecom (34%) and risk functions in financial services (32%):
In terms of the type of AI capabilities adopted, in 2021, the highest embedding rate is natural language text understanding in the high-tech/telecom industry (34%), followed by financial services and robotic process automation in the automotive and assembly industries (33%) and financial services Natural language text understanding for services (32%):
So, what are the possible risks of adopting artificial intelligence?
According to the report, 55% of respondents believe that the most prominent AI application risk in 2021 is cybersecurity, followed by regulatory compliance (48%), explainability (41%) and personal privacy (41%):
“Despite the global deployment of AI, many papers on AI ethics have focused on English-language models and datasets,” the report said.
As AI systems have been deployed around the world, researchers have begun to pay more attention to the interaction between AI and reality, especially the possible harm caused by AI landing, such as racially discriminatory face recognition systems, sexist Resume screening systems, and AI clinical tools for economic income discrimination, etc.
The social prejudice displayed by AI models in the process of implementation has increased researchers’ interest in studying AI ethics, fairness, and prejudice, and motivated relevant practitioners to actively seek rescue measures.
As mentioned, the report also adds to the discussion of AI ethics. The report found:
- The problem of “bias” exhibited by language models is most pronounced, and the new data show that the larger the language model, the more prevalent the bias reflected in the training data. For example, a 280 billion parameter model developed in 2021 is 29% more toxic than a 2018 model with 117 million parameters
- Since 2014, research on fairness and transparency in AI has exploded, with a fivefold increase in related publications, research on algorithmic fairness and bias has gradually become a mainstream research topic, and industry research work in this direction has been published year-on-year 71% increase
- Multimodal models also exhibit diverse and record-setting “biases,” for example, experiments on CLIP show that black images are misclassified as non-human at more than twice the rate of other races
In addition, the report pointed out that although researchers around the world are increasingly interested in research on AI fairness, accountability and transparency, most of the papers at the related conference FAccT were written by researchers in the United States. From 2020 to 2021, the proportion of papers from North American institutions rose from 70.2% to 75.4%:
Compared with fact-checking datasets in other languages, English datasets have the highest proportion with 142, while non-English datasets have only 35 (of which Chinese datasets have only 5):
The Links: IGBT 3HAC046686-001