Click on " Study Algorithm " above, and choose to add " star " or " top "
Heavy dry goods, delivered as soon as possible
Source丨New Wisdom
Edit丨Gokushi Platform
Guided reading
The 2022 AI Index report is out! In this report, China performed well in AI top conference papers, but was lower than the US, EU and UK in terms of citations.
Today, Stanford University released its 2022 Artificial Intelligence Index report.
Professor Li Feifei forwarded the report as soon as it was released.
This year's report is divided into 5 major chapters: Research and Development, Technology Performance, Ethical Challenges of AI Applications, Economics and Education, AI Policy and National Strategy.
Here are 7 report highlights for you:
Sino-US cooperative papers ranked first in the world
In the past 10 years, the number of global AI papers published has doubled, from 162,444 in 2010 to 334,497, and it is increasing year by year.
Specifically, papers in the field of pattern recognition and machine learning have doubled in the six years from 2015 to 2021, while other fields such as computer vision, data mining, and natural language processing have maintained relatively stable development.
In addition, from the perspective of the publication of papers, journal articles account for the largest proportion, 51.5%; 21.5% are top conference papers, and 17.0% are from repositories.
It can be seen that over the past 12 years, the number of papers in journals and repositories has grown by a factor of 2.5 and 30, respectively, but the number of top conference papers has declined since 2018.
In terms of transnational cooperation of papers, from 2010 to 2021, China and the United States jointly published the largest number of artificial intelligence papers in the world, which has increased fivefold since 2010. The number of publications cooperating between China and the United States is 2.7 compared with that of China and the United Kingdom, ranking second in the world.
China dominates the top AI conference, but the number of citations is the lowest
In 2021, China will continue to lead the world in the number of AI journals, top conferences and knowledge base publications. The three publication types combined are 63.2% higher than the US.
At the same time, the United States leads in the number of AI top conference papers and the number of repository citations.
In terms of the number of papers published in AI journals, China has always dominated the list in terms of the number of papers published in AI journals in the past 12 years, with 31.0% in 2021 (18.0% in 2020), followed by the EU and the UK with 19.1%, In the United States it was 13.7%.
In 2021, China will still lead the world in the number of citations in AI journals.
It is worth noting that, in terms of the number of AI journal papers published and the number of citations, the United States dropped from the second place last year to the third place.
So how is the situation of China and the United States publishing papers at the summit?
In 2021, China will have the largest share of the number of papers published in the world's top AI conferences with 27.6%, a larger lead than in 2020, while the EU and the UK will follow with 19.0% and the US with 16.9%. 3.
However, the United States has always maintained a high level of citations in AI top conference papers. In 2021, the total citations will account for 29.52%. The second and third places are the EU, the United Kingdom (23.32%) and China (15.32%).
China fell from the second place last year to the third place. It can be seen from the side that the number of Chinese papers published is the most, but the quality is not as high as that of the United States.
Overall, the number of AI patent applications in 2021 is more than 30 times that of 2015, with a compound annual growth rate of 76.9%.
Specifically, China has applied for more than half of the world's AI patents and has been granted about 6%, roughly the same as the EU and the UK.
Compared with the growing number of AI patent applications and grants, the number of patent applications in China (87,343 in 2021) is much higher than the number of grants (1,407 in 2021).
The most popular GitHub open source library: TensorFlow
As can be seen from the number of users of GitHub's open source AI software library from 2015 to 2021, TensorFlow is still the most popular in 2021, and the cumulative number of GitHub stars is about 161,000, a slight increase from 2020.
In second place is OpenCV, followed by Keras, PyTorch and Scikit-learn
Brush your face with a mask
Let’s start with a set of photos showing the chronological development of the level of face generation.
Compared with 2014, we can only generate a black and white and blurred face from a real person with rich skin tones and expressions, but in 2021, the computer can also reveal more details of black skin, and we can see the black inner band of the skin of the image. Brown, and a grin on his face.
The report shows how computer images are classified. The image below includes various categories such as airplanes, autonomous cars, birds, cats, deer, dogs, frogs, horses, boats, trucks, etc. The image is improved by a classification model and target label for a given image. The intelligence of recognition.
The report pointed out that with the application of deep learning to AI images, the accuracy of image classification has greatly improved.
The following is a comparison chart of the accuracy of AI image recognition. The blue line represents no training data is used, and the green line represents the use of training data. Obviously, the green line (99.02% accuracy) surpasses the blue line (97.9% accuracy), which means that the accuracy of AI image recognition has improved after data training.
Regardless of whether deep training data is used or not, AI image classification (99.02% & 97.90%) has shown a higher level than ordinary people (94.9% accuracy) after 2017.
It seems that "face blindness" only exists in humans, and AI will hardly suffer from it.
The report uses a mask face image set from Beijing University of Posts and Telecommunications. The 6,000 face recognition data set has improved the accuracy of face recognition during the epidemic.
The mask covers the face, which greatly reduces the facial information collected by the face recognition system. However, the AI team from China focused the key information of face recognition on the eyebrows and eyes, and used the correct model for training to achieve " You can even brush your face while wearing a mask.”
Visual common sense reasoning is a bit low, 72 points behind humans
Let's take a look at how computers perform visual commonsense reasoning.
Visual Commonsense Reasoning (VCR) is a frontier hot issue in the AI field. It is a very challenging task, including cognition, learning, reasoning, and data processing from a single visual question answering, image recognition, motion capture, etc. , rising to "cross-media intelligence," representing a new benchmark for computer vision understanding.
The report pointed out that at present, the common sense reasoning ability of computer vision is far behind humans, and the human visual common sense reasoning level has been maintained at 85 points, while the best score of machines in 2021 is only 72 points.
It seems that this is indeed a bit low, but compared to the failing level in 2018 (43 points), the machine has improved by 29 points in 3 years, which is obviously a big step forward.
Natural language processing starts from a high starting point and grows slowly, with an accuracy of 93.1%
In recent years, thanks to the development of speech recognition technology, the service application of complete machine translation has increased significantly, accounting for 46%.
Compared with other applications, the commercial application of machine translation has grown significantly, and the commercial scale has expanded from 21% in 2019 to 38% in 2021, nearly doubling.
The report uses questions and labels for natural language processing provided by Bowman et al., Stanford University, 2015.
Natural language processing is to assume 4 kinds of logic under the premise of given certain tasks, namely error (contradiction contradiction), undecided (neutral neutral), whether it is true (entailment implication), the machine is reasonable and unreasonable. inference.
The report shows that the accuracy of natural language processing in 2021 has reached 93.1%, and it has reached 90% in early 2017, which is a slow growth based on a high starting point.
The Stanford Natural Language Inference (SNLI) dataset contains about 600,000 pairs of labeled sentences, and its performance accuracy is based on the percentage of correctly answered questions.
Other indicators
In terms of the overall volume of this year's report, compared to last year's 7 chapters, this year's report is condensed into 5 chapters, reducing the "AI diversity" part, and combining "AI economy" and "AI education" into "AI economy" and "AI education". Economics and Education" chapter 1.
Although the total number of chapters has decreased, the volume has increased, and the page number of the table of contents has increased from 177 to 196 pages.
Demand for AI workforces has grown significantly over the past nine years in the six countries covered by the Burning Glass data. Among them, AI recruitment positions in Singapore accounted for 2.33% of the total recruitment positions, ranking first, and the United States ranked second with 0.90%.
In terms of investment in artificial intelligence, from 2013 to 2021, private investment in artificial intelligence companies in the United States was more than twice that of China.
AI investment in China has grown, from 10% in 2020 to 17.21% in 2021.
Report download address:
file:///C:/Users/62589/Desktop/2022-AI-Index-Report_Master.pdf
Click to see the paper constantly!