University of Cambridge:
2020 AI Panorama Report
(Full version.pdf )
Only a few of them are shown below
O background reply "20201005" Due to the large demand for today's headline background, there is no reply, please VOX search "artificial intelligence industry chain alliance" to reply.
Text total: 5233 words - 25 figures
Estimated reading time: 14 minutes
NeurIPS accepts papers, 29 percent of authors have undergraduate degrees from Chinese universities, but 54 percent of them go to the U.S. to pursue a graduate ph.D. after graduation, and 90 percent of them choose to work in the U.S. What new observations in the FIELD of AI are worth noting in the 2020 edition of the University of Cambridge's 2020 AI Panorama Report?
Artificial intelligence is a technical field that combines basic science and engineering practice, and in recent years it has integrated more and more other directions. In today's digitalization, AI will play a role in driving technological progress.
Recently, the University of Cambridge's 2020 edition of the "AI Panorama Report" was finally released, which is the third issue of the annual report. As in previous years, the report cites data from well-known tech companies and research groups. The new AI Panorama report presents the recent trends in the field of artificial intelligence in several aspects: research, talent, industry, policy and future prospects.
The survey's two lead authors, Nathan Benaich and Ian Hogarth, are both from Cambridge University.
First, the progress of artificial intelligence research
Although the development of the field of artificial intelligence is accompanied by open source frameworks and active communities, at the beginning of this year's report, we still need to speak in terms of data, first of all, to make it clear: THE openness of AI research is not as high as we think.
On deep learning frameworks, as we've recently felt, researchers have made extensive use of PyTorch in various AI-cap papers, which has preempted much of the original TensorFlow.
On GitHub, PyTorch also has more new research implementations than TensorFlow: According to statistics, about 47% of implementations are now based on PyTorch, compared with about 18% of TensorFlow.
On the other hand, large-scale models are driving technological advances in the field of NLP, and new research such as OpenAI's GPT-3 has pushed the number of parameters of deep learning models to hundreds of billions. Based on the current cloud service hashrate price, training a model per 1,000 parameters costs an average of $1, and GPT-3 with 175 billion parameters can cost millions of dollars, a figure that some experts believe is more than $10 million. The high cost of training has left researchers challenged in exploring new directions.
While AI model training requires more and more computing power, traditional computer architecture is gradually approaching the end of Moore's Law. Research from universities such as MIT says scientists could need tens of billions of dollars if they want to reduce the error rate of image classification tasks in the ImageNet dataset from 11.5 percent to 1 percent.
However, people are also studying ways to improve the efficiency of models, and OpenAI statistics show that since 2012, the computing power required to train deep learning neural networks to classify ImagesNet images to achieve a certain level has halved every 16 months.
There is no doubt that models such as GPT-3 and BERT have brought research in the field of NLP to a new stage. Unsupervised machine translation tools for automatic translation of programming languages are now even emerging. Translate C++ functions to Java on GitHub with 90% accuracy.
The rapid development of technology stems from the high research intensity in the field of artificial intelligence. According to statistics, the number of papers on AI methods (deep learning, NLP, computer vision, reinforcement learning, etc.) around the world has increased by 50% per year since 2017, and in 2020, we may see more than 21,000 new papers in the field of AI.
However, most current machine learning applications function through statistics, ignoring an important method of human learning knowledge - causal reasoning. In tasks such as finding a diagnosis and treatment plan for patients, causal reasoning is a better approach. Ai pioneers such as Judea Pearl and Yoshua Bengio believe that causal reasoning is a new direction that makes machine learning systems better generalized, more robust, and more responsive to decision-making.
Second, AI talent: the United States dominates
The distribution of researchers in the field of artificial intelligence has shown several new trends in recent years.
Talent mobility
First, academia is facing a brain drain, with many research professors in the United States leaving universities for tech companies. From 2004 to 2018, Google, DeepMind, Amazon, and Microsoft have hired 52 tenured professors from American universities. Carnegie Mellon University, the University of Washington, and the University of Berkeley lost 38 professors over the same period. Notably, 41 AI professors left in 2018 alone.
From the perspective of ai summit, researchers who have had educational experience in China have made outstanding contributions to research in the field of AI in recent years. In the case of NeurIPS 2019, 29% of the authors who received papers have obtained undergraduate degrees in China.
But after graduating from domestic college, 54 percent of graduates who continue to publish papers on NeurIPS go to the United States.
In the field of artificial intelligence, the United States is still the center of international research, and 90% of the Doctors who graduated in the United States will stay in the United States to continue their work.
Non-U.S. AI PhD graduates are more likely to graduate for large tech companies, while U.S. PhD graduates are more likely to work for startups or join the academic research ranks.
At the same time, many U.S. PhD graduates in AI graduates will graduate in the U.K. and China.
Next, let's analyze the overview of the AI field from the perspective of research institutions.
In the case of NeurIPS 2019, Google, Stanford, Carnegie Mellon University, MIT, and Microsoft published the top five papers.
Talent is in short supply
As one of the hottest research areas at present, the demand for talents in the FIELD of AI is growing. Many leading universities are also expanding their ai enrollment. In stanford, for example, in recent years, the number of students in stanford's AI field has been ten times higher than in 1999-2004, and the number of students in AI has also doubled compared with 2012-2014. Still, data from Indeed shows that the number of jobs hired is still about three times the number of job seekers.
But it is inevitable that the talent market in the field of artificial intelligence in 2020 will be severely affected by the new crown epidemic. According to data released by LinkedIn, the originally strong growth trend of machine learning jobs in 2020 was hit in February and began to decline.
Third, the rapid development of the industry
Artificial intelligence-designed drugs have begun phase I clinical trials in Japan. In the field of artificial intelligence healthcare, many startups have collected huge amounts of money to achieve a "platform strategy".
During the COVID-19 epidemic, many technology companies have also quickly put AI medical image recognition technology into practical use. Recently, the U.S. Centers for Medicare and Medicaid Services has proposed a cost standard for deep learning-based medical imaging products. AI systems can quickly scan a variety of medical images such as chest X-ray and submit screening results to human experts to rule out non-sensitive factors.
When it comes to artificial intelligence, the most interesting autonomous driving. Since 2018, of the 66 companies in California that have licensed testing for self-driving cars, only 3 have been allowed to test without a safe driver: Waymo (Google), Nuro, and AutoX.
Even in California, where the policies are most open, the mileage of self-driving cars has so far been insignificant compared to that of humans — self-driving car companies in 2019 increased their self-driving miles by 42% compared to 2018. But that's only equivalent to 0.000737% of the miles driven by a california driver with a license in 2019.
Before the use of each manual intervention, the mileage of the car's automatic driving as a criterion for judging is not necessarily the most objective. In some U.S. states, the mileage traveled by a driver with his hands completely off the steering wheel is not recorded.
Recently, though, we've seen a new twist to this data. Baidu's self-driving has reached between 18,050 miles per human intervention, surpassing Waymo's 13,219 miles. For Baidu, which is constantly strengthening its AI investment, recent investment has begun to pay off.
Companies in the field of autonomous driving must have strong financial support. Zoox, which was acquired by Amazon for $1.3 billion, has raised more than $955 million in financing since 2015, and Zoox's latest valuation is about $3.2 billion. Transaction documents show Zoox is burning $30 million a month in early 2020.
Domestic mobility company Didi also recently spun off its autonomous driving business and raised $500 million from institutions such as the SoftBank Vision Fund. In July, Didi launched its self-driving car service in Shanghai.
Currently, most machine learning algorithms in autonomous driving systems focus only on what's around the vehicle and are based on handwritten rules that are massively engineered. Researchers are developing new algorithms similar to AlphaGo, which learn a lot of human driving experience for training. Recently, Waymo, Uber, and Lyft have all demonstrated new technologies for imitation learning and inverse reinforcement learning.
The development of autonomous driving and other fields also requires a lot of computing power, and the new generation of chips launched by companies such as Graphcore and NVIDIA this year has become people's hope.
Policy changes
In addition to the research direction of AI, the rapid landing of artificial intelligence applications has also caused people to worry about privacy and ethics.
Face recognition technology is facing unprecedented controversy
Currently, the use of facial recognition technology is allowed in 50% of the world, and only 3 countries (Belgium, Luxembourg, Morocco) partially ban the technology, only in certain situations.
Those head technology companies are also more cautious about the use of face recognition technology:
Microsoft removed its database of 10 million faces — the largest database currently available. The faces in the database were crawled from the Internet without the consent of the person concerned.
Amazon announced a one-year moratorium on police use of its facial recognition tool, Rekognition, to allow "Congress enough time to enact appropriate regulations."
IBM announced the abandonment of its facial recognition products and technology.
The New York Metropolitan Transportation Administration (MTA) asked Apple to enable FaceID when allowing passengers to wear masks to prevent the spread of the coronavirus.
Since the beginning of the year, the controversy faced by facial recognition technology has been more turbulent than ever.
The U.S. continues to invest heavily in military AI systems
With the advent of machine learning technology, the military is also exploring more and more in this area, although we do not yet know the extent to which this trend will affect the real world.
The U.S. General Services Administration and the U.S. Joint Center for Artificial Intelligence of the Department of Defense awarded Booz Allen Allen Consulting a five-year, $800 million order with keywords such as "data labeling, data management, and artificial intelligence product development."
At the defense level, there are many more AI companies related to this that are receiving lucrative government contracts and venture capital. Pivotal Software, a subsidiary of Dell, has received a $121 million contract from the U.S. Department of Defense, and companies engaged in drones, high-resolution satellite maps, information management, and other businesses have received significant venture capital, such as Anduril, Rebellion, skydio, and so on.
The U.S. Defense Advanced Research Projects Agency (DARPA) organized a virtual air combat competition in which AI systems competed against each other, and the winning "contestant", the top AI developed by Heron Systems, beat the human pilot with a score of 5:0.
From AlphaGo and AlphaStar to AlphaDogfight, ai is beating the top human players in more areas with the help of deep reinforcement learning technology. This also fully demonstrates that the winning techniques trained in the game versus environment can be quickly migrated to the military environment. The defeated pilot said, "As a fighter pilot, our usual standard operating methods are no longer working."
U.S. Defense Secretary Mark T. Esper said the algorithms, trained in simulated combat environments, will be applied to real-world warfare in 2024, including full-size tactical aircraft. Machine learning will have a structural impact on the wars of the world of the future, noting that "those who take advantage of the latest iterations of technology will often have a decisive advantage on the battlefields of the future."
The two ai tops will adopt a new code of ethics
Both NeurIPS and ICLR propose new ethical norms, but do not mandate code and data sharing. Take NeurIPS, the top conference in the field of artificial intelligence, as an example:
NeurIPS will create a dedicated sub-team of experts at the intersection of machine learning and ethics.
NeurIPS now requires authors to submit information about "the broader impact that the work may have, including ethical aspects as well as future societal implications".
Given the growing influence of companies such as Facebook and Google in NeurIPS, "authors must provide clear disclosure of the source of funding and the points of interest in competition."
NeurIPS "strongly encourages" the sharing of data and models, but there is no mandatory provision.
In this regard, machine learning lags behind the life sciences, for example, one of the conditions for publishing papers in nature journals is that authors must "provide materials, data, code, and related protocols to readers in a timely manner."
Huawei's dominance in the smartphone space has increased and it has invested heavily in machine learning technologies
It was also the first time in nine years that a company other than Apple and Samsung had taken the lead in the market, but by mid-September 2020, Huawei was facing a chip supply dilemma due to U.S. sanctions.
Foreign companies that use U.S. chip manufacturing equipment must obtain permission from the U.S. government to provide certain chips to Huawei. Yu Chengdong, CEO of Huawei's consumer business, said at the 2020 Summit of the China Informatization 100 Association: "Our mobile phone business is now very difficult, the chip supply is difficult, and it is very out of stock."
Huawei's Kirin AI chip was foundry by TSMC, which was affected by the US sanctions policy, and the last order accepted by TSMC was on May 15, 2020. At present, Huawei is trying to seek support from SMIC in chip manufacturing.
But TSMC still dominates the industry in R&D spending and semiconductor manufacturing. TSMC's R&D expenses are comparable to SMIC's revenue, the former is currently the only manufacturer of commercial 5nm manufacturing processes (N5) and is currently working on a 3nm manufacturing process with 2 times more power efficiency than 7nm and 33% more performance than 7nm. Not long ago, SMIC also said it would increase capital expenditures to $6.7 billion in 2020 (up from the original target of $3.1 billion).
Future Outlook: 8 Trends
In the final section, the Cambridge 2020 AI Panorama Report gives 8 trend predictions for the next 12 months.
1. The race to build larger language models will continue, and we will witness the birth of the first 10 trillion parametric model.
2. Attention-based neural networks will migrate from the NLP domain to the CV domain to implement the new SOTA.
3. With the adjustment of the parent company's strategy, the AI lab of a large enterprise is about to close.
4. In response to U.S. Department of Defense activities and financing from U.S. military AI startups, a number of Chinese and European defense AI companies will raise more than $100 million over the next 12 months.
5) A leading AI drug discovery startup (such as Recursion, Exscientia) either entered an IPO or was acquired for more than $1 billion.
6. DeepMind will make major breakthroughs in structural biology and drug discovery.
7) Facebook will make a major breakthrough in AR and VR with 3D computer vision technology.
8. NVIDIA will not eventually complete the acquisition of Arm.
recommend:
<h1 class="pgc-h-arrow-right" data-track="187" > [Chinese Style Animation] "Jiang Ziya" behind the brush screen, hiding the inside story of Chinese animation for 100 years</h1>
<h1 class="pgc-h-arrow-right" data-track="188" > [Chinese Style Animation] In addition to "Nezha", these conscientious domestic animations should also be known by more people! </h1>
<h1 class="pgc-h-arrow-right" data-track="189" > [Chinese style animation] "Five Elements of the Misty Mountains" fire, but few people know its predecessor "Year City Glass Heart" A male protagonist with a ten-meter knife died prematurely! </h1>
For more reports
join
"Artificial Intelligence Industry Chain Alliance"
Knowledge Planet, get the relevant reports!
Screenshot of the report section
statement
Source: University of Cambridge, artificial intelligence industry chain alliance recommended to read, does not represent the position of the artificial intelligence industry chain alliance, reprint please indicate, such as involving the copyright of the work, please contact us to delete or do related processing!
EDIT: Zero