"人工智能发展历程及政策分析"

需积分: 7 164 浏览量更新于2024-03-16 收藏 1.75MB PDF 举报

The field of artificial intelligence (AI) has seen significant advancements since its inception in the 1950s. AI can now be defined as computerized systems that are capable of performing tasks that typically require human intelligence, such as visual perception, speech recognition, decision-making, and language translation. These advancements have led to a wide range of potential applications in various industries, including healthcare, finance, transportation, and cybersecurity. However, with these advancements come a plethora of ethical, legal, and policy considerations that must be carefully navigated. One of the key issues surrounding AI is the potential impact on the job market, as automation and machine learning technologies could lead to job displacement and changes in the workforce. Additionally, there are concerns about bias and discrimination in AI algorithms, as these systems can inherit biases from the data they are trained on. In response to these challenges, policymakers are considering a variety of approaches to regulate and govern the use of AI technology. Some have called for the development of ethical guidelines and standards for the responsible use of AI, while others have proposed regulatory frameworks to ensure transparency and accountability in AI systems. One of the key considerations in AI policy is the balance between promoting innovation and ensuring the protection of individual rights and societal welfare. Overall, the field of artificial intelligence holds great promise for improving efficiency, productivity, and quality of life. However, it is crucial for policymakers to carefully consider the potential risks and implications of AI technology and develop policies that promote the responsible and ethical use of these systems. By balancing innovation with safeguarding individual rights and societal values, policymakers can help ensure that AI technology benefits all members of society.

Artificial Intelligence: Background, Selected Issues, and Policy Considerations

Congressional Research Service 6

information; to learn within an environment; to abstract to create new meanings; and to reason in

order to plan and reach decisions.

First wave: handcrafted knowledge. The first wave of AI technologies have abilities primarily

to perceive and reason but no learning capability and poor handling of uncertainty. For such

technologies, researchers and engineers create sets of rules to represent knowledge in well-

defined domains for narrowly defined problems. The TurboTax software, an expert system, is one

example. Rules are built into the application, which then turns input information into tax form

outputs, but it has only a rudimentary ability to perceive and no ability to learn (e.g., about a new

tax law) or to abstract beyond what it is programmed to know.

Second wave: statistical learning. Starting in the 1990s, a second wave of AI technologies were

developed with more nuanced abilities to perceive and learn, with some ability to abstract,

minimal reasoning ability, but no contextual ability. For these systems, engineers create statistical

models for specific problem domains and train them on big data. Generally, while such systems

are statistically powerful, they can be individually unreliable, especially in the presence of

skewed training data (e.g., a face recognition system trained on a limited range of skin tones can

be powerful for similar faces, but highly unreliable for individuals outside of the training

spectrum). As noted by DARPA, these technologies are “dependent on large amounts of high

quality training data, do not adapt to changing conditions, offer limited performance guarantees,

and are unable to provide users with explanations of their results.”

Additional examples of

second wave AI technologies include voice recognition and text analysis.

Third wave: contextual adaptation. The third wave of AI technologies is oriented toward

making it possible for machines to adapt to changing situations (i.e., contextual adaptation).

Engineers create systems that construct explanatory models of real world phenomena, and “AI

systems learn and reason as they encounter new tasks and situations.”

Examples of third wave

technologies would include explainable AI (XAI), as described below.

Recent Growth in the Field of AI

There are many potential indicators of growth in the AI field. This section presents indicators of

growth based on R&D activities and public and private investments in areas of frequent

congressional interest. It also provides a brief discussion of AI hype versus the reality of what AI

technologies are capable of today. It can be challenging to obtain comprehensive and directly

comparable data for the indicators discussed in this section, particularly for AI investments.

Therefore, such data should be evaluated carefully and treated as only indicative of trends.

AI Research and Development

One way to assess the growth in AI R&D is based on the publication of peer-reviewed papers,

including both conference papers and journal articles. According to the AI Index group, between

2000 and 2019, the total number of peer-reviewed AI publications in Elsevier’s Scopus

Arati Prabhakar, former Director of DARPA, “Powerful but Limited: A DARPA Perspective on AI,” presentation at

National Academies of Sciences, Engineering, and Medicine workshop, Robotics and Artificial Intelligence: Policy

Implications for the Next Decade, December 12, 2016, at https://www.nationalacademies.org/event/12-12-2016/

robotics-and-artificial-intelligence-policy-implications-for-the-next-decade (hereinafter “Prabhakar, 2016”).

See “DARPA Announces $2 Billion Campaign to Develop Next Wave of AI Technologies,” September 7, 2108, at

https://www.darpa.mil/news-events/2018-09-07.

Prabhakar, 2016.

Artificial Intelligence: Background, Selected Issues, and Policy Considerations

Congressional Research Service 7

database—the world’s largest abstract and citation database—grew nearly 12-fold.

Authors

based in the Europe Union (EU) published the most peer-reviewed AI publications as a

percentage of the world total from 2000 to 2007 and again from 2012 to 2016, while authors

based in China published the most from 2008 to 2011 and 2017 to 2019.

In 2020, the papers

published by authors in China surpassed those of authors in the United States in the share of AI

journal citations in the world for the first time. However, over the past decade, authors in the

United States have consistently had more cited AI conference papers than authors based in

China.

Further, the number of publications a researcher or country produces does not

necessarily equate to scientific impact or research quality. As one researcher at the University of

Oxford, UK, reportedly stated, “Just pumping out raw numbers of papers that don’t have a lasting

impact isn’t really useful. It’s more important to keep up with the technology frontier.”

Such

evaluations, however, do not discuss the finer points of which studies included teams of

researchers from more than one country, raising the question of how to neatly attribute papers to

regions, organizations, or funding sources.

In addition to published papers, many AI researchers in recent years have published preprint

papers (submitted before peer review) to an online repository called arXiv (pronounced

“archive”). As reported by the AI Index group, between 2015 and 2020, the total number of AI

papers on arXiv increased over six-fold, with more growth in certain subcategories, providing a

rough indication of areas of research activity across a range of AI subfields.

As of 2020, the

most common subcategories of preprint papers were ML and computer vision (Figure 1).

Groups like the AI Index have also attempted to measure progress in AI and its fields of study,

though critics have categorized such efforts as reporting “trends in data that are related to AI”

rather than tracking progress.

Further, recent research has raised concerns about the accuracy of

reported improvements. By some measures, such as training time and cost, areas such as image

classification have improved substantially.

By other measures, researchers assert that progress

has come from tweaks, rather than core innovations, and some purported progress might not have

taken place. For example, some researchers using meta-analyses of algorithms in various fields

and applications—such as pruning algorithms used to make neural networks more efficient and

information retrieval programs used in search engines—have found no clear evidence of

performance improvements over the 10-year period from 2010 to 2019.

AI Index Steering Committee, The AI Index 2021 Annual Report, Human-Centered AI Institute, Stanford University,

Stanford, CA, March 2021, p. 18 (hereinafter, “AI Index 2021”). The AI Index 2021 report authors provided this

information only for the United States, China, and the Europe Union (EU), not individual countries within the EU.

Ibid., p. 20.

Ibid., p. 17.

Neil Savage, “The Race to the Top Among the World’s Leaders in Artificial Intelligence,” Nature Index, December

9, 2020, at https://www.nature.com/articles/d41586-020-03409-8.

Ibid., p. 32.

Ibid., p. 34.

Jeffrey Funk and Gary Smith, “Stanford’s AI Index Report: How Much Is BS?,” Mind Matters News, March 3, 2020,

at https://mindmatters.ai/2020/03/stanfords-ai-index-report-how-much-is-bs/.

AI Index 2021, pp. 48-49. Image classification broadly refers to the assigning of identification labels to images.

Matthew Hutson, “Core Progress in AI Has Stalled in Some Fields,” Science, vol. 368, no. 6494 (May 29, 2020), p.

927, at https://science.sciencemag.org/content/368/6494/927.

剩余46页未读，继续阅读

承让@

粉丝: 8
资源: 380

"人工智能发展历程及政策分析"

人工智能研究中的若干问题

人工智能 员工选择问题

人工智能的历史与背景

PIPEDA 改革以解决人工智能问题的政策建议-研究论文

人工智能背景下的Python教学策略研究.pdf

人工智能背景下中小企业财务分析探析.pdf

人工智能背景下的马克思主义劳动价值论.pdf

人工智能背景下高技术产业技术创新能力研究.docx

人工智能背景下职业教育师资队伍建设的思考.pdf

人工智能行业税收政策建议.zip

最新资源

人工智能员工选择问题