2016人工智能生态报告（英文版）.pdf

《2016人工智能生态报告（英文版）.pdf》由会员分享，可在线阅读，更多相关《2016人工智能生态报告（英文版）.pdf（99页珍藏版）》请在淘文阁 - 分享文档赚钱的网站上搜索。

1、Artificial intelligence is the apex technology of the information era. In the latest in our Profiles in Innovation series, we examine how advances in machine learning and deep learning have combined with more powerful computing and an ever-expanding pool of data to bring AI within reach for companie

2、s across industries. The development of AI-as-a-service has the potential to open new markets and disrupt the playing field in cloud computing. We believe the ability to leverage AI will become a defining attribute of competitive advantage for companies in coming years and will usher in a resurgence

3、 in productivity.Heath P . Terry, CFA (212) 357-Goldman, Sachs & Co.Goldman Sachs does and seeks to do business with companies covered in its research reports. As aresult, investors should be aware that the firm may have a conflict of interest that could affect theobjectivity of this report. Investo

4、rs should consider this report as only a single factor in making theirinvestment decision. For Reg AC certification and other important disclosures, see the DisclosureAppendix, or go to Analysts employed by non-US affiliates are notregistered/qualified as research analysts with FINRA in the U.S.The

5、Goldman Sachs Group, Inc.EQUITY RESEARCH | November 14, 2016PROFILES ININNOVATIONJesse Hulsing (415) 249-Goldman, Sachs & Co.Mark Grant (212) 357-Goldman, Sachs & Co.Daniel Powell (917) 343-Goldman, Sachs & Co.AI, Machine Learning and Data Fuel the Future of ProductivityArtificial IntelligencePiyush

6、 Mubayi(852) 2978-Goldman Sachs (Asia) L.L.C.Waqar Syed(212) 357-Goldman, Sachs & Co.November 14, 2016 Profiles in Innovation Goldman Sachs Global Investment Research 2 Contents Portfolio Manager s summary 3 What is Artificial Intelligence? 9 Key drivers of value creation 11 Fueling the future of pr

7、oductivity 15 AI & The Productivity Paradox: An interview with Jan Hatzius 18 The Ecosystem: Cloud services, open source key beneficiaries of the coming investment cycle in AI 20 Use Cases 41 Agriculture 42 Financials 50 Healthcare 59 Retail 68 Energy 75 Enablers 83 Appendix 90 Disclosure Appendix 9

8、7 Contributing Authors: Heath P. Terry, CFA, Jesse Hulsing, Robert D. Boroujerdi, Jan Hatzius, Piyush Mubayi, Mark Grant, Daniel Powell, Waqar Syed, Adam Hotchkiss, Komal Makkar, Yena Jeon, Toshiya Hari, Heather Bellini, CFA, Simona Jankowski, CFA, Matthew J. Fassler, Terence Flynn, PhD, Jerry Revic

9、h, CFA, Salveen Richter, CFA, Rob Joyce, Charles Long This is the seventh report in our Profiles in Innovation series analyzing emerging technologies that arecreating profit pools and disrupting old ones. Access previous reports in the series below or visit our portal for more, including a video sum

10、mary of this report. Virtual and Augmented RealityDronesFactory of the FutureBlockchainPrecision FarmingAdvanced MaterialsNovember 14, 2016 Profiles in Innovation Goldman Sachs Global Investment Research 3 Portfolio Manager s summary Artificial Intelligence (AI) is the apex technology of the informa

11、tion age. The leap from computing built on the foundation of humans telling computers how to act, to computing built on the foundation of computers learning how to act has significant implications for every industry. While this moment in time may be viewed as the latest cycle of promise and disappoi

12、ntment before the next AI Winter (Exhibit 8), these investments and new technologies will at the very least leave us with the tangible economic benefit to productivity of machine learning. In the meantime, AI, bots, and self-driving cars have risen to the forefront of popular culture and even politi

13、cal discourse. However, our research over the last year leads us to believe that this is not a false start, but an inflection point. As we shall explore in this report, the reasons for the inflection range from the obvious (more and faster compute and an explosion of more data) to the more nuanced (

14、significant strides in deep learning, specialized hardware, and the rise of open source). One of the more exciting aspects of the AI inflection is that “real-world” use cases abound. While deep-learning enabled advances in computer vision and such technologies as natural language processing are dram

15、atically improving the quality of Apples Siri, Amazons Alexa, and Googles photo recognition, AI is not just “tech for tech”. Where large data sets are combined with powerful enough technology, value is being created and competitive advantage is being gained. For example, in healthcare, image recogni

16、tion technology can improve the accuracy of cancer diagnosis. In agriculture, farmers and seed producers can utilize deep learning techniques to improve crop yields. In pharmaceuticals, deep learning is used to improve drug discovery. In energy, exploration effectiveness is being improved and equipm

17、ent availability is being increased. In financial services, costs are being lowered and returns increased by opening up new data sets to faster analysis than previously possible. AI is in the very early stages of use case discovery, and as the necessary technology is democratized through cloud based

18、 services we believe a wave of innovation will follow, creating new winners and losers in every industry. The broad applicability of AI also leads us to the conclusion that it is a needle-moving technology for the global economy and a driver behind improving productivity and ending the period of sta

19、gnant productivity growth in the US. Leveraging the research of Chief GS economist Jan Hatzius, we frame the current stagnation in capital deepening and its associated impact on US productivity. We believe that AI technology driven improvements to productivity could, similar to the 1990 s, drive cor

20、porates to invest in more capital and labor intensive projects, accelerating growth, improving profitability, and expanding equity valuations. Implications While we see artificial intelligence impacting every corporation, industry, and segment of the economy in time, there are four implications for

21、investors that we see as among the most notable. Productivity. AI and machine learning (ML) has the potential to set off a cycle of productivity growth that benefits economic growth, corporate profitability, returns on capital, and asset valuations. According to GS Chief Economist Jan Hatzius “In pr

22、inciple, it AI does seem like something that could be potentially captured better in the statistics than the last wave of innovation to the extent that artificial intelligence reduces costs and See profiles of 5 real-world use cases for AI on pp. 41 to 81. We interview GS Chief Economist Jan Hatzius

23、 about the impact AI/machine learning could have on lagging US productivity growth on p. 18. November 14, 2016 Profiles in Innovation Goldman Sachs Global Investment Research 4 reduces the need for labor input into high value added types of production. Those cost saving innovations in the business s

24、ector are things statisticians are probably better set up to capture than increases in variety and availability of apps for the iPhone, for example. To the extent Artificial Intelligence has a broad based impact on cost structures in the business sector, Im reasonably confident that it would be pick

25、ed up by statisticians and would show up in the overall productivity numbers.” Premium technology. The value of speed in AI and machine learning has the potential to reverse the trend towards cheaper commodity hardware in building data centers and networks. We believe this could drive substantial sh

26、ifts in market share in hardware, software, and services spending. For example, an AWS workload running on a “standard” datacenter compute instance costs as little as $0.0065/hour compared to $0.900 for a GPU instance optimized for AI. Competitive Advantage. We see the potential for AI and machine l

27、earning to reshuffle the competitive order across every industry. Management teams that fail to invest in and leverage these technologies risk being passed by competitors that benefit from the strategic intelligence, productivity gains, and capital efficiencies they create. In the vignettes starting

28、 on page 41 we examine how these competitive advantages are developing in Healthcare, Energy, Retail, Finance and Agriculture. New Company Creation. We have identified over 150 private companies in the AI and ML space founded over the last 10 years (Exhibits 69-75). While we believe that much of the

29、 value in AI will accrue to large companies with the resources, data, and ability to invest, we expect that venture capitalists, entrepreneurs and technologists will continue to drive the creation of new companies that will, in turn, drive substantial innovation and value creation through, at the ve

30、ry least, M&A, though we certainly wouldnt dismiss the potential for a “Google or Facebook of AI” to emerge. In the following pages we delve into AI the technology, its history, the ecosystem being created around machine learning, applications for these technologies across industries and the compani

31、es that are leading the way. What is AI? AI is the science and engineering of making intelligent machines and computer programs capable of learning and problem solving in ways that normally require human intelligence. Classically, these include natural language processing and translation, visual per

32、ception and pattern recognition, and decision making, but the number and complexity of applications is rapidly expanding. In this report, we will focus most of our analysis on machine learning, a branch of AI, and deep learning, a branch of machine learning. We highlight two key points: 1.Simplistic

33、ally, machine learning is algorithms that learn from examples andexperience (i.e., data sets) rather than relying on hard-coded and predefined rules.In other words, rather than a developer telling a program how to distinguishbetween an apple and an orange, an algorithm is fed data (“trained”) and le

34、arnson its own how to distinguish between an apple and an orange.2.Major advances in deep learning are one of the driving forces behind the currentAI inflection point. Deep learning is a sub-set of machine learning. In mosttraditional machine learning approaches, features (i.e., the inputs or attrib

35、utes thatmay be predictive) are designed by humans. Feature engineering is a bottleneckand requires significant expertise. In unsupervised deep learning, the importantfeatures are not predefined by humans, but learned and created by the algorithm.We profile the ecosystem of public and private compan

36、ies enabling the AI revolution on pp. 83 to 89. November 14, 2016 Profiles in Innovation Goldman Sachs Global Investment Research 5 To be clear, were not yet focusing on the kind of True, Strong, or General Artificial Intelligence that is meant to replicate independent human intelligence, and that i

37、s often the AI in popular culture. While there have been certain potential breakthroughs there, like Google DeepMinds AlphaGo system, which not only defeated a Go world champion, but did so using moves no human ever had before, we focus on the more immediately economically tangible areas of developm

38、ent in artificial intelligence. Why is AI development accelerating now? Major leaps in deep learning capabilities have been one of the catalysts behind the AI inflection currently underway. Neural networks, the underlying technology framework behind deep learning, have been around for decades, but t

39、hree things have changed over the last five to ten years: 1.Data. There has been massive growth in the amount of unstructured data beingcreated by the increasingly ubiquitous connected devices, machines, and systemsglobally. Neural networks become more effective the more data that they have,meaning

40、that as the amount of data increases the number of problems thatmachine learning can solve using that data increases. Mobile, IoT, and maturationof low cost data storage and processing technologies (often in the cloud) hascreated massive growth in the number, size, and structure of available data se

41、ts.For example, Tesla has aggregated 780mn miles of driving data to date, andadding another million miles every ten hours through its connected cars, whileJasper (acquired by Cisco for $1.4bn in Feb. 2016) has a platform poweringmachine to machine communication for multiple automobile manufacturers

42、andtelco companies. Verizon made a similar investment in August when it announcedit would acquire Fleetmatics, which connects remote sensors on vehicles to cloudsoftware via increasingly fast wireless networks. The rollout of 5G will onlyaccelerate the rate at which data can be generated and transmi

43、tted. Annual datageneration is expected to reach 44 zettabytes (trillions of GB) by 2020, according toIDCs Digital Universe report, a CAGR of 141% over five years, suggesting that weare just beginning to see the use cases to which these technologies will be applied.Exhibit 1: Annual data generation

44、is expected to reach 44 zettabytes (44 trillion GB) by 2020, according to EMC/IDC annual data generation globally (in ZB) Source: EMC, IDC 051015202530354045502009201020152020AnnualDataGenerationNovember 14, 2016 Profiles in Innovation Goldman Sachs Global Investment Research 6 2.Faster hardware. Th

45、e repurposing of Graphic Processing Units (GPUs), thegeneral availability of lower cost compute power, particularly through cloudservices, and new models for building neural networks have dramaticallyincreased the speed and accuracy of the results neural networks can produce.GPUs and their parallel

46、architecture allow for faster training of machine learningsystems compared to the traditional Central Processing Unit (CPU) based datacenter architecture. By repurposing graphics chips networks can iterate faster,leading to more accurate training in shorter periods of time. At the same time, thedeve

47、lopment of specialized silicon, like the Field Programmable Gate Arrays beingused by Microsoft and Baidu, allows for faster inference by trained deep learningsystems. More broadly, the raw compute power of super computers has increasedexponentially since 1993 (Exhibit 2). In 2016, a single high-end

48、Nvidia video cardfor a gaming PC has sufficient compute power to have classified as the mostpowerful super computer in the world before 2002.Exhibit 2: Raw compute performance of global supercomputers, measured in GFLOPs, has increased exponentially since 1993 Rpeak GFLOPS of #1 ranked global superc

49、omputers on the Top500 list Source: top500.org, compiled by Goldman Sachs Global Investment Research Costs of performance have also declined drastically. Nvidias GPU (GTX 1080) delivers 9TFLOPS of performance for roughly $700, implying a price per GFLOPS of roughly 8 cents. In 1961, stringing togeth

50、er enough IBM 1620s to deliver a single GFLOPS of performance would require over $9 trillion (adjusted for inflation). 1101001,00010,000100,0001,000,00010,000,000100,000,0006/1/19938/1/199410/1/199512/1/19962/1/19984/1/19996/1/20008/1/200110/1/200212/1/20032/1/20054/1/20066/1/20078/1/200810/1/200912