The Economics of Artificial Intelligence

Home > Other > The Economics of Artificial Intelligence > Page 25
The Economics of Artificial Intelligence Page 25

by Ajay Agrawal


  105

  Biolo

  Publications acr

  2006

  2015

  2006

  2015

  2006

  2015

  –

  –

  wth

  wth

  wth

  o

  –

  –

  o

  –

  –

  o

  2004

  2013

  % gr

  2004

  2013

  % gr

  2004

  2013

  % gr

  ning

  able 4.6

  obotics

  T

  Lear

  systems

  R

  Symbol

  systems

  The Impact of Artifi cial Intelligence on Innovation 139

  Table 4.7

  Herfi ndahl- Hirschman index for application sectors

  Application

  H =

  PatShare2

  Chemical applications

  153.09

  Communications

  140.87

  Hardware and software

  86.99

  Computer science peripherals

  296

  Data and storage

  366.71

  Computer science business models

  222

  Medical applications

  290.51

  Electronic applications

  114.64

  Automotive applications

  197.03

  Mechanical applications

  77.51

  Other

  129.20

  This preliminary analysis does not trace out the important knowledge spill-

  overs between innovation in the GPT and innovation and application sec-

  tors, but it is probably far too early to look for evidence of this.

  4.7 Deep Learning as a General Purpose Invention in the Method of

  Invention: Considerations for Organizations, Institutions, and Policy

  With these results in mind, we now consider the potential implications for

  innovation and innovation policy if deep learning is indeed a general pur-

  pose technology (GPT) and/or a general purpose invention in the method

  of invention (IMI). If deep learning is merely a GPT, it is likely to generate

  innovation across a range of applications (with potential for spillovers both

  back to the learning GPT and also to other application sectors), but will not

  itself change the nature of the innovation production function. If it is also

  a general purpose IMI, we would expect it to have an even larger impact

  on economy- wide innovation, growth, and productivity as dynamics play

  out—and to trigger even more severe short- run disruptions of labor markets

  and the internal structure of organizations.

  Widespread use of deep learning as a research tool implies a shift toward

  investigative approaches that use large data sets to generate predictions for

  physical and logical events that have previously resisted systematic empirical

  scrutiny. These data are likely to have three sources: prior knowledge (as in

  the case of “learning” of prior literatures by IBM’s Watson), online transac-

  tions (e.g., search or online purchasing behavior), and physical events (e.g.,

  the output from various types of sensors or geolocation data). What would

  this imply for the appropriate organization of innovation, the institutions

  we have for training and conducting research over time, and for policy, par-

  ticularly, as we think about private incentives to maintain proprietary data

  sets and application- specifi c algorithms?

  140 Iain M. Cockburn, Rebecca Henderson, and Scott Stern

  4.7.1 The Management and Organization of Innovation

  Perhaps most immediately, the rise of general purpose predictive ana-

  lytics using large data sets seems likely to result in a substitution toward

  capital and away from labor in the research production process. Many types

  of R&D and innovation more generally are eff ectively problems of labor-

  intensive search with high marginal cost per search (Evenson and Kislev

  [1976], among others). The development of deep learning holds out the

  promise of sharply reduced marginal search costs, inducing R&D organiza-

  tions to substitute away from highly skilled labor toward fi xed cost invest-

  ments in AI. These investments are likely to improve performance in existing

  “search- intensive” research projects, as well as to open up new opportuni-

  ties to investigate social and physical phenomena that have previously been

  considered intractable or even as beyond the domain of systematic scientifi c

  and empirical research.

  It is possible that the ability to substitute away from specialized labor

  and toward capital (that in principle could be rented or shared) may lower

  the “barriers to entry” in certain scientifi c or research fi elds—particularly

  those in which the necessary data and algorithms are freely available—while

  erecting new barriers to entry in other areas (e.g., by restricting access to

  data and algorithms). As of yet, there are few, if any, organized markets for

  “trained” research tools or services based on deep learning, and few stan-

  dards to evaluate alternatives. Our analysis suggests that the development

  of markets for shared AI services and the widespread availability of relevant

  data may be a necessary precursor to the broad adoption and dissemination

  of deep learning.

  At the same time, the arrival of this new research paradigm is likely to

  require a signifi cant shift in the management of innovation itself. For ex-

  ample, it is possible that the democratization of innovation will also be

  accompanied by a lack of investment by individual researchers in special-

  ized research skills and specialized expertise in any given area, reducing

  the level of theoretical or technical depth in the workforce. This shift away

  from career- oriented research trajectories toward the ability to derive new

  fi ndings based on deep learning may undermine long- term incentives for

  breakthrough research that can only be conducted by people who are at the

  research frontier. There is also the possibility that the large- scale replace-

  ment of skilled technical labor in the research sector by AI will “break

  science” in some fi elds by disrupting the career ladders and labor markets

  that support the relatively long periods of training and education required

  in many scientifi c and technical occupations.

  Finally, it is possible that deep learning will change the nature of scien-

  tifi c and technical advance itself. Many fi elds of science and engineering are

  driven by a mode of inquiry that focuses on identifying a relatively small

  The Impact of Artifi cial Intelligence on Innovation 141

  number of causal drivers of underlying phenomena built upon an under-

  lying theory (the parsimony principle as restated by Einstein states that

  theory should be “as simple as possible but no simpler.”) However, deep

  learning off ers an alternative paradigm based on the ability to predict com-

  plex multicausal phenomena using a “black box” approach that abstracts

  away from underlying causes, but does allow for a singular prediction index

  that can yield sharp insigh
t. De- emphasizing the understanding of causal

  mechanisms and abstract relationships may come at a cost: many major

  steps forward in science involve the ability to leverage an understanding

  of “big picture” theoretical structure to make sense of, or recognize the

  implications of, smaller discoveries. For example, it is easy to imagine a

  deep learning system trained on a large amount of x-ray diff raction data

  quickly “discovering” the double helix structure of DNA at very low mar-

  ginal cost, but it would likely require human judgment and insight about a

  much broader biological context to notice that the proposed structure sug-

  gests a direct mechanism for heredity.

  4.7.2 Innovation and Competition Policy and Institutions

  A second area of impact, beyond the organization of individual research

  projects or the nature of what counts as “science” in a particular fi eld, will

  be on the appropriate design and governance of institutions governing the

  innovation process. Three implications stand out.

  First, as discussed earlier, research over the past two decades has empha-

  sized the important role played by institutions that encourage cumulative

  knowledge production through low- cost independent access to research

  tools, materials, and data (Furman and Stern 2011; Murray and O’Mahony

  2007). However, to date there has only been a modest level of attention to

  the questions of transparency and replicability within the deep learning

  community. Grassroots initiatives to encourage openness organized through

  online hubs and communities support knowledge production. But it is useful

  to emphasize that there is likely to be a signifi cant gap between the private

  and social incentives to share and aggregate data—even among academic

  researchers or private- sector research communities. One implication of this

  divergence may be that to the degree any single research result depends on

  the aggregation of data from many sources, it will be important to develop

  rules of credit and attribution, as well as to develop mechanisms to replicate

  the results.

  This implies that it will be particularly important to pay attention to

  the design and enforcement of formal intellectual property rights. On the

  one hand, it will be important to think carefully about the laws that cur-

  rently surround the ownership of data. Should the data about, for example,

  my shopping and travel behavior belong to me or to the search engine or

  ride- sharing company that I use? Might consumers have a strong collective

  142 Iain M. Cockburn, Rebecca Henderson, and Scott Stern

  interest in ensuring that these data (suitably blinded, of course) are in the

  public domain so that many companies can use them in the pursuit of inno-

  vation?

  On the other hand, the advent of deep learning has signifi cant implica-

  tions for the patent system. Though there has so far been relatively little

  patenting of deep learning innovations, historical episodes such as the dis-

  covery and attempted wholesale patenting of express sequence tags and

  other kinds of genetic data suggests that breakthroughs in research tools—

  often combined with a lack of capacity at patent offi

  ces and confl icting court

  decisions—can result in long periods of uncertainty that has hampered the

  issuing of new patents, and this in turn has led to lower research productiv-

  ity and less competition. Deep learning also presents diffi

  cult questions of

  legal doctrine for patent systems that have been built around the idea of

  creative authors and inventors. For example, “inventorship” has a specifi c

  meaning in patent law, with very important implications for ownership and

  control of the claimed invention. Can an AI system be an inventor in the

  sense envisaged by the drafters of the US Constitution? Similarly, standards

  for determining the size of the inventive step required to obtain a patent

  are driven by a determination of whether the claimed invention would or

  would not be obvious to a “person having ordinary skill in the art.” Who

  this “person” might be, and what constitutes “ordinary skill” in an age of

  deep learning systems trained on proprietary data are questions well beyond

  the scope of this chapter.

  In addition to these traditional innovation policy questions, the pros-

  pect for deep learning raises a wide variety of other issues, including issues

  relating to privacy, the potential for bias (deep learning has been found to

  reinforce stereotypes already present in society), and consumer protection

  (related to areas such as search, advertising, and consumer targeting and

  monitoring). The key is that, to the extent that deep learning is general

  purpose, the issues that arise across each of these domains (and more) will

  play out across a wide variety of sectors and contexts and at a global rather

  than local level. Little analysis has been conducted that can help design

  institutions that will be responsive at the level of application sectors that

  also internalize the potential issues that may arise with the fact that deep

  learning is likely to be a GPT.

  Finally, the broad applicability of deep learning (and possibly robotics)

  across many sectors is likely to engender a race within each sector to establish

  a proprietary advantage that leverages these new approaches. As such, the

  arrival of deep learning raises issues for competition policy. In each appli-

  cation sector there is the possibility of fi rms that are able to establish an

  advantage at an early stage, and in doing so position themselves to be able to

  generate more data (about their technology, about customer behavior, about

  their organizational processes), and will be able to erect a deep- learning-

  driven barrier to entry that will ensure market dominance over at least the

  The Impact of Artifi cial Intelligence on Innovation 143

  medium term. This suggests that rules ensuring data accessibility are not

  only a matter of research productivity or aggregation, but also speak to

  the potential to guard against lock-in and anticompetitive conduct. At the

  present moment there seem to be a large number of individual companies

  attempting to take advantage of AI across a wide variety of domains (e.g.,

  there are probably more than twenty fi rms engaging in signifi cant levels of

  research in autonomous vehicles, and no fi rm has yet to show a decisive

  advantage), but this high level of activity likely refl ects an expectation for

  the prospects for signifi cant market power in the future. Ensuring that deep

  learning does not enhance monopolization and increase barriers to entry

  across a range of sectors will be a key topic going forward.

  4.8 Concluding

  Thoughts

  The purpose of this exploratory chapter has not been to provide a sys-

  tematic account or prediction of the likely impact of AI on innovation, nor

  clear guidance for policy or the management of innovation. Instead, our

  goal has been to raise a specifi c possibility—that deep learning represents a

  new general purpose invention of a method of invention—and to draw out

  some preliminary implicati
ons of that hypothesis for management, institu-

  tions, and policy.

  Our preliminary analysis highlights a few key ideas that have not been

  central to the economics and policy discussion so far. First, at least from the

  perspective of innovation, it is useful to distinguish between the signifi cant

  and important advances in fi elds such as robotics from the potential of a

  general purpose method of invention based on application of multilayered

  neural networks to large amounts of digital data to be an invention in the

  method of invention. Both the existing qualitative evidence and our pre-

  liminary empirical analysis documents a striking shift since 2009 toward

  deep learning- based application- oriented research that is consistent with

  this possibility. Second, and relatedly, the prospect of a change in the innova-

  tion process raises key issues for a range of policy and management areas,

  ranging from how to evaluate this new type of science to the potential for

  prediction methods to induce new barriers to entry across a wide range of

  industries. Proactive analysis of the appropriate private and public policy

  responses toward these breakthroughs seems like an extremely promising

  area for future research.

  144 Iain M. Cockburn, Rebecca Henderson, and Scott Stern

  Appendix

  Table 4A.1

  Artifi cial intelligence keyword allocation

  Symbols

  Learning

  Robotics

  Natural language processing

  Machine learning

  Computer vision

  Image grammars

  Neural networks

  Robot

  Pattern recognition

  Reinforcement learning

  Robots

  Image matching

  Logic theorist

  Robot systems

  Symbolic reasoning

  Bayesian belief networks

  Robotics

  Symbolic error analysis

  Unsupervised learning

  Robotic

  Pattern analysis

  Deep learning

  Collaborative systems

  Symbol processing

  Knowledge representation and reasoning

  Humanoid robotics

  Physical symbol system

  Crowdsourcing and human computation

  Sensor network

  Natural languages

  Neuromorphic computing

  Sensor networks

  Pattern analysis

 

‹ Prev