Book Read Free

The Economics of Artificial Intelligence

Page 29

by Ajay Agrawal


  5.4 A Combinatorial- Based Knowledge Production

  Function with Team Production: An Extended Model

  Our basic model assumes that researchers working alone combine the

  knowledge to which they have access, A, to discover new knowledge. In

  reality, new discoveries are increasingly being made by research teams (Jones

  2009; Nielsen 2012; Agrawal, Goldfarb, and Teodoridis 2016). Assuming

  5. K elements in Kauff man’s original notation.

  162 Ajay Agrawal, John McHale, and Alexander Oettl

  initially no redundancy in the knowledge that individual members bring to

  the team—that is, collective team knowledge is the sum of the knowledge

  of the individual team members—combining individual researchers into

  teams can greatly expand the knowledge base from which new combina-

  tions of existing knowledge can be made. This also opens up the possibility

  of a positive interaction between factors that facilitate the operation of

  larger teams and factors that raise the size of the fi shing- out/ complexity

  parameter, . New meta technologies such as deep learning can be more

  eff ective in a world where they are operating on a larger knowledge base

  due to the ability of researchers to more eff ectively pool their knowledge by

  forming larger teams.

  We thus extend in this section the basic model to allow for new knowledge

  to be discovered by research teams. For a team with m members and no

  overlap in the knowledge of its members, the total knowledge access for the

  team is simply mA. (We later relax the assumption of no knowledge overlap

  within a team.) Innovations occur as a result of the team combining exist-

  ing knowledge to produce new knowledge. Knowledge can be combined by

  the team a ideas at a time, where a = 0, 1 . . . mA. For a given team j with m members, the total number of possible combinations of units of existing knowledge (including singletons and the null set) given their combined

  knowledge access is

  mA

  (16)

  Z =

  mA

  = 2 mA .

  j

  a

  a=0

  Assuming again for convenience that A and Z can be treated as continu-

  ous, the per- period translation of potential combinations into valuable new

  knowledge by a team is again given by the (asymptotic) constant elasticity

  discovery function

  Z

  1

  (17)

  A =

  j

  =

  (2 mA )

  1 for 0 <

  1

  j

  = ln Z = ln(2 mA ) = ln (2) mA for = 0,

  j

  where use is again made of L’Hôpital’s rule for the limiting case of = 0.

  The number of researchers in the economy at a point in time is again L

  A

  (which we now assume is measured discretely). Research teams can poten-

  tially be formed from any possible combination of the L researchers. For

  A

  each of these potential teams, a entrepreneur can coordinate the team.

  However, for a potential team with m members to form, the entrepreneur

  must have relationships with all m members. The need for a relationship

  thus places a constraint on feasible teams. The probability of a relationship

  existing between the entrepreneur and any given researcher is , and thus

  the probability of relationships existing between all members of a team of

  size m is m. Using the formula for a binomial expansion, the expected total number of feasible teams is

  Artifi cial Intelligence and Recombinant Growth 163

  Fig. 5.3 Example of how the distribution of team size varies with ␩

  LA

  L

  (18)

  S =

  A

  m = (1+ ) LA.

  m

  m=0

  The average feasible team size is then given by

  LA

  LA

  mm

  m=0

  m

  (19)

  m =

  .

  LA

  LA

  m

  m=0

  m

  Factorizing the numerator and substituting in the denominator using equa-

  tion (18), we obtain a simple expression for the average feasible team size:

  LA

  LA

  mm

  m=0

  m

  1

  L

  (20)

  m =

  = (1+ ) LA

  A =

  L .

  A

  L

  (1 + ) LA

  1 +

  A

  LA

  m

  m=0

  m

  Figure 5.3 shows an example of the full distribution of teams sizes (with

  L = 50) for two diff erent values of . An increase in (i.e., an improvement A

  in the capability to form teams) will push the distribution to the right and

  increase the average team size.

  We can now write down the form that the knowledge production func-

  tion would take if all possible research teams could form (ignoring for the

  moment any stepping- on- toes eff ects):

  LA

  L

  (2 mA )

  1

  (21)

  A =

  A

  m

  for 0 <

  1.

  m

  m=0

  164 Ajay Agrawal, John McHale, and Alexander Oettl

  We next allow for the fact that only a fraction of the feasible teams will

  actually form. Recognising obvious time constraints on the ability of a given

  researcher to be part of multiple research teams, we impose the constraint

  that each researcher can only be part of one team. However, we assume the

  size of any team that successfully forms is drawn from the same distribution

  over sizes as the potential teams. Therefore, the expected average team size is

  also given by equation (18). With this restriction, we can solve for the total

  number of teams, N, from the equation L = N[/ (1 + )] L , which implies A

  A

  N = (1 + )/ .

  Given the assumption that the distribution of actual team sizes is drawn

  from the same distribution as the feasible team sizes, the aggregate knowl-

  edge production function (assuming > 0) is then given by

  LA

  L

  (2 mA )

  1

  (22)

  A = (1 + ) /

  A

  m

  (1 + ) LA

  m

  m=0

  LA

  L

  (2 mA )

  1

  =

  1

  A

  m

  ,

  (1 + ) L 1

  A

  m

  m=0

  where the fi rst term is the actual number of teams as a fraction of the poten-

  tially feasible number of teams. For = 0 the aggregate knowledge produc-

  tion function takes the form

  LA

  L

  (23)

  A =

  1

  A

  mm ln(2) A

  (1 + ) L 1

  A

  m

  m=0

  =

  1

  (1 + ) L 1

  A

  L

  ln(2) A

  (

  )

  A

  (1 + ) L 1

>   A

  = L ln(2) A .

  A

  To see intuitively how an increase in could aff ect aggregate knowledge

  discovery when > 0, note that from equation (20) an increase in will

  increase the average team size of the teams that form. From equation (16),

  we see that for a given knowledge access by an individual researcher, the

  number of potential combinations increases exponentially with the size of

  the team, m (see fi gure 5.4). This implies that combining two teams of size

  mʹ to create a team of size 2mʹ will more than double the new knowledge output of the team. Hence, there is a positive interaction between and .

  On the other hand, when = 0, combining the two teams will exactly double

  the new knowledge output given the linearity of the relationship between

  team size and knowledge output. In this case, the aggregate knowledge is

  invariant to the distribution of team sizes.

  To see this formally, note that from equation (23) we know that when = 0,

  the partial derivative of A with respect to must be zero since does not

  appear in the fi nal form of the knowledge production function. This results

  Artifi cial Intelligence and Recombinant Growth 165

  Fig. 5.4 Team knowledge production and team size

  from the balancing of two eff ects as increases. The fi rst (negative) eff ect is

  that the number of teams as a share of the potentially possible teams falls.

  The second (positive) eff ect is that the amount of new knowledge production

  if all possible teams do form rises. We can now ask what happens if we raise

  to a strictly positive value. The fi rst of these eff ects is unchanged. But that

  second eff ect will be stronger provided that the knowledge production of a

  team for any given team size rises with . A suffi

  cient condition for this to

  be true is that

  1/

  (24)

  A >

  1

  for all m > 0.

  ln(2) m

  We assume that the starting size of the knowledge stock is large enough so

  that this condition holds. Moreover, the partial derivative of A with respect

  to will be larger the larger is the value of . We show these eff ects for a

  particular example in fi gure 5.5.

  The possibilities of knowledge overlap at the level of the team and dupli-

  cation of knowledge outputs between teams creates additional complica-

  tions. To allow for stepping- on- toes eff ects, it is useful to fi rst rewrite equa-

  tion (20) as

  L

  1

  A

  L

  (2 mA )

  1

  (25) A = 1 +

  L

  A

  m

  .

  1 +

  A (1+ ) L 1

  A

  L

  m

  A

  m=0

  We introduce two stepping- on- toes eff ects. First, we allow for knowledge

  overlap within teams to introduce the potential for redundancy of knowl-

  edge. A convenient way to introduce this eff ect is to assume that the overlap

  166 Ajay Agrawal, John McHale, and Alexander Oettl

  Fig. 5.5 Relationships between new knowledge production, ␩, and ␪

  reduces the eff ective average team size in the economy from the viewpoint

  of generating new knowledge. More specifi cally, we assume the eff ective

  team size is given by

  (26)

  me = m =

  L

  ,

  1 +

  A

  where 0 ≤ ≤ 1. The extreme case of = 0 (full overlap) has each team act-

  ing as if it had eff ectively a single member; the opposite extreme of = 1

  (no overlap) has no knowledge redundancy at the level of the team. Second,

  we allow for the possibility that new ideas are duplicated across teams. The

  eff ective number of non- idea- duplicating teams is given by

  (27)

  N e = N 1

  = 1+

  1

  ,

  where 0 ≤ ≤ 1. The extreme case of = 0 (no duplication) implies that

  the eff ective number of teams is equal to the actual number of teams; the

  extreme case of = 1 (full duplication) implies that a single team produces

  the same number of new ideas as the full set of teams.

  We can now add the stepping- on- toes eff ects—knowledge redundancy

  within teams and discovery duplication between teams—to yield the general

  form of the knowledge production function for > 0:

  L

  1

  A

  L

  (2 mA )

  1

  (28) A = 1 +

  1

  L

  A

  m

  .

  1 +

  A

  (1 + ) L 1

  A

  L

  m

  A

  m=0

  If we take the limit of equation (24) as goes to zero, we reproduce the

  limiting case of the knowledge production function. Ignoring integer con-

  straints on L , this knowledge production function again has the form of

  A

  the Romer/ Jones function:

  Artifi cial Intelligence and Recombinant Growth 167

  1

  LA

  L

  (29) A = 1 +

  1

  L

  A

  m

  ln(2) mA

  1 +

  A

  (1 + ) L 1

  A

  L

  m

  A

  m=0

  (1 + ) L 1

  A

  L

  = 1+

  1

  L

  A

  ln(2) A

  1 +

  A

  (1 + ) L 1

  A

  LA

  = 1+

  1

  ln(2) L A .

  1 +

  A

  We note fi nally the presence of the relationship parameter in the knowl-

  edge production equation. This can be taken to refl ect in part the impor-

  tance of (social) relationships in the forming of research teams. Advances

  in computer- based technologies such as email and fi le sharing (as well as

  policies and institutions) could also aff ect this parameter (see, e.g., Agrawal

  and Goldfarb [2008] on the eff ects of the introduction of precursors to

  today’s internet on collaboration between researchers). Although not the

  main focus of this chapter, being able to incorporate the eff ects of changes

  in collaboration technologies increases the richness of the framework for

  considering the determinants of the effi

  ciency of knowledge production.

  5.5 Discussion

  5.5.1 Something New under the Sun? Deep

  Learning as a New Tool for Discovery

  Two key observations motivate the model developed above. First, using

  the analogy of fi nding a needle in a haystack, signifi cant obstacles to dis-

  covery in numerous domains of science and technology result from highly

  nonlinear relationships of causes and eff ect in high- dimensional data. Sec-

  ond, advances in algorithms such as deep learning (combined with increased

  availability of data and computing power) off er the potential to fi nd relevant

  knowledge and predict combinations that will yield valuable new discoveries.

  Even a cursory review of the
scientifi c and engineering literatures indi-

  cates that needle- in-the- haystack problems are pervasive in many frontier

  fi elds of innovation, especially in areas where matter is manipulated at the

  molecular or submolecular level. In the fi eld of genomics, for example, com-

  plex genotype- phenotype interactions make it diffi

  cult to identify therapies

  that yield valuable improvements in human health or agricultural produc-

  tivity. In the fi eld of drug discovery, complex interactions between drug

  compounds and biological systems present an obstacle to identifying prom-

  ising new drug therapies. And in the fi eld of material sciences, including

  nanotechnology, complex interactions between the underlying physical and

  chemical mechanisms increases the challenge of predicting the performance

  of potential new materials with potential applications ranging from new

  168 Ajay Agrawal, John McHale, and Alexander Oettl

  materials to prevent traumatic brain injury to lightweight materials for use

  in transportation to reduce dependence on carbon- based fuels (National

  Science and Technology Council 2011).

  The apparent speed with which deep learning is being applied in these and

  other fi elds suggests it represents a breakthrough general purpose meta tech-

  nology for predicting valuable new combinations in highly complex spaces.

  Although an in-depth discussion of the technical advances underlying deep

  learning is beyond the scope of this chapter, two aspects are worth highlight-

  ing. First, previous generations of machine learning were constrained by the

  need to extract features (or explanatory variables) by hand before statistical

  analysis. A major advance in machine learning involves the use of “repre-

  sentation learning” to automatically extract the relevant features.6 Second,

  the development and optimization of multilayer neural networks allows

  for substantial improvement in the ability to predict outcomes in high-

  dimensional spaces with complex nonlinear interactions (LeCun, Bengio,

  and Hinton 2015). A recent review of the use of deep learning in computa-

  tional biology, for instance, notes that the “rapid increase in biological data

  dimensions and acquisition rates is challenging conventional analysis strate-

  gies,” and that “[m]odern machine learning methods, such as deep learning,

  promise to leverage very large data sets for fi nding hidden structure within

  them, and for making accurate predictions” (Angermueller et al. 2016, 1).

  Another review of the use of deep learning in computational chemistry

  highlights how deep learning has a “ubiquity and broad applicability to a

  wide range of challenges in the fi eld, including quantitative activity relation-

  ship, virtual screening, protein structure prediction, quantum chemistry,

 

‹ Prev