Call Girls Indiranagar Just Call 👗 7737669865 👗 Top Class Call Girl Service B...
Deep Learning Deep Change NBER conference
1. Deep learning, deep change? Mapping the
development of the Artificial Intelligence
General Purpose Technology
Joel Klinger, Juan Mateos-Garcia and Kostas Stathoulopoulos
NBER Economics of AI Conference
Toronto, 26-27 September 2019
2. Motivation
● We are experiencing a boom in open AI
research
○ 77% of all AI papers published in the last
five years
○ Pre-prints sites and open source playing an
important role in this process
● Can we use this data to produce detailed and
relevant maps of AI R&D?
● And to answer interesting questions about the
economics of AI?
○ Is it an invention in the methods of
invention (or development)?
○ Has its geography been disrupted by the
arrival of new methods (deep learning)?
○ What local factors are conducive to the
development of AI R&D clusters?
2
3. ● We use data from arXiv, a pre-prints dataset widely used by the AI research community to
disseminate its findings. Currently contains 1.59m papers.
○ We focus on papers in the Computer Science and Stats (Machine learning) categories
● We use CrunchBase to measure industrial activity in sectors related to AI
Data
arXiv
134k cs papers
Microsoft
Academic Graph
arXiv enriched
240,000 institution-paper
matches
GRID
Institutions
Citations
Places
Titles
CrunchBase
Geocoding
Supervised machine learning to
identify sectors related to DL
Analysis https://github.com/nestauk/arxiv_ai
Topic modelling to identify ~15,000
DL papers
Co-occurrence analysis to identify
research related to AI
3
4. Evidence of DL as a GPT inside computer science research
Building on Cockburn, Stern & Henderson (2018)
● We find evidence that AI is a GPT in computer science
(“invention in the methods of development”?)
○ Rapid growth (slide 2)
○ Applied in many sectors
○ Impactful - or at least influential in terms of
citations.
4
● This is particularly visible in AI fields (cs.NE, cs.AI,
cs.LG) and application fields involving large,
unstructured datasets
○ CS.CV (Computer Vision)
○ CS.CL (Computer Language)
○ CS.MM (Multimedia)
○ CS.IR (Information retrieval)
5. Geographical evolution
Building on Goldfarb and Trefler (2018)
We have analysed changes in specialisation and
concentration of research in countries & regions.
● We find important changes in the geography of
DL, with the arrival of China and the (relative)
decline of some EU countries.
● DL research is more geographically concentrated
than other arXiv categories.
● After an initial decline in concentration, DL
research starts becoming more concentrated,
consistent with the idea of consolidation after a
‘shake out’.
5
6. Complementarities with local ecosystems
● AI R&D clusters might benefit from co-location with other research disciplines and
deployment sectors.
● We model regional specialisation in DL research post-2012 as:
● And compare the results with equivalent models for other arXiv categories
6
The results suggest that co-location with related disciplines and sectors, and with a
critical mass of research and industrial activity , as well as being in China is associated
with an increase in DL R&D specialisation in a region.
7. Conclusions
Limitation Next step
Experimental datasets Additional validation with other sources
Monolithic definitions Distinguish between theoretical and applied, and measure sectoral relevance of research
Informal analysis Formalise our model of research-industry complementarities
Correlational analysis Develop identification strategy using a technology or policy shock (arrival of TensorFlow?)
Mechanisms? Use additional information such as co-authorships & citations, access to funding, labour flows to
explain the patterns in the data
Economic impacts? Focus on the link between DL research and applied outputs (startups, open software)
Policy implications
● Our results suggest that the arrival of DL has disrupted the geography of AI research creating
opportunities for new regional entrants.
● Evidence of clustering could provide a rationale for targeted (regional) investments
● But the geography of DL seems to be consolidating, and policymakers will need to consider
complementarities with related research / industrial capabilities
New data and analysis can play an important role informing their decisions
7