Measuring AI Development

December 2021
Measuring AI Development
A Prototype Methodology to Inform Policy
CSET Data Brief
AUTHORS
Jack Clark
Kyle Augustus Miller
Rebecca Gelles
Table of Contents
Introduction: Upgrading AI Measurement for the 21st Century ....... 2

Methodology ..................................................................................................... 4
Case Studies: Re-Identification, Speaker Recognition, and Image
Synthesis ............................................................................................................ 8
Topic Overview ........................................................................................................ 8
Re-Identification ................................................................................................................ 8
Speaker Recognition ........................................................................................................ 9
Image Synthesis ............................................................................................................. 10
Results and Analysis ...........................................................................................12
1. Key Takeaways .......................................................................................................... 12
2. Bibliometric Analysis ............................................................................................... 14
3. Performance Metrics Assessment ...................................................................... 29
Leveraging the Methodology .................................................................... 35
Authors ............................................................................................................ 38
Acknowledgments ....................................................................................... 38
Appendices ..................................................................................................... 39
Appendix A: Methodology (Detailed) ..................................................................... 39
Appendix B: Seed Papers ........................................................................................... 42
Appendix C: Notes on Terminology ........................................................................ 48
Appendix D: Progression of Key Machine Learning Technologies............... 50
Appendix E: Organizational Affiliation Data (Detailed) ................................... 51
Appendix F: Metrics Assessment (Detailed) ........................................................ 53
Appendix G: Performance Metrics Data and Sources ....................................... 57
Endnotes.......................................................................................................... 59
Center for Security and Emerging Technology | 1

Introduction: Upgrading AI Measurement for the 21st
Century
This report demonstrates how contemporary approaches to

bibliometric and technical analysis can give policymakers an
adaptable tool to better understand ongoing AI developments.1 By
combining bibliometric analysis (via CSET’s recently developed
Map of Science derived from CSET’s research clusters and merged
corpus of scholarly literature) with qualitative knowledge of
specific AI subfields, we show how detailed pictures of AI progress
can be created to help policymakers understand the current state
of research for specific AI technologies. This report applies our
methodology to three topics: re-identification, speaker recognition,
and image synthesis. Additionally, we offer ideas for how
policymakers can integrate this approach into existing and future
measurement programs.2
By combining a versatile and frequently updated bibliometrics tool

with a more hands-on analysis of technical developments within a
particular AI subfield, we can build out a detailed picture of the
state of published research into specific topics. In particular,
tracking how AI subfields show improvements in performance
benchmarks can send a powerful signal about the state of
technical progress. Benchmarks have played a critical role in
spurring AI development for topics as wide-ranging as image
recognition, self-driving cars, and robotics,3 and they can be used
to supplement bibliometric-based approaches, allowing us to see
whether rapid research growth is resulting in genuine
improvements in performance. This methodology differs from most
contemporary approaches to technology assessment because it is
structured to allow deeper system-led approaches, places a
significant premium on the analysis of fast-moving AI publications
(including bibliometric analysis of preprints), and is generically
applicable to a variety of AI capabilities, rather than custom-
designed for a specific area of study.

Why Does Measurement Matter?
Measurement is uniquely intertwined with policymaking, especially

in the case of AI. For the past few decades, attempts by
policymakers and academic researchers to assess and measure AI
have been linked. The National Institute for Standards and
Technology's (NIST) MNIST dataset of handwritten digits became
a valuable source to help gauge the overall pace of AI progress as
people used it as a simple benchmark to assess capabilities. The
XVIEW dataset, designed by the United States DOD’s Defense
Innovation Unit Experimental (DIUx), played a role in stimulating
the development of computer vision capabilities applied to
satellite imagery and served as a resource that generated data
about AI progress. Similarly, the Defense Advanced Research
Projects Agency’s (DARPA) self-driving car and robotics
competitions have themselves been exercises in measurement and
assessment that catalyzed work in a technical area and gave
researchers a sense of emerging capabilities in a policy-relevant
field. Perhaps the most prominent example of how measurement
and assessment impacted policymakers was the development of
ImageNet, a Stanford University project to help researchers test
the performance of computer vision systems against a large,
deliberately challenging (at the time) dataset.
This report aims to go one step further. The U.S. government today
periodically conducts tests or assessments of AI capabilities, or
develops new tests and datasets at the behest of experts or
agencies to understand AI technologies. We instead propose a
system for continuously monitoring and assessing AI-enabled
capabilities for publication patterns that might highlight
consequential trends to the government, and for continuously
analyzing technical benchmarks that can help the government
detect significant advancements. While many different
organizations regularly try to assess and measure AI capabilities
for a specific policy purpose, we propose a system for measuring
capabilities and patterns within AI as a whole.4
After outlining our methodology, this report discusses how

qualitative expert knowledge combined with bibliometric tools,

like CSET’s Map of Science, can generate insights regarding
developments in the specific research areas of re-identification,
speaker recognition, and image synthesis. We also discuss some
limitations of this approach and end with recommendations for
how policymakers can support the development of similar types of
analytical tools.
Methodology
Bibliometric Clustering: The technical basis of our methodology is

a bibliometric analysis tool, in this case CSET’s Map of Science—a
merged corpus of research representing most of the world’s
scientific literature that groups papers into clusters based on the
citation links between them.5 For each research cluster, the Map of
Science provides detailed metadata, such as keywords and core
subjects, the country or countries in which the authors of the
cluster papers are based, and the types of organizations commonly
listed as funders for research. Additionally, it provides several
CSET-calculated indicators, such as the percentage of papers
within a cluster related to AI, computer vision, natural language
processing, or robotics.6
Bibliometric Analysis: To begin our analysis, we first selected

topics related to machine learning and reviewed some of the
relevant literature to identify 10–20 seed papers that are central to
the given fields.7 The seeds included highly cited works that were
published across multiple years and encompassed key elements of
the topic at hand (the seed papers used in the case studies can be
found in Appendix B). We then searched for these papers in the
Map of Science to identify the clusters that encompassed the
highest number of them. Identifying quality seed papers requires a
degree of domain expertise, but the approach to identifying
clusters can be altered to be more systems-led and operable for
non-experts. For example, one could use keyword searches in the
Map of Science to identify relevant clusters and highly cited
papers, and then manually inspect the papers.
For the purposes of this report, we explored whether we could

narrow each machine learning topic to a single cluster from which
we could extract metadata. These single clusters then became the

core of our subsequent bibliometric analysis. However,
adjustments can be made to this method depending on the
distribution of seed papers across clusters (i.e., research
concentration), ambiguity of the topic, or quality of the data. For
example, the bibliometric analysis could include multiple clusters if
papers related to a topic are highly distributed. We must stress
that individual clusters in the Map of Science represent related
research (via citations between papers) and are not explicitly
structured to encompass a topic. Therefore, in various cases, a
multi-cluster analysis will be optimal or necessary.
Performance Metrics Assessment: In machine learning research,

emerging areas of study often become focused around a
performance metric or group of metrics, which tends to spur
further investment in the field. The clusters used in the bibliometric
analysis provide relevant research to look for such metrics. These
metrics can supplement and corroborate the insights from the
metadata analysis. By manually reviewing the literature within a
research cluster, we can identify when a performance metric for
the field first emerged and how top-level performance on that
metric has changed over time. This can provide policymakers with
a better sense of how rapidly capabilities in a given area are
improving and where such developments are occurring. A
performance assessment works well when there are existing
common benchmarks to evaluate but can be more challenging if
there are no relevant metrics available or they are outdated.
However, by identifying a lack of common metrics, our
methodology can illuminate knowledge gaps in areas that are
relevant to policymakers.

Figure 1. Bibliometric Clustering, Bibliometrics Analysis, and Performance Metrics Assessment
Note: See Appendix A for a more detailed description of bibliometric clustering via the Map of Science, bibliometric analysis, and performance
metrics assessment. Source: CSET.

By applying this methodology, we can measure the publication
growth of specific machine learning–related topics, where it is
occurring, what organizations and individuals are involved, and
when significant technical advancements occur. This can be
adjusted or augmented with additional processes, but we believe
these three elements establish an effective baseline
methodological structure that is applicable to a range of topics and
users.
It should be noted that this approach is not perfect. Clustering

tools like the Map of Science do not necessarily encompass the
entirety of a given subject area, and they may often contain
overlapping, incomplete, or unreliable metadata that must be
manually reviewed. Bibliometrics also do not include private or
closed-source research, and instead rely exclusively on published
scientific literature. Finally, both the selection of seed papers and
the identification of performance metrics are subjective and often
involve heuristic decisions that cannot be automated easily (at
present).8

Case Studies: Re-Identification, Speaker Recognition, and
Image Synthesis
In this section, we apply the methodology to gain insights into the

state of research in two emerging machine learning–enabled
capabilities: re-identification and speaker recognition. We selected
these topics because their emergence may have significant policy
implications. We compare the results of this analysis to a more
mature subfield of machine learning technology: image synthesis.
To demonstrate this methodology clearly and without excessive
bibliometric analytics, we narrowed the scope of analysis to a
single research cluster for each capability (topic); however, this
approach can be altered to include multiple clusters.
The results described below provide meaningful visibility into the

rate, scale, distribution, and improvement of re-identification and
speaker recognition capabilities across countries, organizations,
and individuals. However, our particular application of the
methodology—using one cluster per topic—was unsuccessful with
image synthesis, primarily because research papers related to this
topic are distributed across multiple clusters.
Topic Overview
Re-Identification
Re-identification involves integrating computer vision techniques

with networked surveillance systems to identify and distinguish
objects, track their locations, and correlate their movements
(around obstacles and amidst appearance changes). The input to
these models is typically CCTV footage, and the underlying
methods can be adapted to track people, vehicles, or objects.
Recent progress in re-identification was enabled by major
breakthroughs in computer vision between 2010 and 2015, which
made use of deep neural networks to rapidly improve on image
recognition datasets like ImageNet.9 (See Appendix C for
additional details.)

Figure 2. Re-Identification Process
Source: Mang Ye et al. 10
The development of re-identification systems has important

implications for policymakers: these systems could potentially
offer extremely useful tools to law enforcement, making it cheaper
and easier to surveil large populations, apply other machine
learning models to the resulting data, and subsequently uncover
patterns of activity and association amongst individuals of
interest.11 These capabilities exhibit the potential to augment or
replace humans in the manual analysis of CCTV footage and
dramatically alter the process of imagery intelligence in public
surveillance. Entities in both authoritarian and non-authoritarian
regimes have incentives to develop re-identification for various
purposes—such systems can equally help retailers understand
patterns of traffic in shopping malls as they could help police track
and identify criminal or dissident behavior.
Speaker Recognition
Speaker recognition is an area of research that focuses on using

machine learning to recognize the voices of specific speakers in
audio data.12 Speaker recognition systems can be adapted to
perform a number of more specific tasks, such as verifying the
identity of a known speaker, identifying audio from an unknown
speaker in a group of known speakers, or partitioning audio into
segments depending on when each speaker is talking. 13 As with
re-identification, current progress in speaker recognition is heavily
dependent on previous advances in audio speech recognition
during the 2010–2015 era, as measured by performances on
datasets like Switchboard. 14
Speaker recognition models can be integrated into a range of

different systems that use audio data, including trivial applications

that identify celebrities’ voices in YouTube videos and security
programs that verify voice identities to grant physical access to
sensitive locations. Improvements in speaker recognition merit
attention from policymakers: these systems could enable improved
intelligence processes that automate the analysis of raw phone
audio data at unprecedented scales. (See Appendix C for
additional details.)
Image Synthesis
Image synthesis is a subfield of machine learning research that

focuses on generating realistic-looking images.15 It has been
closely linked to the development of generative adversarial
networks (GANs) since 2014.16 This topic includes the notorious
world of “deepfakes,” or hyper-realistic images of human beings
that have already been used in Hollywood special effects, political
campaigns, and online propaganda. By comparing the fields of re-
identification and speaker recognition to image synthesis, we can
assess how our method of analysis performs on both specific
subfields, as well as more generic ones like image synthesis.
Image synthesis presents a host of issues with policy implications.

In 2019, the president of Gabon was accused of using synthetic
human imagery (or more prosaically, a body double) to stage a
video meant to reassure civilians in the country.17 In India, a
politician made himself appear to give a speech in a language he
did not speak, while propagandists in West Papua used synthetic
facial imagery for fake accounts that aimed to undercut the
independence movement.18 By 2021, image synthesis was widely
used for a variety of legitimate and illegitimate purposes, ranging
from entertainment to making disinformation more convincing.

Figure 3. Image Synthesis Progression, 2014-2019
Sources: Ian J. Goodfellow et al.; Alec Radford, Luke Metz, and Soumith
Chintala; Ming-Yu Liu and Oncel Tuzel; Tero Karras et al.; Tero Karras, Samuli
Laine, and Timo Aila; and Yunjey Choi et al. 19

Results and Analysis
In what follows, we focus first on re-identification and speaker

recognition. We then show how our attempt to replicate the
methodology for image synthesis was unsuccessful when using a
single-cluster analysis because many relevant papers were
distributed across multiple clusters. To baseline our bibliometric
results, we compare the growth of re-identification and speaker
recognition papers against changes in the composition of all AI
papers, as identified by CSET’s Map of Science. We also baseline
relevant performance metrics by comparing them to two broader,
but related, topics of computer vision and natural language
processing.
1. Key Takeaways
Overall:
(1) We identified two respective research clusters for re-

identification and speaker recognition and are confident that
they encompass most publications on the topics. From the
papers in these clusters, we extracted metadata and
performance metrics.20
(2) Publications related to re-identification and speaker

recognition increased substantially between 2010–2020,
and their performance metrics improved significantly. A
continuation of this trend will likely have political, security,
and commercial consequences that warrant policymaker
attention.
(3) Growth of re-identification and speaker recognition

publications occurred alongside the discovery and
proliferation of many key machine learning subcomponents,
including deep learning–based techniques, large datasets,
and increased computational capacity.
(4) Most organizations affiliated with the authors of papers

for both topics were educational, although speaker
recognition had a higher affiliation with private industry.

Re-Identification Speaker Recognition
(5) Re-identification publications (5) Speaker recognition

in cluster #1419 had an average publications in cluster #6855 had
annual growth rate of 40 percent an average annual growth rate of
between 2010–2020. 33 percent between 2010–2020.
Publications increased Publications increased
substantially between 2015– substantially around 2017, which
2018, then plateaued in 2020. continued through to 2020.
(6) There were 64 author-
(6) There were 67 author-
affiliated countries in the speaker
affiliated countries in the re-
recognition cluster, of which the
identification cluster, of which
top two were the United States
China was the most prominent by
and China, which had relatively
a large margin, followed by the
equal affiliation, followed by
United States, the UK, and
India, the Czech Republic, and
Australia.
Japan.
(7) There were over 600 author- (7) There were over 400 author-
affiliated organizations in the re- affiliated organizations in the
identification cluster, most of speaker recognition cluster, most
which are in the education sector. of which are in the education
The top three were the Chinese sector. The top four were Johns
Academy of Sciences, Queen Hopkins University, MIT, Brno
Mary University of London, and University of Technology, and SRI
Sun Yat-sen University. International.
(8) The performance of re- (8) The performance of speaker

identification models improved recognition models improved
significantly in recent years, significantly in recent years.
nearly doubling in performance Improvements were driven by
against our selected benchmark researchers at the University of
since 2018. Multiple Oxford (UK) who designed the
organizations—predominantly metrics (i.e., dataset and
from China—made progress on challenge), Ghent University
one of the most challenging (Belgium), and most recently
testing datasets. SpeakIn Technologies (China).

2. Bibliometric Analysis
From our bibliometric analysis we identified and extracted

metadata from two respective research clusters for re-
identification and speaker recognition. These clusters encompass
most, but not all, papers on the topics, which makes an analysis of
the clusters’ metadata akin to an analysis of the topics. For re-
identification, 12 of our 14 seed papers fell into a single cluster
(#1419), with the other two falling into a cluster related to object
detection and tracking that was relatively light on machine
learning–related papers. Even more promisingly, all 15 of our seed
papers for speaker recognition fell within a single cluster (#6855).
After a spot-check confirming the clusters’ relevance, we felt
comfortable making the focus of our analysis a single cluster for
each topic—most subsequent metadata in this report is that of the
papers in these clusters.
Here we begin with an analysis of publication and citation growth

within each cluster (per topic), then proceed to investigate country
affiliation, organization affiliation, and authors. To provide context,
we compare the growth of our two topic clusters with the growth
of AI broadly (i.e., all publications related to AI). However, unlike
with re-identification and speaker recognition, the AI data does not
stem from clusters. Instead, we used a trained classifier model to
identify and extract all “AI-related” publications from the merged
corpus,21 ultimately giving us approximate figures on the entire
field. Summary statistics for the two topic clusters, as well as AI
publishing overall, are displayed in Table 1. As we describe in
more detail below, image synthesis is omitted from this section as
it required a multi-cluster analysis, while our focus in this paper is
on single-cluster bibliometric analysis.

A. Publications and Citations
Table 1. Papers, Citations, and Growth Rates
Speaker
Re-Identification Recognition
(Cluster #1419) (Cluster #6855) AI
Total Papers (2010-2020) 3,282 1,938 2,015,541
Total Paper Citations (2010-2020)22 96,462 29,096 24,896,447
Average Annual Paper Growth Rate

(2010-2020) 40% 33% 27%
Average Annual Paper Citation

Growth Rate (2010-2020) 21% 15% 23%
Source: CSET Map of Science derived from CSET’s research clusters and merged
corpus of scholarly literature, as of March 2021.
Research into both re-identification and speaker recognition has

grown more quickly than overall AI research. AI grew at an annual
rate of 27 percent from 2010–2020, while the number of papers
related to re-identification and speaker recognition grew at an
average rate of 40 percent and 33 percent per year, respectively.
Although the citation numbers in Table 1 may make it appear that
citations for re-identification and speaker recognition papers have
not grown as quickly as citations for AI papers overall, this is likely
an artifact, in part, of the relatively later increase in attention to
these sub-topics—especially re-identification—around 2015, as
shown in Figures 4 and 5. Such recent papers can take time to
accumulate citations, which helps explain why the citation
numbers do not appear to have grown as rapidly in the two
subfields.
Figure 4 shows a divergence in growth between the topics around

2015, with re-identification publications quickly outpacing speaker
recognition. We suspect that this divergence is related to the fact
that computer vision has been a larger research area with a

broader problem space than voice biometry. However, we are
ultimately uncertain of its cause and further investigation is
warranted, as many trends of the early 2010s—such as the
development of deep neural networks, increased computing
capabilities, and a growth in available data—enabled both types of
research.
Figure 4. Machine Learning Topic Papers Published per Year,

2010–2020
1000
NUMBER OF PAPERS
799
800
600 545
400 370
161 258 256
95 142 212
200 27 62
26
0
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020
YEAR
Speaker Recognition (Cluster #6855)

Re-Identification (Cluster #1419)
Note: For a more detailed analysis of re-identification and speaker recognition

publication progression, see Appendix D. Source: CSET Map of Science, as of
March 2021.
Figure 5. AI Papers Published per Year, 2010–2020
300
271
NUMBER OF PAPERS
250
211
(THOUSANDS)
200 166 168

152
141
150
100
50
0
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020
YEAR
Source: CSET Map of Science, as of March 2021.

Image Synthesis—Topic Necessitates a Multi-Cluster
Investigation: We were unable to identify a single cluster for
image synthesis research because relevant publications were
distributed across multiple clusters. This largely stems from the
fact that clusters in the Map of Science are representations of
related research and are not explicitly intended to represent one
topic in one cluster. Subsequently, topics that have more
distributed research (e.g., image synthesis) require analyzing
multiple clusters, while topics with more concentrated research
(e.g., re-identification) can be analyzed through a single cluster.
Following an extensive spot-check of image synthesis publications
across clusters, we decided to omit this topic from our single-
cluster focus, as we were unable to get reliable metadata on this
topic from one cluster. Future research could explore this topic
through a multi-cluster analysis.
Here we briefly address three related limitations that stem from

restricting this methodology to a single cluster per topic and from
using a cluster of papers (based on citations) to investigate a topic:
- We identified one cluster (#1220) that contained many

publications related to image synthesis; however, it did not
include five of our twelve image synthesis seed papers,
which suggests that it did not sufficiently encompass
publications directly related to the topic.
- Many publications related to image synthesis are dispersed

across research clusters. In our initial attempt to identify a
single cluster for this topic, we had to omit clusters that,
while slightly relevant on a technical level, contained many
papers and research topics not directly related to image
synthesis as a capability. 23 These topics include facial
recognition, point clouds, neural style transfer, and
deepfake detection. This differs from the previous
investigations, during which we identified single clusters
that encompassed most papers directly related to re-
identification and speaker recognition.
- Cluster #1220 included many publications explicitly about

GANs and various extraneous topics (e.g., generative

grammar). To avoid having to manually omit data, and
because GANs are deeply intertwined with the
advancement of image synthesis, we considered partially
incorporating GANs into the investigation and attempted to
analyze the entire cluster. However, using the whole cluster
and expanding the scope of the topic made the investigation
too imprecise, as GANs are used for many purposes beyond
image synthesis, and we could not readily discern the topics
within the cluster metadata. This issue can arise when a
topic is not sufficiently distinct, or it encompasses
technologies with dual-use applications.

B. Countries
CSET’s Map of Science provides country affiliation metadata,

which is derived from organizational affiliations provided by the
authors of papers. The denominator for this country analysis is not
the quantity of papers in the clusters—it is the total country
affiliations of the papers in the clusters. 24 Extracting this metadata
from our re-identification and speaker recognition clusters allows
us to see which countries were most affiliated with paper authors
and which accrued the most paper citations (i.e., how many times a
country-affiliated paper is cited). This is displayed in Tables 2 and
3, which show the top five countries in terms of author country
affiliations and paper citations. 25
Table 2. Top 5 Author Country Affiliations, per Topic
AI Re-Identification Speaker Recognition
1. China 27% China 58% United States 24%
2. United States 16% United States 13% China 22%
3. India 5% UK 6% India 6%
4. UK 5% Australia 5% Czech Republic 4%
5. Japan 4% Italy 4% Japan 4%

Table 3. Top 5 Countries' Share of Paper Citations, per Topic
AI Re-Identification Speaker Recognition
1. United States 35% China 60% United States 48%
2. China 26% United States 24% Canada 15%
3. UK 9% UK 14% France 14%
4. Germany 7% Australia 13% China 11%
5. Canada 6% Italy 8% UK Source: 7%
CSET Map of Science, as of March 2021.
Compared to AI research generally, China was affiliated with a

slightly smaller percentage of papers in speaker recognition but a
substantially larger percentage of re-identification papers. This
finding suggests that Chinese entities may have a particular
interest in re-identification. Conversely, the United States was
affiliated with a slightly smaller percentage of re-identification
papers and a slightly larger percentage of speaker recognition
papers than would be proportionate to its share of all AI research.
It is, however, important to emphasize that much research in each
of these fields is highly international; the top-cited re-identification
paper between 2010–2020, for instance, was written by
researchers at the University of San Antonio, Microsoft, and
Tsinghua University.26
United States-affiliated papers for both topics tended to garner

disproportionately more citations relative to the number of
publications the country is affiliated with. China-affiliated speaker
recognition papers were cited disproportionately little relative to
the country’s affiliated publication output but roughly on par with
its affiliated publication output in re-identification and AI overall.
Notably, the composition of the countries affiliated with re-

identification papers changed between the years 2010–2020. This
is displayed in Figures 6 and 7, which show the share of paper
affiliations and paper citations for the top five countries per year
(from the re-identification cluster).27 While no single country

initially appeared dominant, China has been affiliated with most
papers in this cluster since 2015.
Figure 6. Country Share of Re-Identification Paper Affiliations,

2010–2020
80%
PERCENTAGE OF PAPER AFFILIATIONS
70%
60%
50%
40%
30%
20%
10%
0%
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020
YEAR
China United States United Kingdom
Australia Italy
Figure 7. Country Share of Re-Identification Paper Citations, 2010–

2020
80%
PERCENTAGE OF PAPER CITATIONS
70%
60%
50%
40%
30%
20%
10%
0%
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020
YEAR
China United States United Kingdom
Australia Italy

China-affiliated papers garnered 60 percent of the citations in our
re-identification cluster, 28 while United States-affiliated papers
garnered 24 percent of the citations. UK and Australia-affiliated
papers had relatively low citations of 14 and 13 percent,
respectively. Although China was still the country most affiliated
with papers and paper citations, the citation distribution suggests
that many entities from the other top countries are relatively more
prominent than the paper distribution alone suggests.
Researchers affiliated with Italy, the UK, and the United States
were involved in re-identification publications in our cluster in
2010 (the beginning of our analysis). This was prior to the
development of many key components used in emerging re-
identification systems. By 2013, China leaped ahead of all other
countries,29 maintaining a lead in affiliated publications to 2020.30
Overall, these trends suggest that China is far more active in
published re-identification research than any other nation.
In the speaker recognition cluster, by contrast, we see that the

United States was initially the country most affiliated with
publications, but by 2015 the proportion of papers from Chinese
entities increased and remained roughly on par with publications
by American researchers—though the United States maintained an
edge in its share of paper citations. This is displayed in Figures 8
and 9, which show the share of paper affiliations and paper
citations for the top five countries per year, for the speaker
recognition cluster.31

Figure 8. Country Share of Speaker Recognition Paper Affiliations,
2010–2020
80%
PERCENTAGE OF PAPER AFFILIATIONS 70%
60%
50%
40%
30%
20%
10%
0%
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020
YEAR
United States China India Czech Republic Japan
Figure 9. Country Share of Speaker Recognition Paper Citations,

2010–2020
80%
PERCENTAGE OF PAPER CITATIONS
70%
60%
50%
40%
30%
20%
10%
0%
2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020
YEAR
United States Canada France China United Kingdom

Among speaker recognition paper citations, the United States
garnered 48 percent of the citations in our speaker recognition
cluster. This was followed by Canada and France-affiliated papers
at 15 and 14 percent citations, respectively. Although China was
ranked second for paper affiliations, its affiliated papers only
garnered 10 percent of the citations in the cluster. Similarly,
although India was ranked third for paper affiliation, its affiliated
papers only garnered 3 percent of the citations.
With speaker recognition, researchers affiliated with the United

States, China, and the Czech Republic were involved in
publications between 2010–2011, as illustrated in Figures 8 and
9. As with re-identification, the growth rate of speaker recognition
papers affiliated with the United States and China increased
substantially around 2016. The United States was affiliated with
papers that accrued more citations during most years, even though
China was affiliated with more publications between 2015–2020.
Notably, recent speaker recognition growth appears more
internationally distributed than re-identification. For example,
China and the United States were collectively affiliated with
approximately 75 percent of the re-identification cluster’s
publications in 2019 but only affiliated with 58 percent of the
speaker recognition cluster’s publications the same year.
Explaining the differences in the top affiliated countries between
the two topic clusters is an area for additional research.
C. Organizations
Organizational affiliation metadata is based upon information

provided by authors in their publications and provides a more
granular level of visibility into machine learning publications
trends. Of the over 600 author-affiliated organizations in the re-
identification cluster, 32 76 percent were universities, 9 percent are
government organizations, and 6 percent are private companies.33
Of the over 400 author-affiliated organizations in the speaker
recognition cluster, 65 percent were universities, 11 percent are
private companies, 5 percent are government organizations, and 4
percent are in the non-profit sector. Table 4 displays the ten
organizations with the largest publication output for both re-
identification and speaker recognition. 34 Although the data further

demonstrates the dominance of Chinese organizations in re-
identification research, the results were (marginally) less Chinese-
centric when looking at paper citation counts. Microsoft-affiliated
authors of re-identification papers, for instance, accounted for 5
percent of all citations in the cluster. And yet, even here there is
significant Chinese involvement: Microsoft conducts much of its re-
identification research through the Beijing-based Microsoft
Research Asia, its largest research lab outside of the United
States.35
Table 4. Top Organizations Affiliated with Papers (Per Topic)
Number Share of
Top 10 Organizations Affiliated with Papers of Papers Topic Papers
Australia
University of Technology, Sydney 71 2%
China
Beihang University 70 2%
Beijing University of Posts and

Telecommunications 71 2%
Chinese Academy of Sciences 253 8%
Re-Identification
Peking University 85 3%
Shanghai Jiao Tong University 104 3%
Sun Yat-sen University 113 3%
Tsinghua University 70 2%
Wuhan University 82 2%
UK
Queen Mary University of London 85 3%

China
Chinese Academy of Sciences 71 4%
Hong Kong Polytechnic University 46 2%
Tsinghua University 69 4%
Czech Republic
Speaker Brno University of Technology 62 3%

Recognition
India
Indian Institutes of Technology System (IIT

System) 40 2%
United States
IBM 39 2%
Johns Hopkins University 71 4%
Massachusetts Institute of Technology 53 3%
SRI International 51 3%
University of Texas at Dallas 55 3%
Note: For a more detailed analysis of re-identification and speaker recognition

organization affiliations, see Appendix E. Source: CSET Map of Science, as of
March 2021.
D. Authors
Author metadata provides the most granular level of bibliometric

visibility into machine learning topics, enabling the identification of
prominent individuals involved in open-source publications. In
science, certain authors tend to serve as important “datapoints” in
their own right—for instance, certain authors may run a lab
exclusively dedicated to research in one area, so surfacing them
can help governments identify figures with significant influence in
a given field. Conversely, other authors may be multidisciplinary
and multi-institutional, serving as the connective tissue between
disparate sets of authors and research specialties. As with
organizations, we only intend to illustrate the potential for insight,

and thus the analysis here is limited. However, the data creates
opportunities for deeper analysis, such as identifying prominent
authors of interest for research collaboration and policy
formulation, as well as government investigation into foreign
authors researching machine learning capabilities that could have
security implications.
The data on the authors’ current country locations was manually

collected, as opposed to extracted from the clusters, because at
the time of our analysis such information was not readily available
in the prototype Map of Science.36 Additionally, many authors have
diverse backgrounds that include multiple organizations and work
locations over time, much of which is not portrayed through the
metadata. Notwithstanding these shortcomings, the data can be a
starting point for a deeper analysis of individual researchers, the
organizations they work for, and the professional networks in
which they are embedded.
Of the top twenty authors that published re-identification papers

in cluster #1419, almost half were located in China, with most
others dispersed across the UK, Australia, Italy, and Japan.
Notably, only one of the top authors was located in the United
States despite its relatively high affiliation to papers. Sixteen
authors were affiliated with organizations in the education sector,
primarily public universities, while four work in private industry at
Huawei (China), the Inception Institute of Artificial Intelligence
(UAE), the Samsung AI Center (UK), and Google (United States).
The prominence of authors in the education sector is influenced by
the fact that they publish their work, but we cannot determine to
what extent this skews the data toward education and away from
other sectors.
Of the top twenty authors that published speaker recognition

papers in cluster #6855, five were located in the United States and
China, with the remaining dispersed across Singapore, the Czech
Republic, Sri Lanka, Italy, India, Canada, Australia, and Argentina.
Sixteen top authors were affiliated with organizations in the
education sector, primarily public research universities, while two
were in private industry at Amazon and Apple (United States), two
in the nonprofits SRI International (United States) and the

Computer Research Institute of Montréal (Canada), and only one in
government (Singapore’s Agency for Science Technology and
Research). As previously noted, many authors have diverse
backgrounds that include multiple organizations and work
locations, much of which is not portrayed through the data.
Table 5. Locations of Top 20 Paper Authors
Country Number of Top 20 Authors
China 9
UK 3
Australia 2
Re-Identification Italy 2
Japan 2
United Arab Emirates 1
United States 1
United States 5
China 5
Singapore 2
Czech Republic 2
Canada 1
Speaker
Recognition Australia 1
Sri Lanka 1
India 1
Argentina 1
Italy 1

3. Performance Metrics Assessment
Although bibliometric tools like CSET’s Map of Science provide a

useful means of exploring the growth, geographic distribution, and
affiliations of research in emerging areas, bibliometrics alone do
not say anything about the quality of a growing body of research.
To assess whether the growth in re-identification and speaker
recognition research has generated corresponding improvements
in performance, we need to find useful performance metrics from
within those fields to track. Fortunately, after reviewing papers in
the clusters we identified two such metrics: MSMT17 for re-
identification and VoxCeleb for speaker recognition, which are
described in Table 6. However, if no such metrics existed, we could
scan publications to identify datasets and performance evaluation
methodologies shared across multiple areas, then manually
compile the data to give performance metrics. Most all papers
used to acquire performance metrics are from the same two
clusters used in the bibliometric analysis. 37 Lastly, we must note
that experts in the field often disagree over what datasets to
measure progress on, so a more rigorous application of this
methodology could involve tracking progress on different models
across varying datasets to get a more nuanced picture of
development.

Table 6. Performance Metrics and Datasets
Performance
Dataset Dataset Description Metric Metric Description
MSMT17 MSMT17 is a multi- mean Average mAP calculates the

scene, multi-time, Precision accuracy of a model’s
person re-identification (mAP) imagery detection. A
dataset. It consists of higher mAP score
Re-Identification 180 hours of videos, signifies greater
captured by 12 outdoor accuracy.
cameras and 3 indoor
cameras during 12 time
slots.
VoxCeleb VoxCeleb1 is an audio- Equal Error EER calculates the

(VoxCele visual dataset for Rate (EER) likelihood of a
b1; speaker recognition biometric system
VoxCeleb containing hundreds of incorrectly classifying
2) thousands of “real the subject of interest
Speaker
world” utterances from (e.g., speaker audio)
Recognition
more than 1,000 via false positives or
celebrities. VoxCeleb2 false negatives. A
contains over a million lower EER score
utterances from more signifies less error and
than 6,000 speakers. greater accuracy.
Source: See Appendices B and G.
The MSMT17 dataset was introduced in 2017 by researchers at

Peking University and the University of Texas at San Antonio with
the intention of providing a challenging test relative to existing
benchmarks.38 Because of its recency, as well as the fact that the
best-performing attempts have tended to make use of some of the
latest innovations in computer vision, we are confident that
performance on the MSMT17 dataset is reflective of meaningful
progress.39 Several publications advanced performance on the
MSMT17 dataset, of which we selected five.40 The improvements
in performance, as measured by mean Average Precision, have
doubled since 2018. This is displayed in Figure 10. Researchers
driving this progress correlate with those we would expect to see
from our bibliometric analysis—three of the performance data
points were created by China-affiliated researchers and
organizations, while two were affiliated with the United States.
This corroborates the insights derived from our bibliometric

analysis of re-identification, which indicated that both are the most
prominent countries affiliated with re-identification publications
and citations.
Figure 10. Re-Identification Performance, 2018–2021

100
80 69.4
62.7 67.3
MAP SCORE
60.8
60
40 34
20
0
Sep-18 Jan-19 May-19 Sep-19 Jan-20 May-20 Sep-20 Jan-21
DATE
Source: See Appendix G.
The VoxCeleb dataset, which we used to assess improvements in

speaker recognition, was introduced in 2017 by researchers at the
University of Oxford.41 Unlike MSMT17, VoxCeleb is also
accompanied by an annual competition which sees multiple
groups compete to test the performance of their systems against a
standardized test. We selected six publications that advanced
performance on the VoxCeleb dataset since 2017. These
improvements, as measured by reductions in competitors’ equal
error rate, have been significant and sustained since 2017. This is
displayed in Figure 11. However, as of 2020, the Equal Error Rates
of models tested against VoxCeleb had limited room for
improvement, so testers must design more difficult benchmarks to
gauge performance of newer models. Improvements in state-of-
the-art performance on VoxCeleb have primarily been driven by a
research group at the University of Oxford (UK) that created the
dataset and challenge, Ghent University (Belgium), and most
recently SpeakIn Technologies (China). Compared to re-
identification, these speaker recognition metrics correlated less
with the distribution of the top country affiliations that were
analyzed in the bibliometric analysis—the UK and Belgium were
not amongst the most prominent countries affiliated with speaker
recognition publications.

Figure 11. Speaker Recognition Performance, 2017–2021
9%
8% 7.80%
7%
6%
5% 4.19%
EER
4% 3.22%
2.87%
3%
2%
0.56% 0.52%
1%
0%
DATE
Source: See Appendix G.
In both research areas, we can see improvements between the

period 2017–2020. Performance on the VoxCeleb dataset
improved so dramatically—the last record-breaking model had
only a 0.52 percent equal error rate—that researchers are
currently trying to develop harder benchmarks to measure
continued progress. These results indicate that the bibliometric
findings of rapid growth in both research areas have in fact
translated to major gains in performance.42
The performance improvements illustrated above are similar to

those seen in other AI areas such as computer vision and natural
language processing. As displayed in Figure 12, the performance
of computer vision systems against ImageNet (a prominent image
classification benchmark dataset used to test a range of models)
advanced rapidly between 2013–2015, then exceeded human
performance in 2016.43 By 2018, improvement rates plateaued as
it neared the upper stratum of performance (97–98 percent), with
later performance increases coming from the addition of even
larger datasets and more powerful computational resources. This
is higher than re-identification performance (~69 mAP score),
likely because re-identification is a more complicated task.

Figure 12. Computer Vision Performance, 2013–2021 (ImageNet
Challenge: Top-5 Accuracy)
Source: Papers with Code, 2020; AI Index, 2021 | Chart: 2021 AI Index.
Natural language processing capabilities show similar

improvements in two prominent benchmarks: “GLUE” and the
updated “superGLUE.”44 These benchmarks only cover the years
2019–2021 because natural language processing only recently
emerged as an area of rapid, competitive AI development using
contemporary techniques.45 GLUE was an attempt to create a
benchmark that captured progress in natural language processing
systems broadly by testing single systems against multiple suites
of tests. However, as displayed in Figure 13, GLUE performance
was rapidly saturated, forcing the creators to make superGLUE, a
harder variant. Even this metric has become saturated recently. In
this way, the replacement of benchmarks highlights the progress
happening in the field and serves as a signifier for recent rapid
progress.

Figure 13. Natural Language Processing Performance, 2019–2021
95
89.9 90.3 90.6
90 88.4
90.3 90.6
89.3
85
SCORE
80.5
80
73.5
75
71.5
70
Jul-19 Nov-19 Mar-20 Jul-20 Nov-20 Mar-21 Jul-21
DATE
Score (GLUE) Score (superGLUE)
Source: Gluebenchmark “Leaderboard” and “Leaderboard Version: 2.0.”

Leveraging the Methodology
The previous sections described a bibliometric and technical

performance-based measurement approach that policymakers can
use to understand changes in AI. A key question remains—how
could the U.S. government leverage such a process to support
policymaking? Determining the optimal degree of government
involvement is outside the scope of this report, but we believe that
it should, at the very least, play a leading role in organizing and
resourcing private and academic entities that wish to participate,
as well as steering research efforts toward areas relevant to
policymakers. Notwithstanding the degree of involvement, we
offer three overarching suggestions for how this type of
methodology might be utilized and further improved with
government support.
1. Augment Technology Assessment with More Continuous

Analysis
Many attempts to measure AI involve periodic or ad-hoc

assessments, rather than more continuous analyses. There has
been a sustained surge in the publication of AI papers, and the
popularity of preprint repositories like arXiv means the publishing
cycles within this domain have accelerated. The methodology
outlined in the previous sections can help the U.S. government
take advantage of the information surfaced from these AI papers
and provide a way to generate a more continuous stream of
information about the state of particular AI subfields.
As an example for how this approach could complement ad-hoc

assessments, consider NIST's Face Recognition Vendor Test
(FRVT). NIST conducts periodic assessments of deployed facial
recognition systems' state-of-the-art technical capabilities. To
enhance these assessments, NIST could leverage the bibliometric
and technical performance-based methodology outlined above to
gain a more continuous understanding of facial recognition
research, instead of relying solely on intermittent assessments.
This process could be conducted biannually for bibliometrics to
incorporate new publications and quarterly for performance
metrics to incorporate new or improved benchmarks (in addition to

NIST-created benchmarks). Such an approach could also help
policymakers understand the gap between capabilities being
deployed in the world and those being demonstrated by
researchers, and might provide an early warning system to flag
major progress on a technical metric.
2. Create Shared Infrastructure
The bibliometrics analysis in the prior sections relies on CSET’s

Map of Science. But it is easy to imagine that a similar type of
resource, housed within or resourced by the United States
government, could become a useful tool for policymakers across
multiple agencies. We could imagine such a tool being used by
agencies such as the National Science Foundation, the National
Institutes of Health, DARPA, and IARPA to surface trends in
research activity. All these users will have different needs, but a
shared infrastructure could allow multiple agencies to coordinate
on areas of interest. It could also create multiple “demand signals”
for the developers of such a bibliometric system. A shared
platform that enables the collection, cleaning, and sharing of
bibliometrics data provides an important resource for federal
agencies and independent researchers to perform their own
analysis.46
The performance metrics assessment in the prior section relied on

the availability of common benchmarks. However, there are
machine learning topics that lack reliable metrics altogether, and
there is no single resource to readily attain existing metrics. This is
an area that could benefit from a shared repository for capturing
available metrics—just as bibliometrics are resourced into a single
database (e.g., the Map of Science).47 Such an infrastructure could
be continuously updated to incorporate new metrics and identify
short-lived or saturated benchmarks. The repository could also
identify policy-relevant machine learning technologies that
currently lack performance benchmarks. Candidates to support
and maintain the repository include a government organization like
NIST or a consortium of public and private participants. Ultimately,
this repository could be integrated with a bibliometrics
infrastructure, creating opportunities to apply methodologies like
the one described in this report.48

3. Enable Open-Source Analysis
Governments periodically convene technical experts to provide

advice on matters of technical change, opportunity, and hazard.
The methodology outlined in this report provides a means to
enable more nuanced policymaker-advisor communication and
deeper insight. Governments could construct or support optimized
research publication infrastructures, maintain or resource a
database of performance metrics, pair it with periodic assessment
of bibliometric trends, and then provide the system to technical
experts. By supporting the construction and sustainment of such a
tool and then sharing access to it with technical experts,
governments could more easily receive expert insights on an as-
needed basis, enable the development of low-cost public “insight
competitions” where members of the public could compete to
generate meaningful data from the tool, and share information
with the public to provide a sense of progress in a given area.

Authors
Jack Clark is a non-resident research fellow with CSET, where Kyle

A. Miller is a student research analyst, and Rebecca Gelles is a
data scientist.
Acknowledgments
The authors are grateful to C. Raymond Perrault at SRI

International, Kuansan Wang at Microsoft, and Robert Stojnic at
Meta AI for reviewing the report and providing feedback and
suggestions. We also thank our CSET colleagues John Bansemer
and Catherine Aiken for their guidance and comprehensive
reviews, Ben Buchanan (currently on leave), Igor Mikolic-Torreira,
Micah Musser, and Drew Lohn for their support in the research and
writing process, as well as Heeu Millie Kim for assistance with
data editing. Melissa Deng and Alex Friedland provided editorial
support.
© 2021 by the Center for Security and Emerging Technology. This

work is licensed under a Creative Commons Attribution-Non
Commercial 4.0 International License.
To view a copy of this license, visit

https://creativecommons.org/licenses/by-nc/4.0/.
Document Identifier: doi: 10.51593/20210008

Appendices
Appendix A: Methodology (Detailed)
Database—CSET’s Map of Science (Bibliometric Clustering): The

Map of Science is an aggregation of multiple bibliometrics datasets
that encompasses most scientific literature and is updated
quarterly to incorporate new publications and accrued citations,
thus enabling regular assessments. The literature is grouped into
research clusters, which are collections of research publications
linked by citations. Clustering creates representations of related
research, but the nature of the relationships within or between
clusters is indeterminate and requires manual evaluation.49 The
database is considered curated because it calculates indicators
that can inform analysis, such as a cluster’s relation to AI,
computer vision, and natural language processing. While the Map
of Science is currently under development and has limitations, this
report seeks to demonstrate its potential to automate and expedite
the identification and extraction of publication metadata on specific
machine learning–related topics. Metadata provides a range of
information, including keywords and phrases, sources and
citations, dates, authors, country and organization affiliations,
publication mediums, and subject tags. Once fully developed, the
database will likely include functions that automate much of the
manual data structuring required in this report.
Sub-Method—Bibliometric Analysis: This sub-method is used to

identify, extract, and analyze metadata from the Map of Science,
enabling the measurement of trends in specific machine learning–
related topics. The first step involves selecting a distinct topic to
investigate (before using the Map of Science), framing what it
encompasses, reviewing publications related to it on open-access
repositories (e.g., arXiv), and collecting 10–20 “seed papers” that
encapsulate it. This requires a degree of domain expertise, as the
classification of “quality” seeds varies depending on the topic.
Seed papers should include highly cited works from prominent
authors that were published over different years and must
encompass all key elements of the topic being investigated—
which often requires “incorporating” any highly relevant terms or

topics into the investigation to establish sufficient distinction and
framing.
Step two involves correlating the seed papers with clusters in the
Map of Science, identifying the cluster or clusters that sufficiently
encompass the topic, scanning neighboring clusters to corroborate
topic distinction and provide context, omitting extraneous
information, and then analyzing the extracted metadata for
strategic insight. Adjustments can be made to the bibliometric
analysis depending on the ambiguity of the topic, distribution of
seed papers across clusters, quality of the data, or expertise of the
user. For example, instead of collecting seed papers at the outset,
non-experts could use keyword searches in the Map of the Science
to identify relevant clusters, leverage automated data analytics
functions in the Map of Science to identify prominent (e.g., highly
cited) papers in the clusters, and then manually inspect the papers.
This sub-method can be further optimized, such as by recruiting
domain experts to tailor versions of bibliometric analysis for
different machine learning research areas and policy-related
inquiries, or incorporating natural language processing models to
automate and scale the analysis of full-body publication text.
These improvements are a vector for further research, as they are
outside the scope of this report.
Sub-Method—Performance Metrics Assessment: This sub-method

is used to assess the technical performance of a given machine
learning–enabled system or model, corroborate the insight from
bibliometric analysis, provide decisionmakers with information that
relates to the capabilities of a particular aspect of machine
learning, and more directly aid in identifying significant changes
and advancements. The process revolves around manually
tracking the technical performance metrics in publications (from
clusters identified during the bibliometric analysis). Metrics are
designed by researchers to assess particular machine learning
systems and solve specific problems, and are often released
publicly to be used and updated by the community. This allows us
to select a distinct topic for investigation, manually collect and
compare particular metrics in different publications across
repositories like arXiv and Papers with Code, chart a timeline of

performance development, and assess current performance
benchmarks—if the metrics are available and reliable. A
performance assessment works well when there are existing
benchmarks with track records of testing and evaluation, but can
be more challenging if there are no relevant metrics available or
they are outdated (i.e., saturated). However, identifying topics that
lack metrics can help illuminate knowledge gaps in policy-relevant
areas that require further research. Although this process requires
extensive domain expertise and manual inspection of publications,
it could be more streamlined if a maintained database of
categorized metrics was available.

Appendix B: Seed Papers
Re-Identification Seed Papers:
● Rodolfo Quispe and Helio Pedrini, “Top-DB-Net: Top

DropBlock for Activation Enhancement in Person Re-
Identification,” arXiv preprint arXiv:2010.05435v1 (2020),
https://arxiv.org/abs/2010.05435v1.
● Xingyang Ni, Liang Fang, and Heikki Huttunen,

“AdaptiveReID: Adaptive L2 Regularization in Person Re-
Identification,” arXiv preprint arXiv:2007.07875v1 (2020),
● Changxing Ding et al., “Multi-task Learning with Coarse

Priors for Robust Part-aware Person Re-identification,”
arXiv preprint arXiv:2003.08069v1 (2020),
● S.V. Aruna Kumar et al., “P-DESTRE: The P-DESTRE: A

Fully Annotated Dataset for Pedestrian Detection, Tracking,
Re-Identification and Search from Aerial Devices,” arXiv
preprint arXiv:2004.02782 (2020),
https://arxiv.org/abs/2004.02782.
● Jianyang Gu et al., “Long-Term Cloth-Changing Person Re-

identification,” arXiv preprint arXiv:2012.13498 (2020),
● Jianyang Gu et al., “1st Place Solution to VisDA-2020: Bias

Elimination for Domain Adaptive Pedestrian Re-
identification,” arXiv preprint arXiv:2012.13498 (2020),
● Ammarah Farooq et al., “A Convolutional Baseline for

Person Re-Identification Using Vision and Language
Descriptions,” arXiv preprint arXiv:2003.00808 (2020),
● Zhedong Zheng et al., “Joint Discriminative and Generative

Learning for Person Re-Identification,” arXiv preprint

arXiv:1904.07223v1 (2019),
● Guan'an Wang et al., “Color-Sensitive Person Re-

Identification,” 2019,
https://www.ijcai.org/Proceedings/2019/131.
● Hui Li et al., “Pedestrian re-identification based on Tree

branch network with local and global learning,” arXiv
preprint arXiv:1904.00355 (2019),
● Shi-Zhe Chen, Chun-Chao Guo, and Jian-Huang Lai, “Deep

Ranking for Person Re-Identification via Joint
Representation Learning,” arXiv preprint arXiv:1505.06821
(2016), https://arxiv.org/abs/1505.06821.
● Brian H. Wang et al., “Deep Person Re-identification for

Probabilistic Data Association in Multiple Pedestrian
Tracking,” arXiv preprint arXiv:1810.08565 (2018),
● Jianfu Zhang, Naiyan Wang, and Liqing Zhang, “Multi-shot

Pedestrian Re-identification via Sequential Decision
Making,” arXiv preprint rXiv:1712.07257v1 (2017),
● Ejaz Ahmed, Michael Jones, and Tim K. Marks, “An Improved

Deep Learning Architecture for Person Re-identification,”
2015 IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), 2015,
https://ieeexplore.ieee.org/document/7299016.
Speaker Recognition Seed Papers:
● Xugang Lu, Peng Shen, Yu Tsao, and Hisashi Kawai,

“Integrating a Joint Bayesian Generative Model in a
Discriminative Learning Framework for Speaker
Verification,” arXiv preprint arXiv:2101.03329 (2021),

● Jixuan Wang et al., “Speaker Attribution with Voice Profiles
by Graph-based Semi-supervised Learning,” arXiv preprint
arXiv:2102.03634 (2021), https://arxiv.org/abs/2102.03634.
● Wei Yao et al., “Multi-stream Convolutional Neural Network

with Frequency Selection for Robust Speaker Verification,”
Version 2, arXiv preprint arXiv:2012.11159v2 (2021),
● Wei Yao et al., “Multi-stream Convolutional Neural Network

with Frequency Selection for Robust Speaker Verification,”
Version 1, arXiv preprint arXiv:2012.11159v1 (2020),
● Arsha Nagrani et al., “Voxceleb: Large-scale speaker

verification in the wild,” October 2019,
https://www.robots.ox.ac.uk/~vgg/publications/2019/Nagra
ni19/nagrani19.pdf.
● Hemlata Tak et al., “End-to-end anti-spoofing with

RawNet2,” arXiv preprint arXiv:2011.01108 (2020),
● Deep CNNs With Self-Attention for Speaker Identification,

(2019),
https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=
8721628.
● VoxCeleb2: Deep Speaker Recognition, (2018),

https://www.robots.ox.ac.uk/~vgg/publications/2018/Chung
18a/chung18a.pdf.
● Weicheng Cai, Jinkun Chen, and Ming Li, “Exploring the

Encoding Layer and Loss Function in End-to-End Speaker
and Language Recognition System,” arXiv preprint
● David Snyder et al., “X-Vectors: Robust DNN Embeddings

for Speaker Recognition," 2018 IEEE International
Conference on Acoustics, Speech and Signal Processing
(ICASSP), 2018, https://www.semanticscholar.org/paper/X-

Vectors%3A-Robust-DNN-Embeddings-for-Speaker-
Snyder-Garcia-
Romero/389cd9824428be98a710f5f4de67121a70c15fd3.
● Suwon Shon, Hao Tang, and James Glass, “Frame-Level

Speaker Embeddings for Text-Independent Speaker
Recognition and Analysis of End-to-End Model,” arXiv
preprint arXiv:1809.04437 (2018),
● D Snyder et al., “Deep Neural Network Embeddings for

Text-Independent Speaker Verification,” (2017),
INTERSPEECH, http://refhub.elsevier.com/S0885-
2308(19)30271-2/sbref0055.
● Gautam Bhattacharya, Jahangir Alam, and Patrick Kenny,

“Deep Speaker Embeddings for Short-Duration Speaker
Verification,” INTERSPEECH, August 2017,
https://www.isca-
speech.org/archive/Interspeech_2017/pdfs/1575.PDF.
● Mitchell McLaren et al., “The Speakers in the Wild (SITW)

Speaker Recognition Database,” INTERSPEECH, 2016,
https://www.semanticscholar.org/paper/The-Speakers-in-
the-Wild-(SITW)-Speaker-Recognition-McLaren-
Ferrer/3fe358a66359ee2660ec0d13e727eb8f3f0007c2.
● L. Burget et al, “Discriminatively Trained Probabilistic Linear

Discriminant Analysis for Speaker Verification,” 2011 IEEE
International Conference on Acoustics, Speech and Signal
Processing (ICASSP), 2011,
https://www.semanticscholar.org/paper/0cf0a9f92e252f10
bb31a56da19fbe3b71bf1151.
Image Synthesis Seed Papers:
● Simranjeet Singh, Rajneesh Sharma, and Alan F. Smeaton,

“Using GANs to Synthesise Minimum Training Data for
Deepfake Generation,” arXiv preprint arXiv:2011.05421

● Shilong Shen, “Correspondence Learning for Controllable
Person Image Generation,” arXiv preprint arXiv:2012.12440
● Janet Rafner et al., “The Power of Pictures: Using ML

Assisted Image Generation to Engage the Crowd in
Complex Socioscientific Problems,” arXiv preprint
● Xiangrui Xu, Yaqin Li, and Cao Yuan, “Conditional Image

Generation with One-Vs-All Classifier,” arXiv preprint
● Thanh Thi Nguyen et al., “Deep Learning for Deepfakes

Creation and Detection: A Survey,” arXiv preprint
arXiv:1909.11573v1 (2019),
● Sakshi Agarwal and Lav R. Varshney, “Limits of Deepfake

Detection: A Robust Estimation Viewpoint, arXiv preprint
● Andrew Brock, Jeff Donahue, and Karen Simonyan, “Large

Scale GAN Training for High Fidelity Natural Image
Synthesis,” arXiv preprint arXiv:1809.11096 (2019),
● Jun-Yan Zhu et al., “Visual Object Networks: Image

Generation with Disentangled 3D Representation,” arXiv
preprint arXiv:1812.02725 (2018),
● Pavel Korshunov and Sebastien Marcel, “DeepFakes: A New

Threat to Face Recognition? Assessment and Detection,”
arXiv preprint arXiv:1812.08685 (2018),
● Mevlana Gemici, Zeynep Akata, and Max Welling, “Primal-

Dual Wasserstein GAN,” arXiv preprint arXiv:1805.09575

● Alexey Dosovitskiy and Thomas Brox, “Generating Images
with Perceptual Similarity Metrics based on Deep
Networks,” arXiv preprint arXiv:1602.02644 (2016),
● Alec Radford, Luke Metz, and Soumith Chintala,

“Unsupervised Representation Learning with Deep
Convolutional Generative Adversarial Networks,” arXiv
preprint arXiv:1511.06434 (2015),

Appendix C: Notes on Terminology
Re-Identification Terminology: Researchers use a variety of terms

to discuss re-identification (RE-ID), as is the case with many
emerging machine learning–enabled capabilities and systems. Due
to this varying taxonomy, we must “incorporate” any highly
relevant terms or topics into the investigation to maintain sufficient
distinction and framing. In addition to the term RE-ID, we
incorporated seven related terms into the investigation: Person
RE-ID, Pedestrian RE-ID, Object RE-ID, Vehicle RE-ID, Human RE-
ID, Pedestrian Tracking, and Pedestrian Retrieval. For the context
of this report, we consider all of them synonymous and refer to
them collectively through the term “re-identification.” Additionally,
we must note that RE-ID is distinguished from “gait recognition,”
which is the task of identifying individuals based on their unique
manner of moving and walking (although both capabilities can be
integrated). Domain expertise and an analysis of prominent RE-ID
papers informed the selection of these terms.
Table 7. Re-Identification Cluster #1419 Topic Relevance Spot-

Check (via Keywords in Publication Titles)
RE-ID (Cluster #1419) Terminology Keyword Search Matches
Re-Identification 2,563
Person Re-Identification 2,128
Pedestrian Re-Identification 72
Object Re-Identification 5
Vehicle Re-Identification 157
Human Re-Identification 25
Pedestrian Tracking 4
Pedestrian Retrieval 9
Total Matches 4,963

Speaker Recognition Terminology: In addition to the term speaker
recognition (SR), we incorporated four related labels into the
investigation: Speaker Verification, Speaker Identification, Voice
Identification, and Speaker Diarization. We must note that there
are specific elements that distinguish these terms. “SR” broadly
relates to identifying speakers, “Speaker Verification” is verifying a
speaker identity that was previously learned, “Speaker
Identification” is identifying an unknown speaker that is in a group
of known speakers, “Voice Identification” is identifying a known
voice, and “Speaker Diarization” is a process for partitioning audio
to identify who spoke when.50 For the context of this report, we
refer to all these labels collectively through the term “SR.”
Additionally, we must note that SR is distinguished from “speech
recognition,” which is the task of identifying and converting words
from audio to text. As with RE-ID terminology, domain expertise
and the analysis of SR papers informed the selection of these
terms.
Table 8. Speaker Recognition Cluster #6855 Topic Relevance

Spot-Check (via Keywords in Publication Titles)
SR (Cluster #6855) Terminology Keyword Search Matches
Speaker Recognition 539
Speaker Verification 653
Speaker Identification 97
Voice Identification 0
Speaker Diarization 46
Total Matches 1,335

Appendix D: Progression of Key Machine Learning Technologies
Re-Identification Progression: The early years of 2010–2013 had

steady growth rates, but the overall quantity of papers and
citations was rather limited. During that period the key
components (e.g., cost and availability of data and computational
capacity) and model architectures (e.g., deep learning approaches
to computer vision) for re-identification were in their technological
infancy. Significant advancements in machine learning computer
vision occurred around 2012, which was followed by increased
growth in re-identification publications between 2015–2018. This
six-year period saw the discovery and proliferation of many key
machine learning subcomponents that enable re-identification,
including deep convolutional neural networks (CNN), residual
networks, and general techniques like pre-training, where you
train a general-purpose computer vision neural network on a large
amount of generic data, then fine-tune it on more specific datasets.
Following these advancements, the research community went
from producing dozens of re-identification papers over multiple
years to hundreds annually (although growth started to plateau in
2020).51
Speaker Recognition Progression: The early years of 2010–2014

had steady growth rates, but the overall quantity of papers and
citations was limited, a dynamic that parallels re-identification
development. Many of the same machine learning advancements
that enabled the proliferation of re-identification also enabled
speaker recognition, including deep CNNs and increased data and
compute capacity. Developments in natural language processing
learning techniques, such as Transformers, played a significant
role in the accelerated speaker recognition growth after 2017.
During this timeframe, the research community went from
producing dozens of speaker recognition papers to producing
hundreds annually. Alongside, and intertwined with, these
advancements was the proliferation of “smart” devices (e.g.,
smartphones), programs (e.g., Apple’s Siri), and services (e.g.,
audio recording) that could use speaker recognition.52

Appendix E: Organizational Affiliation Data (Detailed)
Re-Identification Organizational Affiliations: There were over 600

organizations affiliated with an author of at least one re-
identification-related publication, of which approximately 76
percent were in the education sector, 9 percent in government, and
6 percent in private industry. We identified six “primary
organizations,” meaning they were ranked in the top ten for both
publication and citation count: (1) Chinese Academy of Sciences,
(2) Queen Mary University of London, (3) Sun Yat-sen University,
(4) University of Technology, Sydney, (5) Peking University, and (6)
Tsinghua University. All these entities are educational and highly
affiliated with their respective governments.
China was home to the most organizations affiliated with the

authors of re-identification papers in our cluster. Of the top 50
organizations affiliated with re-identification publications, 35 were
located in China, five in the United States, and two in both
Australia and Italy. Of the top 50 cited organizations, 22 were
located in China, twelve in the United States, three in both France
and Italy, and two in Australia, Austria, and Singapore.
Most of the top organizations were public research universities

that have significant government affiliations, along with varying
affiliations with private entities. For example, the Chinese
Academy of Sciences is a national academy fully supported by
Chinese government entities. When we focus on the rankings
amongst private industry organizations, the data indicates that
Microsoft (United States), Alibaba Group (China), Tencent (China),
Vision Semantics Limited (UK), and SenseTime (China) were the
most affiliated with re-identification papers. Notably, Microsoft
conducts much of its re-identification-related research through the
Beijing-based “Microsoft Research Asia,” the company’s largest
research lab outside of the United States. This information can aid
in a range of governmental regulatory and collaborative efforts, as
well as inform security considerations regarding the international
proliferation of technologies designed by domestic companies
(e.g., Microsoft).

Speaker Recognition Organizational Affiliations: There were over
400 organizations affiliated with an author of at least one speaker
recognition publication, of which approximately 65 percent were in
the education sector, 11 percent in private industry, 5 percent in
government, and 4 percent non-profit. This is a similar dynamic to
re-identification, although with fewer total organizations and
greater private affiliation. Speaker recognition also has different
primary organizations ranked in the top 10 for both paper
affiliations and citations of affiliated papers: (1) Johns Hopkins
University, (2) MIT, (3) Brno University of Technology, and (4) SRI
International. In contrast to re-identification, speaker recognition
primary organizations were a mix of private, public, and nonprofits,
most of which are not directly affiliated with their respective
governments.
The geographic distribution of speaker recognition papers was

more heterogeneous than re-identification. Of the top 50
organizations affiliated with authors of speaker recognition papers,
13 were located in the United States, 11 in China, three in Canada,
and two in Switzerland, Spain, South Korea, Singapore, India,
Australia, and Argentina. Of the top 50 organizations ranked by
the citation counts of their affiliated papers, 17 were in the United
States, six in China, four in Spain and Canada, and three in the UK.
Most of the top organizations were public universities with

significant government affiliations, notwithstanding outliers such
as the private Johns Hopkins University (most publications) and
the nonprofit Computer Research Institute of Montreal (most-
cited). For example, Tsinghua University, the University of Texas at
Dallas, and Hong Kong Polytechnic University were each public
universities in the top ten for speaker recognition publications.
When focused on rankings among private industry organizations,
we see that IBM, Microsoft, NEC Corporation, Tencent, and Google
were the most affiliated with speaker recognition.

Appendix F: Metrics Assessment (Detailed)
Re-Identification Metrics Assessment: The assessment of re-

identification concentrates on performance against the MSMT17
dataset using the mean Average Precision (mAP) metric. We chose
this dataset because: it is relatively new (introduced in 2017 by
researchers at Peking University and the University of Texas at San
Antonio), has not yet been saturated, involves some of the latest
techniques, and at the time of its creation was designed to be
challenging relative to existing benchmarks.53 Therefore, we are
confident that the performance on MSMT17 is reflective of real
progress, rather than progress on an easy benchmark.
Five publications informed our assessment of re-identification

advancement between 2018–2021 (see Appendix G for details),
four of which were affiliated with China, two with the United
States, and one with both France and Australia. These were
chosen because the researchers tested their systems against
MSMT17.
We extracted two insights from the progress of re-identification

models’ mAP performance against the MSMT17 dataset:
1. Re-identification performance has improved significantly,

with a high rate of progress between 2018–2019 but
slower improvements between 2020–2021. Notably, the
metric improvements correlated with the increased rate of
re-identification publication, as seen through the
bibliometric metadata analysis. Finding out where further
improvements will come from could be important, as
subsequent investigations of yet-to-exist results will tell us
whether performance growth is stagnating or if re-
identification is able to successfully harness further
breakthroughs from computer vision writ large to further
expand performance (both in test environments and its real-
world application). This insight warrants attention from
policymakers as a continuation of this trend will see re-
identification systems advance to a degree that makes them
more efficient and reliable—increasing incentives for both
government and private entities to adopt the technology.

2. Researchers driving this progress correlate with those we
would expect to see from our bibliometric analysis—four of
the performance data points were created by China-
affiliated researchers and organizations, while two were
affiliated with the United States. This corroborates the
insights derived from our bibliometric analysis of re-
identification, which indicated that both are the most
prominent countries affiliated with re-identification
publications and citations. This also suggests that studying
technical metrics can “sanity check” the insights derived
from bibliometrics. However, a more comprehensive
application of the performance metrics assessment is
necessary to determine the extent to which metrics
affiliations correlate with the distribution of country
affiliations seen in the bibliometrics metadata.
Speaker Recognition Metrics Assessment: The assessment of

speaker recognition concentrates on performance against the
VoxCeleb dataset using the Equal Error Rate (EER) metric. We
chose VoxCeleb because, like MSMT17, it is relatively new
(introduced in 2017 by researchers at the University of Oxford).
Unlike MSMT17, VoxCeleb is also accompanied by an annual
competition that sees multiple groups compete to test the
performance of their systems against a standardized test (in many
ways, VoxCeleb is to speaker recognition what ImageNet is to
object detection)—this gives us a pre-structured way to gather
data for the technical analysis and provides some inbuilt auditing
of technical results by competition administrators.
Six publications informed our assessment of speaker recognition

advancement between 2017–2020 (see Appendix G for details),
four of which were affiliated with the UK and one with both
Belgium and China. These were chosen because they were the
highly ranked solutions for the VoxCeleb competitions.
We extracted two insights from the progress of speaker

recognition models’ EER performance against the VoxCeleb
dataset:

1. Speaker recognition performance has improved
significantly, with the EER decreasing (i.e., increasing model
precision) from ~8 percent in 2017 to ~0.5 percent in 2021.
This performance is so good that VoxCeleb evaluations
have, in recent years, moved to testing against harder and
more complicated underlying datasets. This reflects a
similar pattern of development that occurred in computer
vision where, after results started to stagnate on ImageNet,
researchers developed harder variants of the dataset to test
against (e.g., ImageNet Adversarial).54 Notably, some of
these improvements occurred alongside the development of
key natural language processing learning techniques (e.g.,
using Transformers) and correlated with the increase in
speaker recognition publications, as seen through our
bibliometric analysis. This insight warrants attention from
policymakers, as a continuation of this trend will see
speaker recognition systems advance to a degree that could
supplement intelligence capabilities in profound ways (e.g.,
the recognition of millions of speakers in perpetuity).
2. Improvements in state-of-the-art performance on VoxCeleb

have primarily been driven by a research group at the
University of Oxford (UK) that created the dataset and
challenge, Ghent University (Belgium), and most recently
SpeakIn Technologies (China). Compared to re-
identification, these speaker recognition metrics did not
correlate as much with the distribution of the top country
affiliations that were analyzed in the bibliometric analysis—
the UK and Belgium were not amongst the most prominent
countries affiliated with authors of speaker recognition
publications. This is an area for further analysis, as it neither
corroborates nor invalidates the insight from publication
metadata. However, it does illustrate how metadata alone
cannot capture important technical elements of machine
learning research and further highlights how pairing
technical metrics with bibliometric analysis can “sanity
check” the findings of different analyses.

Further metrics assessment is needed to better determine country
affiliation. First, the results might superficially indicate that the UK
and Belgium are leading performance on speaker recognition
applied to VoxCeleb1, but if we analyze more results and adopt a
cohort-based analysis technique—that is, look at the people that
came in second and third place in the VoxCeleb challenge each
year—we can get a sense of the broader landscape. Here, teams
associated with the United States and China appeared within the
competition as well. This suggests that, although a performance
assessment of state-of-the-art speaker recognition models can
yield insight, a more comprehensive application of the metrics
assessment is necessary to get a clearer signal and context.
Second, researchers from various countries may be testing out
their systems against VoxCeleb without publishing the results,
due to the commercial and security value of speaker recognition
capabilities.

Appendix G: Performance Metrics Data and Sources
Table 9. Re-Identification Performance Metrics Data and Sources
Country
Date mAP Method Organization(s) Affiliation Publication Title
TransReID:
Transformer-based
Alibaba Group; Object Re-
3/26/2021 69.4 TransReID Zhejiang University China Identification
Deep Miner: A Deep

and Multi-branch
Network which Mines
Rich and Diverse
Digeiz AI Lab; Ecole Features for Person
2/18/2021 67.3 Deep Miner Polytechnique France Re-identification
Multi-task Learning
Multi-task South China with Coarse Priors for
Part-aware University of Robust Part-aware
Network Technology; China; Person Re-
9/21/2020 62.7 (MPN) University of Sydney Australia identification
Texas A&M
University;
University of Science
and Technology of
China; Walmart
Technology; ABD-Net: Attentive
Wormpex AI United but Diverse Person
8/9/2019 60.8 ABD-Net Research States; China Re-Identification
Person Transfer GAN

Peking University, to Bridge Domain Gap
University of Texas China; United for Person Re-
9/13/2018 34 GLAD at San Antonio States Identification
Source: Chuting He et al.; Abdallah Benzine et al.; Changxing Ding et al.;

Tianlong Chen et al.; Wei et al. 55

Table 10. Speaker Recognition Performance Metrics Data and
Sources
Country
Date EER Method Organization Affiliation Publication Title
The SpeakIn System for

SpeakIn VoxCeleb Speaker
Technologies Recognition Challenge
9/5/2021 0.52% QMF Co. Ltd. China 2021
The IDLAB VoxSRC-20

Submission: Large Margin
Fine-Tuning and Quality-
Speech Aware Score Calibration in
duration Ghent DNN Based Speaker
10/20/2020 0.56% QMF University Belgium Verification
Ours + Voxceleb: Large-Scale

Relation University of Speaker Verification in the
10/16/2019 2.87% Module Oxford UK Wild
Utterance-level
University of Aggregation for Speaker
5/17/2019 3.22% GhostVLAD Oxford UK Recognition In The Wild
University of VoxCeleb2: Deep Speaker

6/27/2018 4.19% ResNet50 Oxford UK Recognition
VoxCeleb: A Large-Scale
CNN256D- University of Speaker Identification
6/26/2017 7.80% Embedding Oxford UK Dataset
Note: A lower EER score signifies less error and greater accuracy.
Source: Nagrani et al.; Joon Son Chung, Arsha Nagrani, and Andrew Zisserman;
Weidi Xie et al.; Arsha Nagrani et al.; Jenthe Thienpondt, Brecht Desplanques,
and Kris Demuynck; Miao Zhao et al. 56

Endnotes
1
For more literature on the field of bibliometrics, see the following sources:
Office of Management, “Learn More About Bibliometrics,” National Institutes of
Health (NIH), https://www.nihlibrary.nih.gov/services/bibliometrics/learn-more-
about-bibliometrics; Ashok Agarwal et al., “Bibliometrics: Tracking Research
Impact by Selecting the Appropriate Metrics,” Asian Journal of Andrology 18,
no. 2 (January 2016): 296-309,
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4770502/; Anthony van Raan,
“Measuring Science: Basic Principles and Application of Advanced
Bibliometrics,” in Springer Handbook of Science and Technology Indicators
(2019): 237-280, https://link.springer.com/chapter/10.1007/978-3-030-02511-
3_10; Bibliometrics involves the “quantitative evaluation of scientific articles and
other published works, including the authors of articles, the journals where the
works were published, and the number of times they are later cited,” A.W.
Jones, “Forensic Journals: Bibliometrics and Journal Impact Factors,”
Encyclopedia of Forensic and Legal Medicine (Second Edition) (2016),
https://www.sciencedirect.com/science/article/pii/B9780128000342001816#.
2
This is not an attempt to forecast future developments, an in-depth analysis of
a particular aspect of AI, nor a comprehensive review or critique of other
research efforts to measure AI. Rather, it is an outline of a systematic process to
measure AI developments and a demonstration that we hope will inspire
policymakers to implement and improve the methodology. Additionally, though
none of the elements of the methodology are novel, by integrating them
together we can outline a process to improve measurement.
3
National Institute for Standards and Technology (NIST), “The EMNIST
Dataset,” U.S. Department of Commerce, https://www.nist.gov/itl/products-and-
services/emnist-dataset; Defense Advanced Research Projects Agency
(DARPA), “The DARPA Grand Challenge: Ten Years Later,” March 13, 2014,
https://www.darpa.mil/news-events/2014-03-13; Defense Advanced Research
Projects Agency (DARPA), “DARPA Robotics Challenge (DRC),”
https://www.darpa.mil/program/darpa-robotics-challenge.
4
The efforts of this report are similar to parts of the Intelligence Advanced
Research Projects Activity’s (IARPA) Foresight and Understanding from
Scientific Exposition (FUSE) program. The FUSE program ran from 2010–2017
and was intended to “enable reliable, early detection of emerging scientific and
technical capabilities across disciplines and languages found within the full-text
content of scientific, technical, and patent literature” and “discover patterns of
emergence and connections between technical concepts at a speed, scale, and
comprehensiveness that exceeds human capacity.” Like this report, FUSE
emphasized the need for a methodology that incorporated continuous
assessment of bibliometrics over ad-hoc assessment. However, the FUSE

program intended to design a fully refined and automated methodology that
involved six “research thrusts” (theory development, document features,
indicator development, nomination quality, evidence representation, and system
engineering), while this report’s objectives are far more limited, as we only
intend to demonstrate the potential for insight through a prototype
methodology that can be adopted and updated by different end users.
Additionally, the methodology in this report incorporates performance metrics
assessment and an AI-curated corpus of scientific literature, which are elements
not included in IARPA’s FUSE program. For more details, see Intelligence
Advanced Research Projects Activity (IARPA), “Foresight and Understanding
from Scientific Exposition (FUSE),” Office of the Director of National Intelligence,
https://www.iarpa.gov/index.php/research-programs/fuse.
5
The CSET Map of Science is a merged corpus of Digital Science’s Dimensions,
Clarivate’s Web of Science, Microsoft Academic Graph (MAG), China National
Knowledge Infrastructure (CNKI), arXiv, and Papers with Code. As of September
2021, it included approximately 130 million papers, 1.4 billion citation linkages,
and 123 thousand research clusters. See also the Map of Science user interface,
Jennifer Melot and Ilya Rahkovsky, “CSET Map of Science” (Center for Security
and Emerging Technology, October 2021),
https://cset.georgetown.edu/publication/cset-map-of-science/. We must note
that the public user interface was released in October 2021 but the data in this
report was extracted from the Map of Science in March 2021, therefore the
clusters addressed in this report do not reflect clusters in the current iteration of
the Map of Science; Autumn Toney, "Creating a Map of Science and Measuring
the Role of AI in it" (Center for Security and Emerging Technology, June 2021),
https://cset.georgetown.edu/publication/creating-a-map-of-science-and-
measuring-the-role-of-ai-in-it//.
6
For more information on the methodology behind CSET’s research clusters, see
Ilya Rahkovsky et al., “AI Research Funding Portfolios and Extreme Growth”
(Center for Security and Emerging Technology, April 2021),
https://www.frontiersin.org/articles/10.3389/frma.2021.630124/full; For an
explanation of the methodology for classifying papers as AI-relevant, see James
Dunham, Jennifer Melot, and Dewey Murdick, “Identifying the Development and
Application of Artificial Intelligence in Scientific Text,” arXiv preprint
arXiv:2002.07143 (2020), https://arxiv.org/abs/2002.07143; For more
information on clustering methodology and the clustering model, see Kevin W.
Boyack, Caleb Smith, and Richard Klavans, “A detailed open access model of the
PubMed literature,” Scientific Data 7 (November 2020),
https://www.nature.com/articles/s41597-020-00749-y; For a “Data Snapshot”
that explores the underlying data and analytic utility of the CSET Map of
Science, see Toney, “Creating a Map of Science and Measuring the Role of AI in
it.”

7
We primarily use the term “machine learning” instead of “artificial intelligence”
for the remainder of the report because AI is too broad (e.g., this investigation
does not encompass expert systems).
8
There are preliminary systems being built that help automate this, ranging
from expert-led syntheses of the field (e.g., the AI Index) to community-curated
resources aided by automated extraction of performance metrics from papers
(e.g., Papers with Code). Additionally, CSET is currently working to extract more
data from papers and clusters in the Map of Science, including but not limited to
compute use, libraries use, source code, and datasets—all of which could further
automate and improve the methodology presented in this report.
9
“ImageNet,” https://www.image-net.org/.
10
Mang Ye et al., “Deep Learning for Person Re-identification: A Survey and
Outlook,” arXiv preprint arXiv:2001.04193 (2020),
11
Re-identification systems can track individuals in anonymous and non-
anonymous capacities. There is also a growing field of research into
unsupervised re-identification—that is, systems that automatically identify and
track individuals seen in streaming video data without needing a prior label and
image.
12
More specifically, speaker recognition is an emerging voice biometry capability
that is enabled by deep machine learning models trained on labeled and
unlabeled audio data, while speaker recognition systems are designed to
recognize the voices of specific speakers in an audio sample or live feed. The
speaker recognition process is first focused on training the model through
ingesting labeled data, such as the voice audio of a speaker that is linked to said
speaker’s name or some other identifier. This process “enrolls” a speaker’s voice
into a unique signature that the model can recognize. Once speaker signatures
are learned to an ascertainable degree, speaker recognition systems can then be
deployed in a variety of capacities to identify audio inputs and correlate them to
known individuals. For more information, see “What is Speaker Recognition
(Preview)?” Microsoft, November 2021, https://docs.microsoft.com/en-
us/azure/cognitive-services/speech-service/speaker-recognition-overview.
13
Note that speaker recognition differs from speech recognition, which focuses
on identifying and converting words from audio to text.
14
J.J. Godfrey, E.C. Holliman, and J. McDaniel, “Switchboard: Telephone Speech
Corpus for Research and Development,” ICASSP-92: 1992 IEEE International
Conference on Acoustics, Speech, and Signal Processing, 1992,
https://www.semanticscholar.org/paper/SWITCHBOARD%3A-telephone-

speech-corpus-for-research-Godfrey-
Holliman/d80000d84223e177d070a01a734dba56d5f5c069.
15
More specifically, image synthesis is a matured computer vision capability
enabled by deep generative machine learning models and frameworks (e.g.,
GANs and variational autoencoders) that are trained on large imagery datasets.
It is a relatively broad term that encompasses many techniques and systems for
creating computer-generated images. The general purpose of this capability
contrasts with re-identification and speaker recognition, which involves the
discrimination, identification, and recognition of phenomena—not the creation of
synthetic media.
16
A GAN is an architecture of two “competing” neural networks, a generator
and a discriminator, that can be used for many purposes. In image synthesis, the
model is trained on labeled and unlabeled imagery data, which is fed into a
generator neural network that extracts features and learns to generate synthetic
images based on them. These generated images are iteratively tested against a
second neural network that is designed to discriminate between synthetic and
organic media and acts as a feedback mechanism to optimize the generator.
17
Sarah Cahlan, “How Misinformation Helped Spark an Attempted Coup in
Gabon,” The Washington Post, February 13, 2020,
https://www.washingtonpost.com/politics/2020/02/13/how-sick-president-
suspect-video-helped-sparked-an-attempted-coup-gabon/.
18
Charlotte Jee, “An Indian Politician Is Using Deepfake Technology to Win New
Voters,” MIT Technology Review, February 19, 2020,
https://www.technologyreview.com/2020/02/19/868173/an-indian-politician-
is-using-deepfakes-to-try-and-win-voters/; Benjamin Strick, “West Papua: New
Online Influence Operation Attempts to Sway Independence Debate,”
Bellingcat, November 11, 2020,
https://www.bellingcat.com/news/2020/11/11/west-papua-new-online-
influence-operation-attempts-to-sway-independence-debate/.
19
Ian J. Goodfellow et al., “Generative Adversarial Networks,” arXiv preprint
arXiv:1406.2661 (2014), https://arxiv.org/abs/1406.2661; Alec Radford, Luke
Metz, and Soumith Chintala, “Unsupervised Representation Learning with Deep
Convolutional Generative Adversarial Networks,” arXiv preprint
arXiv:1511.06434 (2015), https://arxiv.org/abs/1511.06434; Ming-Yu Liu and
Oncel Tuzel, “Coupled Generative Adversarial Networks,” arXiv preprint
arXiv:1606.07536 (2016), https://arxiv.org/abs/1606.07536; Tero Karras et al.,
“Progressive Growing of GANs for Improved Quality, Stability, and Variation,”
arXiv preprint arXiv:1710.10196 (2017), https://arxiv.org/abs/1710.10196; Tero
Karras, Samuli Laine, and Timo Aila, “A Style-Based Generator Architecture for
Generative Adversarial Networks,” arXiv preprint arXiv:1812.04948 (2018),
https://arxiv.org/abs/1812.04948; Yunjey Choi et al., “StarGAN v2: Diverse

Image Synthesis for Multiple Domains,” arXiv preprint arXiv:1912.01865 (2019),
20
All re-identification and speaker recognition publication metadata displayed in
this report was queried via BigQuery and extracted from the Map of Science
clusters in March 2021. Subsequent updates to the Map of Science after this
date will likely alter the clusters. All papers published before January 2010 and
after December 2020 were omitted from the bibliometric analysis.
21
The model determines which publications are “related” to AI broadly by
predicting the categories that authors assign to their publications. It is “a
strategy for identifying the universe of research publications relevant to the
application and development of artificial intelligence. The approach leverages
the arXiv corpus of scientific preprints, in which authors choose subject tags for
their papers from a set defined by editors. We compose a functional definition of
AI relevance by learning these subjects from paper metadata, and then inferring
the arXiv-subject labels of papers in larger corpora: Clarivate Web of Science,
Digital Science Dimensions, and Microsoft Academic Graph,” Dunham et al.,
“Identifying the Development and Application of Artificial Intelligence in
Scientific Text.”; The merged corpus includes more papers than the Map of
Science. Papers in the merged corpus that lack citations or references are not
assigned to a research cluster and therefore are not included in the papers
represented in the Map of Science.
22
The amount of times a given set of papers was cited by other papers.
23
Four clusters (#91424, #57343, #95593, and #14681) were omitted.
24
Papers can have multiple country affiliations if the authors indicate more than
one affiliation in a paper. Therefore, there are more country paper affiliations
than papers (in the clusters).
25
Table 2 shows the distributions of AI, re-identification, and speaker
recognition paper author affiliations of the top 5 countries (by paper affiliations).
Table 3 shows the citations of AI, re-identification, and speaker recognition
papers of the top 5 countries (by citations of the country-affiliated papers).
26
The most cited re-identification paper in our cluster is Liang Zheng et al.,
“Scalable Person Re-identification: A Benchmark,” in 2015 IEEE International
Conference on Computer Vision (ICCV), December 2015,
https://ieeexplore.ieee.org/document/7410490; Additionally, the most cited
speaker recognition paper in our cluster was produced by authors in French
academia (School of Engineers in Computer Intelligence), United States
academia (MIT), and Canadian academia (Universities of Quebec) and nonprofits
(Computer Research Institute of Montréal). The paper is Najim Dehak et al.,
“Front-End Factor Analysis for Speaker Verification,” IEEE/ACM Transactions on

Audio, Speech, and Language Processing 19, no. 4 (May 2011),
https://ieeexplore.ieee.org/document/5545402.
27
Percentages in Figures 6 and 7 are out of the total re-identification paper
affiliations and paper citations of all countries in cluster #1419.
28
China-affiliated re-identification paper citations may be moderately
undercounted because papers from the CNKI dataset, which is a Chinese
publication repository that CSET merged into the Map of Science, can lack
citation data. There were at least 88 CNKI-sourced papers in cluster #1419 that
had null citation values, although many of them were recently published. For
the purpose of this report, we assess that the quantity of missing citations does
not skew the data to a problematic degree.
29
The sudden and sustained rise in China-affiliated re-identification publication
is tied to many variables, but we assess that it largely stems from the
proliferation of deep learning architectures, datasets, and compute resources,
combined with political, security, and commercial incentives associated with
China’s rapid modernization and urbanization.
30
The disparity in re-identification paper affiliations between China and other
countries widened rapidly after 2016, with the ratio of U.S.-to-China affiliated
papers being over 1:3 in 2017, 1:4 in 2018, and 1:5 in 2019. The affiliated paper
citation disparity was 1:2 in 2017 and 1:3 in 2018, which is less stark but still
significant. However, Chinese-affiliated papers sometimes lack data on citations
because they are sourced from the CNKI dataset (which can lack data on
citations). Therefore, the lower citations disparity is somewhat skewed, but it is
not highly problematic.
31
Percentages in Figures 8 and 9 are out of the total speaker recognition paper
affiliations and paper citations of all countries in cluster #6855. Note that the
top countries affiliated with papers are not always the same as the top countries
affiliated with papers that are more frequently cited.
32
The figure on the number of organizations is an approximation based on MAG
affiliation IDs. Although reaching an exact figure is possible, it can require
extensive manual data cleaning in the current version of the Map of Science.
33
Approximately 20 percent of re-identification publications in cluster #1419
and 25 percent of speaker recognition publications in cluster #6855 are not
linked to “organization type” metadata. A fully developed and operational Map
of Science will likely address this issue, but it must be recognized that there are
currently blind spots within the data. Areas that lack such visibility are readily
identifiable and can often be addressed manually.

34
Much of the data displayed in Table 4 was manually organized due to current
limitations of the CSET Map of Science metadata (e.g., there were often
overlapping names for the same entities).
35
A comprehensive organizational analysis is outside the scope of this report, as
we primarily intend to demonstrate a process for obtaining the information that
enables subsequent policy inquiries. However, the metadata opens multiple
vectors for further analysis and insight. This includes identifying key clusters of
research among organizations (along with their geographic distributions),
fostering collaborative research initiatives with universities and research centers
for specific machine learning cross-sector collaboration (e.g., FFRDCs) and
tracking organizational affiliations with foreign regimes.
36
The public Map of Science will provide data on individual author countries (in
cases where such information is available).
37
We must note that two of the papers used to acquire more recent metrics
were published after March 2021 (the month we extracted metadata from the
clusters), therefore they are not included in our version of the clusters but would
be if we repeated the methodology using updated clusters from the most recent
version of the Map of Science.
38
For more information on the MSMT17 dataset, see Longhui Wei et al., “Person
Transfer GAN to Bridge Domain Gap for Person Re-Identification,” arXiv preprint
39
For example, the use of large-scale “pre-training” on much broader vision
recognition datasets.
40
See Appendix G for details on performance metrics data and sources.
41
For more information on the VoxCeleb dataset, see Arsha Nagrani, Joon Son
Chung, and Andrew Zisserman, “VoxCeleb: A Large-Scale Speaker Identification
Dataset,” arXiv preprint arXiv:1706.08612 (2018),
42
For a more detailed assessment of re-identification and speaker recognition
performance metrics, see Appendix F. For more details on metrics data and
sources, see Appendix G.
43
Figure 12 is sourced from Daniel Zhang et al., “Artificial Intelligence Index
Report 2021” (Stanford University Human-Centered Artificial Intelligence (HAI),
March 2021), https://hai.stanford.edu/research/ai-index-2021.
44
“Leaderboard Version: 2.0,” Gluebenchmark,
https://super.gluebenchmark.com/leaderboard/; “Leaderboard,” Gluebenchmark,
https://gluebenchmark.com/leaderboard; “The GLUE benchmark, introduced a

little over one year ago, offers a single-number metric that summarizes progress
on a diverse set of such tasks, but performance on the benchmark has recently
surpassed the level of non-expert humans, suggesting limited headroom for
further research;” SuperGLUE “is a new benchmark styled after GLUE with a
new set of more difficult language understanding tasks, a software toolkit, and
a public leaderboard … As with GLUE, we seek to give a sense of aggregate
system performance over all tasks by averaging scores of all tasks. Lacking a
fair criterion with which to weigh the contributions of each task to the overall
score, we opt for the simple approach of weighing each task equally, and for
tasks with multiple metrics, first averaging those metrics to get a task score,”
Alex Wang et al., “SuperGLUE: A Stickier Benchmark for General-Purpose
Language Understanding Systems,” arXiv preprint arXiv:1905.00537v1 (2019),
45
GLUE and SuperGLUE consist of multiple natural language processing-
related tasks that models are tested against, the results of which are
aggregated into a single score. They were developed because of the emergence
of a new class of highly capable natural language processing models that
leveraged techniques that had previously improved computer vision
(specifically, large-scale pre-training and the use of Transformer architecture
models).
46
By having multiple types of users, it will become easier to identify and triage
areas for further investment and investigation. For instance, significant demand
for greater country-level granularity regarding research paper affiliation could
spur governmental efforts to develop the capability. Having such demand
signals provides a cheap way to identify areas of shared need.
47
Papers with Code is an example of such a resource.
48
For example, one could identify a cluster or clusters related to a topic, extract
all relevant metrics from the cluster publications, ingest them into the
repository, then use both elements (the cluster bibliometrics and the repository
metrics) to analyze the topic more holistically.
49
“The cluster-level view in the Map of Science is the most granular
aggregation of scientific research publications that provides insight into specific
areas of research without analyzing publications individually,” Autumn Toney
and Melissa Flagg, “Comparing the United States’ and China’s Leading Roles in
the Landscape of Science” (Center for Security and Emerging Technology,
2021), https://cset.georgetown.edu/publication/comparing-the-united-states-
and-chinas-leading-roles-in-the-landscape-of-science/. They are also assigned
broad subject areas and specific research fields using MAG.
50
“A typical speaker diarization system usually contains multiple steps. First, the
non-speech parts are filtered out by voice activity detection. Second, the speech

parts are split into small homogeneous segments either uniformly or according
to the detected speaker change points. Third, each segment is mapped into a
fixed dimensional embedding. Finally, clustering methods or end-to-end
approaches are applied to generate the diarization results,” Jixuan Huang et al.,
“Speaker Diarization with Session-Level Speaker Embedding Refinement Using
Graph Neural Networks,” arXiv preprint arXiv:2005.11371 (2020),
51
Regarding the mediums through which re-identification papers are published
(i.e., document type metadata) and distributed (i.e., database metadata): Most
re-identification papers in our cluster came from conferences between 2010–
2018, but journal publications have increased significantly since 2019. Re-
identification publications’ growth likely began to stagnate in 2020 because
COVID-19 disrupted publishing by forcing changes in the dates of conferences,
through which approximately 41 percent of the re-identification papers in our
cluster have been released. Additionally, this metadata was extracted from the
Map of Science in March 2021, so the datasets may not be fully updated to
reflect all the papers published (and cited) in 2020.
52
Approximately 19 percent of the papers in cluster #6855 have null citations
metadata. Although some data is missing, many null values are likely because
approximately 12 percent of the papers were published since the year 2019 and
have had limited time to accrue citations that would be reflected in the Map of
Science.
53
For example, systems that do well on MSMT17 tend to do large-scale "pre-
training" on non-re-identification datasets, mirroring the performance trends in
computer vision more broadly.
54
Dan Hendrycks et al., “Natural Adversarial Examples,” arXiv preprint
55
Chuting He et al., “TransReID: Transformer-Based Object Re-Identification,”
arXiv preprint arXiv:2102.04378 (2021), https://arxiv.org/abs/2102.04378;
Abdallah Benzine et al., “Deep Miner: A Deep and Multi-Branch Network which
Mines Rich and Diverse Features for Person Re-identification,” arXiv preprint
arXiv:2102.09321 (2021), https://arxiv.org/abs/2102.09321; Changxing Ding et
al., “Multi-Task Learning with Coarse Priors for Robust Part-Aware Person Re-
Identification,” arXiv preprint arXiv:2003.08069 (2021),
https://arxiv.org/abs/2003.08069; Tianlong Chen et al., “ABD-Net: Attentive but
Diverse Person Re-Identification,” arXiv preprint arXiv:1908.01114v3 (2019),
https://arxiv.org/abs/1908.01114v3; Wei et al., “Person Transfer GAN to Bridge
Domain Gap for Person Re-Identification,” arXiv preprint arXiv:1711.08565

56
Nagrani et al., “VoxCeleb: A Large-Scale Speaker Identification Dataset”; Joon
Son Chung, Arsha Nagrani, and Andrew Zisserman., “VoxCeleb2: Deep Speaker
Recognition,” arXiv preprint arXiv:1806.05622 (2018),
https://arxiv.org/abs/1806.05622; Weidi Xie et al., “Utterance-Level
Aggregation for Speaker Recognition in the Wild,” arXiv preprint
arXiv:1902.10107 (2019), https://arxiv.org/abs/1902.10107; Arsha Nagrani et
al., “Voxceleb: Large-Scale Speaker Verification in the Wild,” Computer Speech
& Language 60 (March 2010),
https://www.sciencedirect.com/science/article/pii/S0885230819302712; Jenthe
Thienpondt, Brecht Desplanques, and Kris Demuynck, “The IDLAB VoxSRC-20
Submission: Large Margin Fine-Tuning and Quality-Aware Score Calibration in
DNN Based Speaker Verification,” arXiv preprint arXiv:2010.11255v1 (2020),
https://arxiv.org/abs/2010.11255v1; Miao Zhao et al., “The SpeakIn System for
VoxCeleb Speaker Recognition Challenge 2021,” arXiv preprint

Measuring AI Development

Uploaded by

Copyright:

Available Formats

Measuring AI Development

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Measuring AI Development

Uploaded by

Copyright:

Available Formats

December 2021

Introduction: Upgrading AI Measurement for the 21st Century ....... 2

Center for Security and Emerging Technology | 1

This report demonstrates how contemporary approaches to

By combining a versatile and frequently updated bibliometrics tool

Center for Security and Emerging Technology | 2

Measurement is uniquely intertwined with policymaking, especially

After outlining our methodology, this report discusses how

Center for Security and Emerging Technology | 3

Bibliometric Clustering: The technical basis of our methodology is

Bibliometric Analysis: To begin our analysis, we first selected

For the purposes of this report, we explored whether we could

Center for Security and Emerging Technology | 4

Performance Metrics Assessment: In machine learning research,

Center for Security and Emerging Technology | 5

Center for Security and Emerging Technology | 6

It should be noted that this approach is not perfect. Clustering

Center for Security and Emerging Technology | 7

In this section, we apply the methodology to gain insights into the

The results described below provide meaningful visibility into the

Re-identification involves integrating computer vision techniques

Center for Security and Emerging Technology | 8

Source: Mang Ye et al. 10

The development of re-identification systems has important

Speaker recognition is an area of research that focuses on using

Speaker recognition models can be integrated into a range of

Center for Security and Emerging Technology | 9

Image synthesis is a subfield of machine learning research that

Image synthesis presents a host of issues with policy implications.

Center for Security and Emerging Technology | 10

Center for Security and Emerging Technology | 11

In what follows, we focus first on re-identification and speaker

(1) We identified two respective research clusters for re-

(2) Publications related to re-identification and speaker

(3) Growth of re-identification and speaker recognition

(4) Most organizations affiliated with the authors of papers

Center for Security and Emerging Technology | 12

(5) Re-identification publications (5) Speaker recognition

(8) The performance of re- (8) The performance of speaker

Center for Security and Emerging Technology | 13

From our bibliometric analysis we identified and extracted

Here we begin with an analysis of publication and citation growth

Center for Security and Emerging Technology | 14

Table 1. Papers, Citations, and Growth Rates

Total Papers (2010-2020) 3,282 1,938 2,015,541

Total Paper Citations (2010-2020)22 96,462 29,096 24,896,447

Average Annual Paper Growth Rate

Average Annual Paper Citation

Research into both re-identification and speaker recognition has

Figure 4 shows a divergence in growth between the topics around

Center for Security and Emerging Technology | 15

Figure 4. Machine Learning Topic Papers Published per Year,

Speaker Recognition (Cluster #6855)

Note: For a more detailed analysis of re-identification and speaker recognition

Figure 5. AI Papers Published per Year, 2010–2020

200 166 168

Source: CSET Map of Science, as of March 2021.

Center for Security and Emerging Technology | 16

Here we briefly address three related limitations that stem from