Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming...

65
UNCLASSIFIED LLNL-MI-750860 UNCLASSIFIED Lawrence Livermore National Laboratory P.O. Box 808 Livermore, CA 94550 Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs May 2018

Transcript of Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming...

Page 1: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

UNCLASSIFIED

Lawrence Livermore National Laboratory • P.O. Box 808 • Livermore, CA 94550

Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

May 2018

Page 2: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 2

UNCLASSIFIED

Disclaimer

The views expressed are those of the author and may not reflect those of Lawrence Livermore National

Laboratory, the Department of Energy, the National Nuclear Security Administration, or any other U.S.

government entity. This document was prepared as an account of work sponsored by an agency of the

United States government. Neither the United States government nor Lawrence Livermore National

Security, LLC, nor any of their employees makes any warranty, expressed or implied, or assumes any legal

liability or responsibility for the accuracy, completeness, or usefulness of any information, apparatus,

product, or process disclosed, or represents that its use would not infringe privately owned rights. Reference

herein to any specific commercial product, process, or service by trade name, trademark, manufacturer, or

otherwise does not necessarily constitute or imply its endorsement, recommendation, or favoring by the

United States government or Lawrence Livermore National Security, LLC. The views and opinions of

authors expressed herein do not necessarily state or reflect those of the United States government or

Lawrence Livermore National Security, LLC, and shall not be used for advertising or product endorsement

purposes.

This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore

National Laboratory under contract DE-AC52-07NA27344. Lawrence Livermore National Security, LLC.

Z-18-194

Page 3: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 3

UNCLASSIFIED

Coming of Age

Artificial Intelligence and the Continuing

Revolution in Military Affairs

Contents

Executive Summary .................................................................................................................................... 5

Introduction ................................................................................................................................................. 6

AI Technology Trend Lines ........................................................................................................................ 9

Commercial investments will shape AI market “winners,” but AI markets are not yet profitable ...... 9

A new generation of narrow AI technologies are beginning to outperform humans on varied tasks 10

Beyond deep learning: toward a potential hybrid Bayesian approach ............................................... 17

Potential Military Applications of AI ....................................................................................................... 18

Emerging narrow AI technologies are well-suited to varied DOD applications ................................ 18

Logistics: AI useful across a broad range of tasks ............................................................................. 19

Sensor-to-shooter situational awareness: AI is a force-multiplier for “big data” exploitation ........... 20

Unmanned vehicles: more varied and complex missions likely as key technologies mature ............ 22

Cyber and electronic warfare: AI critical to maintaining the edge ..................................................... 24

Wargaming: potentially better longer-term prospects ........................................................................ 25

Autonomous weapons: intelligent systems are available but policy-constrained............................... 27

Command and control: “AlphaWar” probably decades away ............................................................ 28

Outlook and Considerations ...................................................................................................................... 30

AI forecast: foggy with a chance of technology breakthroughs ......................................................... 31

Calibrating expectations ..................................................................................................................... 39

Appendix A: Challenges Associated with Credibly Estimating Commercial Investment in AI-Related

Technologies and Applications ................................................................................................................. 41

Appendix B: Deep Learning and Deep Reinforcement Learning Fuel the Latest AI Breakthroughs ....... 43

Appendix C: Candidate Near-Term Implementation Measures to Strengthen DOD’s AI Posture ........... 48

Page 4: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 4

UNCLASSIFIED

This page is intentionally left blank.

Page 5: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 5

UNCLASSIFIED

Coming of Age:

Artificial Intelligence and the Continuing Revolution in

Military Affairs

Jason D. Ellis, Barry Y. Chen, Mary D. Gullett, and Benjamin W. Bahney

Lawrence Livermore National Laboratory

Executive Summary

Advancements in artificial intelligence (AI) are enabling machines to outperform humans in an

increasing variety of narrow tasks, leading national security planners to consider both the breadth

of potential military uses of AI and their broader transformative implications for the nature of

future warfare. As originally conceptualized, the revolution in military affairs (RMA)—empowered

by advances in microelectronics over the past five decades—emphasized the prospective role of

autonomous systems for battlefield advantage. In our view, AI technologies will soon be able to

help militaries improve precision strike through real-time target acquisition; in time will be able

to aid in information dominance through use in automated electronic and cyber warfare; and may

eventually be able to enhance command and control and strategic planning.

While an artificial general intelligence remains far off, so-called “narrow” AI technologies have

matured sufficiently for varied defense applications. Machines for the first time are exceeding

human expert performance in areas such as object detection in imagery, speech recognition, and

modestly complex video and board games. There are two AI research subfields responsible for

these recent breakthroughs:

• Deep learning. This maturing set of technologies is well suited to automating and

accelerating the first half of the “observe, orient, decide, act” (OODA) loop, a set of

capability enhancements that, when combined with distributed sensors, could facilitate real-

time reconnaissance strike acquisition and targeting.

• Deep reinforcement learning. While less mature, these technologies train machines to

optimize action sequences by providing reward signals for successful outcomes and could—

once more fully developed—have substantial implications for autonomous platforms, rapid

decisionmaking, and new methods of operation, potentially revolutionizing the second part

of the OODA loop over the longer term.

These cutting-edge technologies have some important limitations that could hinder their broad

implementation in the Department of Defense (DOD). AI algorithms typically require substantial

labeled training data, may lack useful prediction uncertainty characterizations, may be vulnerable

to spoofing or poisoned data streams, and often give operators limited insight into their logic—a

key shortcoming where DOD leaders may require greater transparency into the logic structures

and associated decision processes.

• Defenses against spoofing are still in early-stage development, but at least two—adversarial

training and defensive distillation—are potentially effective.

• The opacity of today’s machine learning methods could lead AI systems to perform

unanticipated actions, but Bayesian models could lead to a more “explainable AI.”

Page 6: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 6

UNCLASSIFIED

As narrow AI technologies continue to mature, they will likely become relevant to additional

military applications. Their potential utility, technical feasibility, and potential adoptability by

DOD varies over the short, medium, and long terms:

• For applications such as logistics, investment in relevant AI technologies and associated

autonomous systems could deliver new capabilities in the near to medium term but reflect

diminishing returns over the long term as supply chain efficiencies are realized.

• In other areas, such as automated command and control, AI-associated investments could

have substantial payoff over the long term but probably would provide minimal near-term

advantage because even the most promising deep reinforcement learning techniques

significantly underperform humans in real-time strategy activities.

Indeed, cutting-edge AI technologies are not yet achieving superhuman performance in strategy

games that simulate real-world activities, so there is reason to be circumspect about their long-

term potential. So-called artificial “general” intelligence—that is, an AI capable of performing

any task a human might undertake—is likely to remain an aspirational goal for many years, if not

decades. In this context, DOD can both capitalize on commercial successes and, in parallel,

develop increasingly capable AI military systems. This might include DOD adopting a five-part

approach over the next decade to acquire, develop, and use AI technologies, such as the following:

1. Acquire or adapt existing commercial technologies for routine administrative tasks, such as

human resources, travel, and other “back office” support.

2. Acquire or adapt commercial products or services to improve select combat support and

combat service support functions, such as logistics and supply chain management.

3. Conduct limited development of existing commercial AI tools for specific defense missions,

such as providing enhanced support to intelligence, surveillance, and reconnaissance (ISR).

4. Develop and deploy new or improved AI products that enhance fielded and developmental

combat systems, such as electronic or cyber warfare.

5. Design, develop, prototype, and experiment with a combination of AI technologies, military

platforms, and new operating concepts to develop new or enhanced combat capability.

Introduction

Revolutions in military affairs (RMAs) are characterized by a combination of significant technological

change, an evolution in military systems, operational innovation, and organizational adaptation.1 The

ongoing RMA, driven by continued advances in microelectronics, is changing how militaries organize and

fight around the concept of reconnaissance strike complexes that can find, fix, and finish targets using

sensors and stand-off weapons. The defense strategists who first conceived of this RMA in the early 1990s

conceived a broad role for “automation,” which some artificial intelligence (AI) technologies are now

beginning to realize. Namely, they foresaw that militaries would pursue the ability to find deep targets in

real time, with stand-off weapons that potentially guide themselves, increasingly moving toward striking at

the speed of light with cyber or directed energy, and using organizational concepts and strategies developed

by complex computing systems.2

Page 7: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 7

UNCLASSIFIED

Hype around AI abounds, for both its prospective commercial and national security applications. But to

date, the Department of Defense (DOD) has only a handful of successes in leveraging these commercially

driven technologies to aid the warfighter. Indeed, the extents to which commercial firms continue to

improve AI technologies and to which the DOD can fully capitalize on these developments are timely and

important questions. This paper evaluates the emerging prospects for AI-related military applications over

the near term, namely within five years; the mid term, which we define as five to ten years; and potentially

over longer-term windows of opportunity. We first consider the underlying technology trends in AI, then

explore potential use cases for DOD, and finally describe candidate measures to strengthen DOD’s ability

to successfully exploit the emergent revolution in AI technologies.

Recent advancements in AI have already enabled machines to outperform humans in specific tasks, and

continued progress is likely. Looking ahead, this raises questions about the extent to which the application of

AI technologies in the military domain could fundamentally transform the nature of warfare. The ability to

wage war at machine speeds could conceivably reduce decision time while a corresponding growth in the

availability of “big data,” underwritten by continuing advances in high-performance computing, could

enhance the quality of military decisionmaking. At the same time, parallel advances in technologies such as

increasingly capable autonomous systems both draw on and will further advance the state of the art in machine

learning, computer vision, and other AI technologies.* Should present trends continue, AI-enabled systems

will have implications for each element of the “observe, orient, decide, act” (OODA) decision loop.3

Autonomous systems and AI technologies are neither new nor unprecedented. Over the past half-century,

many nation states have designed, developed, and fielded military systems capable of autonomous or semi-

autonomous operations. Today, more than 30 countries field such hardware for tasks such as identifying,

tracking, prioritizing, and cueing targets; maneuvering and homing in on targets; and weapon detonation

timing.4,5 DOD has long embraced the selective application of autonomous technologies and seeks to

capitalize on promising AI developments. Among other things, in April 2017 it established an “algorithmic

warfare” task force designed “to turn the enormous volume of data available to DOD into actionable

intelligence and insights at speed.”6 Its first task: “to put an algorithm into a combat zone” by the end of 2017

in support of DOD’s counter-Islamic State campaign.7 More broadly, the department’s third offset strategy

centers on countering foreign anti-access area denial advancements by fielding autonomous deep-learning

systems, fostering human-machine collaboration, machine-assisted human operations, advanced human-

machine combat teaming, and network-enabled semi-autonomous weapons.8 Ultimately, the success of these

initiatives will be tied to DOD’s ability to effectively exploit developments in AI and related technologies.

In contrast to the first and second offsets pursued by DOD, industry is leading the charge in AI research

and development; the private sector is shaping the technology art-of-the-possible through substantial and

growing corporate investments. As a result, AI technical areas such as deep neural networks, machine

learning, natural language processing, and computer vision are markedly improving. International Data

Corporation (IDC) forecasts a rise from about $8 billion in 2016 global revenue from AI technologies to

more than $46 billion in 2020, a 54.4 percent compound annual growth rate.9,† Forrester Research

anticipates a greater than 300 percent increase in commercial investment this year in AI over 2016 levels

across all market sectors.10 Bank of America (BofA) Merrill Lynch Global Research sees an “Industry 4.0”

* AI is a broad scientific and computational research area, with subfields such as deep learning and deep reinforcement learning. As used in this document, AI technologies include enabling capabilities (such as machine learning, computer vision, or natural language processing) while AI-related technologies include applied capabilities (such as robotics).

† Market estimates vary considerably, and not all are as optimistic as the IDC estimate. But even the lower-end estimate provided by Tractica, which baselines AI worldwide 2016 revenue at $1.4 billion, projects this to rise to more than $10 billion by 2020 and to reach $59.8 billion by 2025. [Reference: Press release | Tractica website | Artificial Intelligence Software Revenue to Reach $59.8 Billion Worldwide by 2025 | 2 May 2017 | https://www.tractica.com/newsroom/press-releases/artificial-intelligence-software-revenue-to-reach-59-8-billion-worldwide-by-2025/].

Page 8: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 8

UNCLASSIFIED

revolution underway with a move toward intelligent and perceptive robotic systems—fueled by

improvements in AI, computation, and other technologies—with a potential multitrillion dollar market

impact through the 2020s.11,12 AI and related technologies are widely viewed as transformational in nature.13

Artificial Intelligence “101”

Artificial intelligence is “a science and a set of computational technologies that are inspired by—but

typically operate quite differently from—the ways people use nervous systems to sense, learn, reason,

and take action.”S1

• Weak or narrow AI—which seeks to solve limited or discrete tasks a human might perform—

reflects the state of the art.

• Strong or general AI—a much more expansive form of AI capable of performing virtually

any task a human might undertake—is more in the realm of fiction than science today and is

likely to remain an aspirational goal for two or more decades.

Machine learning “provides the foundational mathematical and statistical algorithms that are used in

AI’s application areas.”S2

Deep learning “describes a class of machine learning approaches centered on the construction of

artificial neural networks.”S3

Reinforcement learning is a “technique whereby learning proceeds by adaptively constructing a

sequence of actions that collectively maximize some long-term reward.”S4

Autonomy “results from delegation of a decision to an authorized entity to take action within specific

boundaries.” An autonomous system would “have the capability to independently compose and select

among different courses of action to accomplish goals based on its knowledge and understanding of

the world, itself, and the situation.” AI is foundational to autonomy.S5

References

S1. One Hundred Year Study on Artificial Intelligence: Report of the 2015-2016 Study Panel, Stanford University,

Stanford, CA | Artificial Intelligence and Life in 2030 | September 2016 | p. 4.

S2. JSR-16-Task-003 | JASON committee | Perspectives on Research in Artificial Intelligence and Artificial General

Intelligence Relevant to DOD | January 2017 | p. 11.

S3. Techsight Snapshot Report | Office of the Assistant Secretary of Defense for Research and Engineering | Deep Learning |

October 2017 | p. 2.

S4. A. Ilachinski, Center for Naval Analyses | AI, Robots, and Swarms: Issues, Questions, and Recommended Studies |

January 2017 | p. 49.

S5. Defense Science Board | Summer Study on Autonomy | June 2016 | pp. 3–4.

U.S. efforts to harness the military potential of AI are not unique. Both Russia and China have publicly

underscored their interest in developing AI military systems. Among other things, this includes Russian

development of missiles, drones, and aircraft able to select targets and outmaneuver defensive systems and

a turret-mounted weapon featuring a “fully automated combat module.”14,15,16 China views AI as integral

to its ability to compete successfully in “intelligentized” warfare.17 In this context, the People’s Liberation

Army (PLA) seeks to tap AI for missions ranging from integrated reconnaissance and strike, to cyber and

electronic warfare operations, to development and deployment of networked autonomous systems. Taking

Page 9: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 9

UNCLASSIFIED

the long view, Yang Wei, deputy director of the scientific and technological commission at the Aviation

Industry Corporation of China, argues that China is well positioned to “overtake others [in AI], because we

are all at the same starting line.”18 Whether this ambitious Chinese vision will ultimately materialize is

unclear, but its prospects rise with China’s State Council establishing a goal for the state to develop a $150

billion domestic AI industry and to become a recognized global AI innovation center by 2030.19,20,21

Taken together, anticipated AI-related technology developments and known foreign interest in more capable

AI military systems present an opportunity but also pose a significant challenge to U.S. national security. Unlike

the second offset strategy, which conferred a multidecade warfighting advantage to U.S. forces, DOD’s ability

to achieve a definitive and sustainable warfighting advantage through AI is in question. Rather, as with areas

such as cybersecurity or submarine warfare, the likelihood of AI-related foreign military innovation

underscores the prospective development of a multifaceted offense/defense dynamic. While a first-mover

advantage in this context will be transitory, it may prove a critical enabler for military operations.

AI Technology Trend Lines

The field of AI (see sidebar) is focused on the study of “agents [e.g., machines and computers] that receive

percepts from the environment and perform actions.”22 Cognitive abilities are required for agents to make

sense of percepts to ultimately take actions that increase their rewards. These cognitive abilities (such as

pattern recognition, reasoning, planning, or decisionmaking) are typically associated with humans. “Weak”

or “narrow” AI refers to the nonsentient AI focused on solving limited tasks. “Strong” or “general” AI is

the type that is most frequently depicted in science fiction, where robots can reason, plan, and act as if they

were human. While general AI would represent the pinnacle of AI research, its arrival is uncertain at best,

and it will only come potentially two or more decades from now. What is clear is that narrow AI has arrived.

Over the past several years, AI systems have achieved tremendous breakthroughs on tasks once thought to

be nearly impossible for computers or only realizable in the distant future.

Commercial investments will shape AI market “winners,” but AI markets are not yet profitable

DOD is investing significantly in artificial intelligence, big data, and cloud technologies. Data science and

analytics firm Govini estimates aggregate DOD outlays in 2017 of more than $2.4 billion across a range of

AI systems, learning and intelligent technologies, and advanced computational enablers.23,* As sizable as

these investments are, however, they comprise less than one-tenth of corporate AI investments. McKinsey

Global Institute estimates 2016 industry investments at $26 billion to $39 billion, more than three-quarters

of which represent internal investments made by technology giants such as Google, Facebook, and Baidu

(appendix A).24 At the same time, global revenue of just $8 billion in commercial AI underscores that the

rate of adoption for AI technologies at scale have thus far been low. Not surprisingly, adoption tends to be

highest in those market sectors already characterized by strong digital adoption, such as in the high-

technology, telecommunications, automotive, and financial services industries, and lowest in sectors such

as education, tourism, and health care.25 Certainly, widespread estimates of a sharp near-term rise in global

market growth suggest that certain AI technologies are maturing rapidly and will find market support across

multiple industries.

* Estimates of Federal AI spending vary considerably. For example, Govini’s $2.4 billion 2017 estimate is significantly higher than the $1.1 billion estimate provided by the National Science and Technology Council (NSTC) in 2015, a difference more likely grounded in accounting variations—that is, what specifically should be counted as “AI” spending—than in short-term programmatic growth. [Reference: Report | NSTC | Preparing for the Future of Artificial Intelligence | October 2016 | https://obamawhitehouse.archives.gov/sites/default/files/whitehouse_files/microsites/ostp/NSTC /preparing_for_the_future_of_ai.pdf | p. 25.]

Page 10: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 10

UNCLASSIFIED

While internal corporate investments can be difficult to track, McKinsey estimates substantial allocations

in 2016 to machine learning ($5–7 billion) and computer vision ($2.5–3.5 billion), and much lower levels

in areas such as natural language ($600–900 million), autonomous vehicles ($300–500 million), smart

robotics ($300–500 million), or virtual agents ($100–200 million).26 Among other things, both AI-related

startup companies and industry mergers and acquisitions are changing the corporate landscape.27,28 In turn,

Forrester Research anticipates moderate or significant success within the next five years in AI technology

areas, including deep learning platforms, natural language generation, biometrics, speech recognition, text

analytics and natural language processing, and image and video analysis.29 Within five to ten years, they

project comparable success in areas such as semantic technology, virtual agents, machine learning

platforms, robotic process automation, AI-optimized hardware, and decision management.30

Ultimately, DOD’s ability to achieve a future warfighting posture that incorporates AI will be grounded in

large part by commercial innovation. To the extent that innovation in AI and related technologies progresses

as anticipated, DOD may be able to leverage corporate innovation in “winning” areas. For example, the

$7.5 billion investment forecast for global military robotics spending in 2018 is just 15 percent of the

aggregate $50.5 billion in total robotics spending worldwide.31 In other areas, where a compelling business

need is lacking, a commercial market is less likely to develop, or expected return on investment is low—

for example, swarm intelligence or video analytics—DOD could focus its research and development

resources to groom desired technologies. In this context, the Joint Staff’s director of intelligence for

warfighter support, Lt. Gen. Jack Shanahan, sees that “everything that industry is working on has some

applicability throughout the entire department,” in applied mission areas ranging from intelligence

collection, sensor fusion, and targeting support for operations through back-office functions such as

logistics flow.32 For applications such as data triage, “the state of the art is good enough for the

government,” observes the head of DOD’s Algorithmic Warfare Task Force, Colonel Drew Cukor.33

Underscoring this good enough mindset, Will Roper, then-director of DOD’s Strategic Capabilities Office,

suggested in 2017 that DOD avoid “letting perfect, exquisite, government-only solutions be the pacing

function for the military.”34

A new generation of narrow AI technologies are beginning to outperform humans on varied tasks

As a point of departure, narrow AI technologies have matured sufficiently that they could help DOD achieve

select enterprise and mission objectives. Among the possible applications: automating, and in the process

significantly accelerating, the OODA loop. Based on the current state of the art, AI techniques are especially

well-suited for making sense of the world (“orient”) and evaluating the best courses of action (“decide”).

Organizations able to collect and make sense of large datasets to facilitate better decisions more rapidly

will have a competitive advantage in the information era. Those with faster OODA loop execution times

have, all things being equal, a competitive edge. AI could thus be a key enabler to achieving a higher

operational tempo than near peer adversaries.

Narrow AI technologies have begun to demonstrate their potential. Over the past several years, some have

achieved—whether individually or in combination—what are typically referred to as “superhuman” results.

Computers have for the first time exceeded human expert performance in several challenging applications,

including object detection in imagery,35 speech recognition,36 and moderately difficult video and board games.37

Google DeepMind’s AlphaGo AI program recently beat the world’s number one ranked Go board game player

three out of three times in 2017—a feat subject matter experts previously considered would not happen for

perhaps another decade.38,39 While AI encompasses many different research subfields (table 1 provides a

representative set), arguably the most promising areas are those responsible for recent breakthroughs: “deep

learning” and “deep reinforcement learning.” (We discuss these more fully in appendix B.)

Page 11: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 11

UNCLASSIFIED

Table 1. Select AI research subfields and enabling technologies.

Select AI research subfields

Representative capabilities

Key limitations Prognosis

Semantic modeling and reasoning

• Natural language processing

• Robotics

Significant time and monetary investment often required to build the ontology; requires structured data.

Remains useful in domains with complex rules; rule-based approach provides transparency/auditability; useful where inputs and outputs can be highly specified (if/then rules) for regularized tasks (e.g., countermeasures for heat-seeking missiles).

First generation neural networks

• Computer vision Inaccurate and prone to overfitting; large networks were computationally prohibitive.

Replaced by newer generation neural networks.

Probabilistic graphical models (Bayesian modeling)

• Natural language processing

• Robotics & autonomy

• High-performance computing

• Modeling and simulation

Computationally intensive/expensive, especially for models involving many variables.

More flexible in some ways than artificial neural networks, but inferior for many types of pattern recognition. Graphical structure provides transparency/ auditability—for example, “conclusions” come with associated probabilities.

Deep learning

• Computer vision

• Speech and natural language processing

• Video and image analytics

• Machine intelligence

Reliant on big data and can be misled by visual noise imperceptible to a human; does not scale well to domains for which it is not trained; logic can be unclear.

Capable of surpassing human performance where there is sufficient training data, computational horsepower, and clearly defined objectives. Excels at pattern recognition; one of the most successful applications has been in image recognition.

Deep reinforcement learning

• Robotics

• Machine intelligence

• Autonomy

• Virtual agents

• High-performance computing

Reliant on big data and can be misled by visual noise imperceptible to a human; does not scale well to domains for which it is not trained; logic can be unclear.

Capable of surpassing human performance where there is sufficient training data, computational horsepower, and clearly defined objectives. Well-suited for planning and action, particularly applicable to robotics.

Quantum and Neuromorphic computing

• Computer vision

• Speech and natural language processing

• Robotics

No demonstration yet of a high-volume application where neuromorphic computing has outperformed the available alternative(s).

Potentially requires less computing power to process AI algorithms.

Ble

ed

ing

Ed

ge

C

utt

ing

Ed

ge

M

atu

re T

ech

no

log

ies

Page 12: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 12

UNCLASSIFIED

The success of deep learning is predicated on the availability and fidelity of labeled data

Algorithmic advances fuel superhuman performance specifically when neural networks, which are the

backbone of deep learning, are coupled with (1) the existence and usability of massive labeled training

sets—that is, large amounts of labeled data created by crowds of human labelers; and (2) the availability of

substantial computational horsepower. Among the various AI approaches, neural networks readily scale

both in model complexity and computational tractability. This means that the traditional bias/variance

tradeoff for overall error reduction is no longer as meaningful a tradeoff;40 the ability to train large neural

networks (to reduce bias) using massive amounts of training data (to reduce variance) results in networks

that simultaneously achieve low bias and low variance and, as such, very low error rates. While the

mathematical limit is not yet clear, deep learning can continue improving performance with larger models

and additional training data—up to some unknown limit—while other learning algorithms cannot (see

figure 1).

The proliferation of both inexpensive digital sensors (such as smartphones and digital cameras) and data

sharing platforms (such as Flickr, Facebook, and YouTube) has resulted in massive amounts of available

training data for AI. Immense quantities of picture, audio, and video data are collected, as well as data on

consumer product and other reviews or preferences. For instance, Tesla and other companies are collecting

billions of miles of actual driving data from video, ultrasonic, positioning, and accelerometer sensors.41

Figure 1. “Why deep learning?” (courtesy of

Andrew Ng). Unlike traditional learning

algorithms, deep learning effectively takes

advantage of increased quantities of training

data to continually improve predictive

performance. One reason for this continued

gain in performance is the feasibility of training

larger, more complicated deep learning models

on increasing volumes of data.

Deep learning’s achievements are predicated on the availability of massive amounts of labeled training

data—that is, the availability of large numbers of 𝑋, 𝑌 pairs where 𝑌 is the human-desired prediction when

given input datum 𝑋. In the ImageNet Large-Scale Visual Recognition Challenge (ILSVRC)—an annual

competition where research teams evaluate their algorithms for visual object recognition—the availability

of millions of human-annotated images for training has helped deep learning developers effectively train

massive neural networks to attain subhuman error rates.42 In fact, it is common for ILSVRC teams to

augment their training sets by rotating, flipping, and scaling their training images, resulting in potentially

millions of additional labeled images that can be used for training purposes. While such an approach helps

generate additional training data, it also points to potential limits in the ability of deep learning algorithms

to generalize from training features.

Page 13: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 13

UNCLASSIFIED

Deep learning requires significant computational horsepower

Computational power is integral to the success of AI. The dirty little secret in the development of AI

learning systems generally, and deep learning specifically, is the vast space of possible model

configurations controlled by tuning knobs, also known as hyperparameters. In deep learning, examples of

hyperparameters include model architecture parameters (such as the number of layers, number of neurons

per layer, or choice of nonlinear activation functions) and optimization parameters (such as the learning

rate). Discovering optimal hyperparameter settings is a time-consuming endeavor but is necessary to

achieve the best possible performance. This is an area where high-performance computing can accelerate

the training of neural networks. The faster one can train and evaluate neural network performance, the faster

one can optimize hyperparameter settings. It is an open question whether modern deep neural network

architectures are sufficient to fully realize the promise of deep learning. It is conceivable that more domain

knowledge or more complexity will need to be added to such architectures in the future.

The availability of graphics processing units (GPUs) and their ability to parallelize and quickly process

matrix-to-matrix multiplications and element-wise operations make the fast training of large neural

networks possible. For example, training AlexNet—a typical ILVSRC deep network built on 1.2 million

images—took a 16-core Xeon central processing unit (CPU) 43 days, but only three days using an Nvidia

TitanX GPU in 2015. Facebook recently demonstrated the ability to train on this same dataset in one hour

using 256 Nvidia Tesla P100 GPUs.43 Using high-performance computing clusters of GPU nodes with low-

latency, high-bandwidth interconnect, it may be possible to further accelerate the training of huge neural

networks on massive training data. This is the goal of the Livermore Big Artificial Neural Network

(LBANN) project at Lawrence Livermore National Laboratory (LLNL).44 The ability to quickly explore

various neural network configurations through accelerated training will make it increasingly likely that

narrow AI proliferates to nearly every application where labeled training data can be collected.

The democratization of deep learning serves as a technology accelerator

One important trend within in the deep learning community is what might be called the “democratization” of

deep learning: the widespread availability of training software frameworks, neural network models, and even

in some cases training data enables broad participation in development and training of tailored neural networks.

The availability of suitable computation resources (such as Amazon’s EC2 or independent GPU workstations)

together with myriad deep learning training frameworks (such as Caffe, Torch, Theano, Keras, TensorFlow, or

MXNet) has led new entrants to the field and has arguably begun to accelerate progress in AI’s further

development.45 To achieve superhuman performance, neural network architectures must be tailored to each

targeted application. Searching the space of promising network architectures is a time-consuming task, as there

are nearly infinite architecture combinations that could be tested. The democratization of deep learning means

that more technologists are going to be able to search the space of possible neural network models. As a result,

further breakthroughs in model architectures and an increasing number of tasks where deep learning exceeds

human performance are more likely in future developmental efforts in AI.*

Breakthroughs in unsupervised and few-shot learning

A key challenge for deep learning, and for AI more broadly, is the ability to learn from unlabeled or small

quantities of labeled examples. Humans do not need to be shown thousands of images of an airplane to

learn what an airplane looks like. In few-shot learning, researchers are developing new algorithms that try * Note that while further advances in narrow AI through deep learning are likely, this does not mean that deep learning can ultimately scale to general AI capabilities. [Reference: Website | Deep Learning: A Critical Appraisal | arXiv | https://arxiv.org/abs/1801.00631.]

Page 14: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 14

UNCLASSIFIED

to mimic this ability of humans to learn from small data sets. A related area of study is unsupervised

learning, where the machine is provided only the input measurements 𝑋 during training. In the deep learning

community,* the unsupervised training task typically focuses on learning good feature representations of

input data.46,47 Such feature representations can often be transferred to other tasks where there is little

labeled training data. Learning features in an unsupervised manner on massive amounts of data is a

promising approach to reducing the amount of labeled training data required in the transfer task. The

creation of more effective unsupervised and few-shot learning techniques has the potential to further reduce

the time required to develop deep learning systems on new applications by reducing the need to gather and

annotate vast amounts of training data.

The success of deep reinforcement learning is integral to making AI work for military decisionmaking

In reinforcement learning, the AI is tasked with discovering optimal actions to take in a situation based on its

understanding of the dynamics in play and is provided reward signals for successful actions. More formally,

the AI learns to pick the sequences of actions to maximize its expected sum of future rewards. This

reinforcement-learning setting fits well with the OODA loop model, where an agent observes the world, orients

itself and its relationship to other agents in the world, decides which action would lead to the “best” outcome,

and acts accordingly (see figure 2).48 Deep reinforcement learning refers to the use of deep neural networks to

implement the models used in reinforcement learning that infer the state of the world from input percepts or

infer the best action to take in this context. Using deep reinforcement learning to automate the orient and decide

steps, researchers can develop remarkably effective AI for game play—in effect, a complex simulated

environment that in some cases can serve as a surrogate for real-world applications.

Figure 2. The

“OODA Loop”

framework:

observe, orient,

decide, and act.

This version

developed by

Andrew Ilachinski,

Center for Naval

Analyses.

* The neural networks used in unsupervised feature learning do use supervised training pairs 𝑋 and 𝑌, but human annotation effort is not required to produce the labels for 𝑌. It is in this sense that neural network training is “unsupervised.”

Page 15: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 15

UNCLASSIFIED

The current leader in deep reinforcement learning is Google DeepMind. In 2013, the company created

Deep Q-Learning Network (DQN), a deep neural network that has proven able to play many Atari video

games at or above human-level performance.49 Overall, DQN outperformed human players in 29 of 49

different Atari games.50 The actions it selected were designed to maximize its overall score, learning

directly from the input video of the game itself as much as ten times per second. Despite the limited

number of actions possible (“move joystick up (or down),” “move joystick up (or down) and press fire,”

etc.), by playing select games millions of times and observing changes in scores across multiple

gameplay sessions, DQN learned which actions were higher in quality and, as such, had greater success

prospects.

Google DeepMind achieved another impressive milestone in May 2017 by developing a deep reinforcement

learning system called AlphaGo that beat the world’s top-ranked player in the board game Go.51,52 Given

the complexity of the game, the AI research community considered such an accelerated achievement to

be at least ten years in the future. AlphaGo’s success both underscores the rapidity of progress and

highlights the difficulties of credibly anticipating the pace of progress. In Go, players take turns placing

stones on a 19 × 19 grid with a goal of capturing as much territory as possible. Go is much more complex

than Chess: it features roughly 250 possible moves per round and typically 150 rounds per game, versus

35 and 80, respectively, in Chess. This means there are on the order of 250150 possible move sequences—

making an exhaustive enumeration of game tree options practically infeasible.

To achieve this feat, Google DeepMind created two different types of deep neural networks. The first

type, called policy networks, learn the probability that an action will eventually lead to a win. These

networks can be trained by watching the games of human players or by the computer playing itself

repeatedly. The second type, called value networks, predict the probability of a win given the current

state of the game. AlphaGo combines these two types of networks in an innovative way to pick the best

action at each state. It uses policy networks to suggest the best moves at any time, double-checks whether

the moves will result in high-value states as estimated by the value network, and then quickly plays out

several moves in advance to further assess the utility of the proposed move. While Go does not feature

the dynamic complexity of a real-world decision environment in which multiple actors pursue parallel

objectives, AlphaGo’s success asks the question of whether deep reinforcement learning could ultimately

“beat” adversaries in real-world environments of relevance to DOD.

In this context, it is important to consider games where AI still significantly underperforms humans—for

example, the real-time strategy game StarCraft. In StarCraft, one or more players must balance resource-

gathering goals with warfighting objectives, such as building an effective combination of units for

defensive, offensive, and reconnaissance operations. This is a case where imbalances in parallel

objectives can be exploited by competent adversaries. For example, without enough resource-gathering

units, the speed with which a player can build warfighting units is limited. If, however, units assigned to

defend factories and resources are given short shrift, a capable adversary can curtail a player’s longer-

term offensive warfighting potential. Even using the most promising deep reinforcement learning

techniques, AI has not yet beat intermediate human players in StarCraft.

At least three factors explain AI’s comparatively limited performance in this context. First, the number

of distinct world-states combined with the universe of possible actions at any particular state is

staggering. In comparison with Go’s 250150 possible action sequences, StarCraft is orders of magnitude

more complex at 1,000,000300 (that is, the branching factor raised to the number of actions taken in a

five-minute game). Second, unlike Go and Chess, StarCraft incorporates the “fog of war”: the state of

the world is only partly visible. Adversary movements may be obscured, and terrain visibility is limited

to positions near friendly units. Planning an optimal course of action depends on timely and reliable

knowledge of the state of the world; if this is unknown, the AI must consider a wider range of possible

Page 16: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 16

UNCLASSIFIED

actions to account for uncertainty. Finally, successful StarCraft players must simultaneously make good

tactical and strategic decisions. This means that an AI must be able to effectively balance between

potentially competing short- and long-term planning and execution objectives.

Long-term planning becomes exponentially more complex as the number of possible actions to consider

increases. Compared to Go, where the number of possible actions to consider at any time is only roughly

250, StarCraft allows for millions of actions, so long-term planning is much more difficult in StarCraft.

From many sample AI-versus-human matches,53 it is quickly evident how rigid and inflexible typical AI

unit-build strategies are. These poor build strategies stem from poor long-term planning by the AI and are

easily countered by human players. Tactically, however, AI has the advantage over human players because

the number of actions that an AI can execute per window of time is much greater. In a well-balanced battle,

the AI can micromanage its army and quickly position each unit to successively attack the enemy and retreat

before getting damaged by a counterattack. In this way, the AI can defeat human players in a skirmish but

often loses the war; the human can out-strategize the machine.

Deep learning is not foolproof—adversaries can spoof or poison data streams

As deep learning successes mount, some researchers are actively seeking vulnerabilities in trained neural

networks to either attack or defend from attack. One area of research involves the study of adversarial

examples, or specially modified input data that look like normal examples to humans but appear vastly

different to a neural network. Figure 3 shows an image of a panda, which the AI correctly identified at a

57.7 percent confidence level. When an adversarial signal was added to the dataset,54 the neural network

incorrectly estimated with 99.3 percent confidence that the image was a gibbon, even though it appeared

unaltered to humans.

Figure 3. The original image of a panda (left) is correctly classified by a neural network with 57.7

percent confidence. The adversarial example image (right) is created by adding small (epsilon) amounts

of an adversarial signal to every input pixel of the original image. Now the neural network estimates with

99.3 percent confidence that the image is a gibbon.

Page 17: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 17

UNCLASSIFIED

The research community is aware of this limitation in AI performance and is seeking to develop appropriate

countermeasures. Several methods for generating adversarial examples have been proposed, requiring

varying amounts of information about the targeted neural network—for example, its architecture, gradient

information, input/output behavior, or the full model.55,56 Defenses against such attacks are still in early-

stage development, but two have shown some effectiveness: adversarial training and defensive distillation.

In adversarial training, adversarial examples are generated and added to the training set to inoculate the

network. In turn, defensive distillation trains smaller networks that learn the input/output behavior of larger

networks to improve its generalization capabilities beyond their training dataset and, as such, augment their

resiliency to potentially spurious data.57 As research in generating and defending against adversarial

examples continues to evolve, it will be important for DOD to understand, develop, and apply effective

techniques to either defend or attack deep learning-based AI systems.

Beyond deep learning: toward a potential hybrid Bayesian approach

As actual battlefield conditions are chaotic and hard to predict, and as adversaries try to be unpredictable,

it is likely that there will be a dearth of training data in key application areas for military-use deep learning.

In these cases where relevant data is scarce—where there is insufficient labeled training data—it may be

possible to augment, supplement, or adapt contemporary deep learning approaches.* Deep reinforcement

learning is one possibility to address this problem, but there are others. For example, Bayesian inference

has long been used for data exploitation problems where accuracy and uncertainty quantification matter. In

short, Bayesian classification minimizes the information cost for describing input data.

Recent advances in deep learning have established a tight correspondence between deep neural network

architectures and statistical processes. Because a single-layer neural network of infinite width is equivalent

to a statistical process,58,59 the nonlinear magic of neural networks is, in some sense, equivalent to linear

regression.60,61,62 A recent result from Google Brain showed that a neural network statistically corresponds

to a sequence of numbers that can be trained with linear algebra.63

The statistical analog to a deep neural net is a significant advance because supervised learning outputs (i.e.,

classifications or predictions) are naturally described with probabilities, and a statistical process enables

globally optimal solutions via standard linear algebra operations. A statistical representation of neural

networks also naturally provides uncertainty estimates, which appear to be meaningful in terms of

prediction errors.64 The uncertainties provided by such predictions have also been shown to help mitigate

adversarial attacks.65 The statistical representation of neural nets thus gives a Bayesian and enhanced

method for supervised learning.

In principle, the demonstrated performance of deep learning methods in some previously intractable data

exploitation tasks could provide a pathway to hybrid options that blend the better attributes of each field.66

This potential path is founded on the expectation that the whole is greater than the sum of its parts: a new

Bayesian deep learning technique could transform data-training, spoofing resilience, and interpretability of

analyses—an improvement over the capabilities of emerging narrow AI technologies and another step

toward the longer-term development of an artificial general intelligence.

* The authors are grateful to Michael D. Schneider, LLNL, for his contributions in this area.

Page 18: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 18

UNCLASSIFIED

Potential Military Applications of AI

Emerging narrow AI technologies are well-suited to varied DOD applications

AI and associated technologies have dozens of potential defense-related applications (some being actively

considered, in development, or already deployed).* We group these applications into seven broad categories:

1. Logistics.

2. Sensor-to-shooter situational awareness.

3. Unmanned vehicles.

4. Cyber and electronic warfare.

5. Wargaming.

6. Autonomous weapons.

7. Command and control.

Our qualitative evaluation of the value of AI for specific DOD applications is predicated on three key

variables: the utility to DOD if suitable AI technologies were available, the technical feasibility of such

developments, and their potential adoptability in the DOD context. This is informed by the following:

• Scope of applicability.

• Maturity of the principal enabling capabilities.

• Levels of commercial investment in related applications.

• Mission enhancement prospects.

• Potential for cost savings.

For each application category, which we describe in the numbered subsections below, we introduce AI’s

function, the main AI technologies involved, and what we consider to be the scope of AI’s potential impact

over the short term (within five years), medium term (within 10 years), and long term (beyond 10 years).

In this context, we emphasize three caveats: (1) AI technologists have historically struggled to accurately

forecast the pace and scope of AI developments; (2) the various categories overlap and, in some cases,

demonstrate cross-dependencies; and (3) effective DOD implementation requires mitigation of acute

organizational, cultural, and other nontechnical challenges. While the first two items reflect challenges

associated with technology development, the latter underscores various technology adoption hurdles that

DOD would need to overcome to fully exploit AI’s mission or enterprise potential. The time horizons we

present generally correspond with how we view the level of technical difficulty involved and the

corresponding achievability of operationally relevant progress within the envisioned timeframes. Across

the seven categories, we identify relevant contemporary DOD and commercial spending and point to their

* Military applications of AI technologies proposed or otherwise discussed in open-source publications include, but are not limited to, the following (listed alphabetically): air defense; automated planning; autonomous weapons; collection management; command and control; communications; conflict prevention; cyber operations; equipment maintenance; indication and warning; intelligence collection; intelligence exploitation; logistics; manpower allocation; medical care; mission handoff; modeling; reconnaissance; robotics; search and rescue; sensor fusion; sensor-to-shooter situational awareness; simulation; surveillance; targeting; training; translation; unmanned vehicles; wargaming; and weapon development and optimization.

Page 19: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 19

UNCLASSIFIED

relative near-, mid-, and long-term value across the spectrum of potential defense-mission applications for

AI. Within each category, we suggest how the state-of-the-possible may evolve over these timeframes, and

we estimate China’s current level of interest for initial benchmarking purposes.

To illustrate, investment in logistics-relevant AI technologies is both operationally useful and technically

achievable in the near to medium terms, but the payoffs from long-term investments reflect diminishing

returns. At the same time, existing or developmental AI-infused technologies would support greater

autonomous weapon system operation, but adoption prospects may be comparatively limited for policy or

other nontechnical reasons. In turn, while AI-associated investments in command-and-control technologies

would probably have substantial payoff over the long term, the current and prospective state of technology

over the next decade (and probably longer) will not likely support robust application in this area.

1. Logistics: AI useful across a broad range of tasks

AI applied to logistics (Figure 4).67,68 AI in the realm of military logistics refers to the use of AI and related

technologies such as computer vision, natural language processing, and robotics to streamline or otherwise

improve the full scope of DOD logistics operations, including acquisition, inventory management,

maintenance, production, routing, supply, transportation, and warehousing. Efficient, timely, and reliable

logistical support is a critical element for operational success, on par with military manpower, budgets,

infrastructure, or combat systems.69 Businesses are investing heavily in intelligent logistics applications,

resulting in a growing number of mature commercial applications that can potentially be leveraged or

adapted for DOD purposes.

Figure 4. Potential utility of increased AI use in logistics over the short, medium, and long terms,

compared with current levels of DOD and commercial investment and China’s estimated level of

interest. Information on China is derived primarily from Elsa B. Kania, Battlefield Singularity.70

Page 20: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 20

UNCLASSIFIED

Scope of AI’s potential for logistics. DOD has successfully employed narrow AI to automate specific

functions, such as fixing and repurposing aircraft parts,71 but a more widespread application of the full suite

of technologies could better ensure asset optimization, reduce supply and resupply costs, and diminish

operational risk.72 The commercial robotics industry is thriving, and as the demand for these systems

increases across different sectors, control and autonomous logic will improve along with the dexterity and

capability of the systems.73 With some modifications, many of these systems could improve a range of

different logistical functions for the military.

• AI in the short term can help automate a growing range of generally predictable logistical tasks,

such as equipment maintenance, inventory management, and resupply on preplanned missions.74

Algorithms can provide autonomous assessments and predictions of equipment degradation and

the remaining life of machines and their components, adjusting functionality according to health

status. Some of the same technologies could enable the adjustment of transit, shipping, and sourcing

routes based on weather patterns or other environmental conditions. Improvements in robotics can

better enable off-road ground-vehicle autonomy for use in resupply and assist with autonomous

route and obstacle clearance and other work considered repetitive, dangerous, or difficult for

humans.75,76

• Improvements in computer vision and robotics in the medium term could enable the development

of robotic transport systems that could carry supplies and help soldiers and tactical units avoid

threats, maneuver and clear objects efficiently, and initiate contact under favorable conditions.77

Civilian robotic applications, such as those that are designed to operate in urban areas and those

designed to interact with humans, are likely to see significant innovation over the medium term and

could be adapted to the military domain.78 Logistical drone swarms could supply food or

ammunition packages, enabling supply or resupply in dispersed locations without risking human

lives. The Navy, for example, seeks to deploy unmanned helicopters that could be called up on

demand to autonomously navigate obstacles and threats, select a safe landing zone, and unload

supplies or evacuate casualties, a capability that could potentially be realized over the medium

term.79 AI could help optimize resource allocations in other similarly complex tasks, such as

enabling life-saving efficiencies during a medevac operation, automatically supplying medical

information about troops in combat, helping diagnose injuries, and analyzing routes, weather, and

local landing sites to determine the best form, course, and method of evacuation.80

• Over the long term, AI-enabled logistical systems will become more interoperable, allowing for

more effectively integrated logistics. With improved logistical autonomy, soldiers will be better

able to sustain high-tempo operations and focus increased attention on the combat “tooth” over the

support “tail.”

2. Sensor-to-shooter situational awareness: AI a force multiplier for “big data” exploitation

AI applied to sensor-to-shooter situational awareness (Figure 5).81,82,83 AI applied to sensor-to-shooter

situational awareness centers around using computer vision and natural language processing to augment or

automate the collection, processing, exploitation, and dissemination of data. Mushrooming volumes of live

video and other data streaming from sensors in the skies and on the ground have increased the need for new

tools and capabilities to help automate the processing and exploitation of data and help deliver either raw

or actionable information to operators at all levels in real- or near-real-time.84 Since image classification is

a task in which existing deep learning systems already have exceeded human performance, the potential for

cutting-edge AI technologies to galvanize these processes is significant.

Page 21: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 21

UNCLASSIFIED

Figure 5. Potential utility of increased AI use in sensor-to-shooter situational awareness over the

short, medium, and long terms, compared with current levels of DOD and commercial investment

and China’s estimated level of interest. Information on China is derived primarily from Elsa B. Kania,

Battlefield Singularity.85

Scope of AI’s potential for sensor-to-shooter situational awareness. One of the most advantageous uses

of cutting-edge AI capabilities is to sort through big data, suggesting significant near-term potential for AI

technologies in this area. DOD is in the early stages of exploring AI’s potential for this application. The

Algorithmic Warfare Cross-Functional Team (AWCFT) established in 2017 seeks to “accelerate DOD’s

integration of big data and machine learning” by “turn[ing] the enormous volume of data available to DOD

into actionable intelligence and insights at speed.” Over the past several years, personnel supporting the

campaign against ISIS in Iraq and Syria have observed video feeds and imagery and manually recorded

information about activities, objects, or people of interest. To make this process more efficient, the AWCFT

is working to couple existing computer vision algorithms with machine learning techniques to

autonomously detect and classify objects in video streams in order to truncate the cycle time associated

with such operations to enable actionable information.86,87

• The current state of deep learning technology is such that computers can vastly accelerate the

processing and exploitation of information where large, labeled datasets exist; improved accuracy

comes with larger training datasets.* One of the most successful applications of deep learning has been

in image recognition, so readily exploitable technologies exist.88 While AI can serve as a significant

force multiplier, some important limitations persist. These include imperfections in the technology—

including the risk of machine-learning algorithms developing biases89—and the inability of most deep-

* As an example of the size of the training datasets that generally are required, the image classifiers that competed in the 2014 ILSVRC trained on a set of 1.2 million images distributed among 1,000 categories. The training required significant human effort to provide a large enough sample space of “correct” labels. [Reference: A. Ilachinski, Center for Naval Analyses | AI, Robots, and Swarms: Issues, Questions, and Recommended Studies | January 2017 | p. 63.]

Page 22: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 22

UNCLASSIFIED

learning-enabled machines to explain the logic behind their findings, a critical shortcoming where

battlefield and DOD senior leaders may require further transparency to justify action.

• Over the medium term, improved machine learning training models have the potential to further

reduce manpower requirements and integrate a much wider range of different data types to spot

objects, patterns, and abnormalities for indications, warning, and decision advantage, particularly

with further progress in unsupervised and few-shot learning techniques. However, the prospect of

increasing volumes of tainted data, including sophisticated data designed to mislead these systems,

may pose a growing challenge. As Greg Allen and Taniel Chan of the Belfer Center for Science

and International Affairs explain, AI “can assist intelligence agencies in determining the truth, but

it also makes it easier for adversaries to lie convincingly.”90

• Over the longer term, it is possible that relevant AI technologies could be developed to support the

full spectrum of man-in-the-loop, man-on-the-loop, and fully autonomous operations. For now,

intelligent systems face challenges in identifying credible threats and routing the information to an

appropriate authority or initiating the desired response. Facebook’s director of global policy

management noted in June 2017 that computers “are not very good at identifying what constitutes

a credible threat that merits escalation to law enforcement.”91 While deep-learning-enabled and

deep reinforcement-learning-enabled computer vision algorithms excel at detecting simple objects

or events that match specific predefined criteria, they struggle to infer abstract meanings from

images, video footage, and real-life situations.92 For example, cutting-edge computer vision

systems can recognize some simple human actions such as walking, running, or hand-waving, but

they cannot reliably determine the intent behind those actions.93 We expect that significant

improvements on this front are at least a decade away.

3. Unmanned vehicles: more varied and complex missions likely as key technologies mature

AI applied to unmanned vehicles (Figure 6).94,95 AI applied to unmanned vehicles involves using

computer vision and robotics to provide unmanned mobile robot platforms—including aerial, ground,

maritime, and space vehicles—with the capability to independently compose and select among different

courses of action to accomplish their goals.96,97 Although the DOD has significantly increased its adoption

of unmanned vehicle systems over the past decade, the vast majority of these systems are remotely operated

rather than truly autonomous or intelligent.98 Reliance on remote operators is creating a growing manpower

burden on the U.S. military, and their reliance on data links constitutes a potential vulnerability via cyber

warfare, electronic warfare, or counterspace attack.99

Scope of AI’s potential for unmanned vehicles. The development of autonomous vehicles for the military

has been slow and incremental, trailing advances made in autonomous commercial systems such as aerial

drones and driverless cars.100,101 DOD relies on autonomous systems for select missions, such as

reconnaissance and explosive ordnance disposal; use across the Armed Forces remains largely relegated to

niche roles.102 Military applications for autonomous swarming techniques, meanwhile, remain

comparatively under-explored.103,104 Increasingly intelligent unmanned systems could provide significant

operational benefits, giving the military greater reach and persistence into denied areas, enabling more

daring concepts of operation, and allowing a shift in manpower from operator roles to oversight

responsibilities.105,106 The Center for a New American Security’s Paul Scharre posits that advances in swarm

technology could usher in a paradigm shift in warfare where “mass once again becomes a decisive factor

on the battlefield” and where “having the most intelligent algorithms may be more important than having

the best hardware.”107

Page 23: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 23

UNCLASSIFIED

Figure 6. Potential utility of increased AI use in unmanned vehicles over the short, medium, and long

terms, compared with current levels of DOD and commercial investment and China’s estimated level

of interest. Information on China is derived primarily from Elsa B. Kania, Battlefield Singularity.108

• Short-term advancements in computer vision are likely to enable greater unmanned system

autonomy, allowing a single operator to more easily control multiple unmanned systems and

improving route planning capabilities.109,110 Available technologies are likely to aid onboard

information processing, potentially leading to significant size, weight, and power modifications.

Adoption prospects for highly capable autonomous systems may be somewhat greater in the near

term in the air and sea domains because they are comparatively less cluttered than ground

applications, where the structure of the terrain can vary significantly and the systems are more

likely to encounter other agents—machines or humans—whose behavior may be unpredictable.111

• Continued advancements in computer vision, natural language processing, and robotic systems

over the medium term are likely to enable a broader array of autonomous and human-machine

and machine-machine teaming operations, such as longer transit and navigation in comparatively

more dynamic or denied-access environments, autonomous support to manned vehicles, and

limited heterogenous autonomous swarming.112 Small unmanned air vehicles could be used for

focused reconnaissance missions, searching for targets and relaying coordinates back to human

controllers.113 Multiple unmanned systems designed for collaborative operations that are either

under development or in a demonstration phase could fully mature over the medium term.114,*

Superior swarming capabilities could enhance the prospect of decisive advantage over

adversaries by overwhelming enemy defenses and jamming enemy radars. Intelligent swarms

* Examples of such systems include the UTAP-22—an unmanned aerial system developed by Kratos that in test flights has shown the capability to fly in formation with other systems in several different scenarios and configurations—and small, low-cost systems for ISR missions. For additional examples, see page 29 to 34 of V. Boulanin and M. Verbruggen of the Stockholm International Peace Research Institute | Mapping the Development of Autonomy in Weapon Systems | November 2017 | www.sipri.org/sites/default/files/2017-11/siprireport_mapping_the_development_of_autonomy_in_weapon_systems_1117_1.pdf.

Page 24: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 24

UNCLASSIFIED

could also serve as a low-cost, expendable means to impose costs on an adversary by forcing the

adversary to counter large numbers of systems,115 although the affordability and accessibility of

such devices makes it likely that these capabilities will become increasingly diffuse in the

medium term.

• AI enhancements over the long term are likely to provide for more seamless collaboration and

interoperability. Over time, unmanned systems probably will become more capable of

autonomously making combat decisions within established legal and policy constraints.116 As with

other AI-enabled applications, a key challenge going forward will be encounters with intelligent

adversaries deploying deception, assault, and counter-autonomy technologies—such as the

adversarial examples for deep learning models discussed above.117

4. Cyber and electronic warfare: AI critical to maintaining the edge

AI applied to cyber and electronic warfare (Figure 7).118,119 AI applications for cyber and electronic

warfare center on the prospective use of large-scale machine learning to automate offensive and defensive

operations. In both areas, AI would be a critical enabler because as networks, sensors, and other electronic

devices become more sophisticated and generate more data and more agile waveforms, machines

exceeding human performance will be able to parse thousands of logs per second in an effort to identify

vulnerabilities to either exploit or defend. Given the widespread occurrences in recent years of costly

network penetrations, data thefts, and computer viruses, commercial entities already seek to use AI to

protect information against cyber intruders—a trend that is likely to continue to drive innovation and, as

such, enable new or enhanced attack or defense options. PriceWaterhouseCoopers found in 2015, for

instance, that 79 percent of those responding to its industry survey had detected a cybersecurity incident

during the year.120 Indeed, these incidents provide useful machine-learning training data, and businesses

are investing heavily in AI information security technologies. Not surprisingly in this context, one

industry association projects investment of more than $800 million in 2017 in AI in cybersecurity, up

from just $71 million in 2012.121

Scope of AI’s potential for cyber and electronic warfare. Current offensive and defensive cyber and

electronic warfare technologies generally rely on supervised machine learning. In general, they can sift

through data, search for abnormalities, and sometimes patch vulnerabilities automatically, but they typically

supply information or provide feedback to human decisionmakers for action.122 Significant commercial AI

investment in the cyber domain, in particular, will likely accelerate technical progress in this area. At the

same time, DOD is likely to face increasing pressure to automate electronic warfare and other digital combat

systems to compete successfully against adversaries operating at machine speeds.123

• AI in the short term can serve as a significant force multiplier. Improvements in AI technologies

can support enhanced cyber and electronic attack and defensive options. DOD has increased its

investment in technologies designed to automatically detect, patch, and exploit existing software

vulnerabilities after observing its success against automated software at the DARPA 2016 Cyber

Grand Challenge.124 (In this case, although the software was ultimately defeated by human hackers,

it proved able to rapidly identify some vulnerabilities.125,126) Data integrity and network security

are key challenges in this context, as developing technologies support potential enhancements to

both U.S. and foreign systems.

• Over the medium term, AI technologies may be able to autonomously mitigate and remediate

known threats. In particular, intelligent electronic warfare systems may be able to isolate unknown

or unanticipated radar systems in the presence of friendly, neutral, and hostile signals; rapidly

transmit countermeasure signals; evaluate the effectiveness of their response; and adjust

accordingly.127 In parallel, intelligent software agents may enable machines to “introspect,”

Page 25: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 25

UNCLASSIFIED

generating novel attack vectors against themselves and then modifying their own code to generate

defenses against such attacks.128 In this context, appropriately managing false positives is a

challenge to overcome; current-generation technology excels at flagging potential problems but is

far less adept at determining whether something actually represents a meaningful threat.129

• Over the long term, potentially unpredictable AI-on-AI autonomous system interactions may

materialize. In parallel, growing linkages between digital and physical systems are likely to expand

the number of possibilities for lethal digital operations.130

Figure 7. Potential utility of increased AI use in cyber warfare over the short, medium, and long

terms, compared with current levels of DOD and commercial investment and China’s estimated

level of interest. Information on China is derived primarily from Elsa B. Kania, Battlefield Singularity.131

Note: while this chart emphasizes cyber rather than cyber and electronic warfare, given the greater public

availability of data, the offense/defense competition applies to both.

5. Wargaming: potentially better longer-term prospects

AI applied to wargaming (Figure 8).132,133 AI in a wargaming context revolves around using reinforcement

learning to increase the chances for military victory by accurately predicting how events could unfold,

developing potential options in this context, and identifying best possible courses of action.134 For such

wargaming systems to be effective, they must be realistic; when humans are involved, machines must

generally behave as a human might. The better they can handle dynamic scenarios involving partial

knowledge, overlapping and potentially competing objectives, and numerous actors—a fluid environment

where the enemy also acts in a manner of its choosing—the greater their effectiveness.

Page 26: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 26

UNCLASSIFIED

Figure 8. Potential utility of increased AI use in wargaming over the short, medium, and long

terms, compared with current levels of DOD and commercial investment and China’s estimated

level of interest. Information on China is derived primarily from Elsa B. Kania, Battlefield Singularity.135

Scope of AI’s potential for wargaming. AI-based computational wargames have existed for decades but

have advanced considerably over the past 10 to 15 years. Nevertheless, most systems are programmed to

exhibit only tactical behavior and are more limited in their ability to represent the full range of military

decisionmaking. Advancements in deep reinforcement learning hold the prospect of more realistically

portraying how tactical, operational, and strategic decisions may be affected by a variety of different players

and new technologies.

• In the short term, AI can aid in the development of tactical plans against a narrow problem set or

within a limited set of parameters, some of which may be unknown and must be discovered and

learned. Significant data inputs and time and monetary investments are required to develop and

train such systems, a limitation that is unlikely to be overcome in the short term.

• AI advancements over the medium term probably will enable machines to better recognize a human

player’s tactics and some limited aspects of the player’s overall strategy and more effectively play

out scenarios to second- or third-order consequences. Success prospects are enhanced to the extent

that the technical AI community can further mature deep-reinforcement-learning techniques,

augment relevant training data, and resolve known difficulties in ambiguous and dynamic

environments. It is possible—but not yet certain—that deep-reinforcement-learning techniques will

ultimately defeat human players in more complex games such as StarCraft (see appendix B).

• Over the long term, AI technologies may enable wargaming platforms to incorporate, process, and

exploit real-time data and emerging technology concepts. However, more basic technology

development is required to find out whether AI can successfully add value to decisionmaking in

life-like decision environments.

Page 27: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 27

UNCLASSIFIED

6. Autonomous weapons: intelligent systems are available but are policy-constrained

AI applied to weapon systems (Figure 9).136,137 Much of the public discussion regarding AI in the military

domain has focused on debates about intelligent autonomous weapons, an application area that—depending

on the system—could involve any one or a combination of computer vision, machine learning, and robotic

technologies. Existing policy guidance requires autonomous systems to employ “appropriate levels of

human judgment” unless the weapon system has been explicitly authorized by specified leadership.138,* Not

surprisingly, various nongovernmental organizations seek to restrain government development and

deployment of autonomous weapon systems. In August 2017, for example, more than 116 founders of

robotics and AI companies warned the United Nations that once autonomous weapons are developed, “they

will permit armed conflict to be fought at a scale greater than ever, and at timescales faster than humans

[could] comprehend.”139 Others have characterized such warnings as misguided, noting that the

development of autonomous weapons may be inevitable and potentially desirable—if improved speed and

accuracy lead to fielding autonomous systems more capable and discriminate than those controlled by

humans.140,141,142,143 The international community is just beginning to debate these issues.

Figure 9. Potential utility of increased AI use in autonomous weapons over the short, medium, and

long terms, compared with current levels of DOD and commercial investment and China’s

estimated level of interest. Information on China is derived primarily from Elsa B. Kania, Battlefield

Singularity.144

* Paragraph 4.d of DOD Directive 3000.09 states that “autonomous or semi-autonomous weapon systems intended to be used in a manner that falls outside the policies in subparagraphs 4.c.(1) through 4.c.(3) must be approved by the Under Secretary of Defense for Policy, the Under Secretary of Defense for Acquisition, Technology, and Logistics; and the CJCS before formal development and again before fielding.” [Reference: DOD | Directive 3000.09 | Autonomy in Weapon Systems | Last revised 8 May 2017 | http://www.esd.whs.mil/Portals/54/Documents/DD/issuances/dodd/300009p.pdf.]

Page 28: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 28

UNCLASSIFIED

Scope of AI’s potential for weapon systems. Multiple lethal autonomous systems exist today, but nearly

all are used for defensive purposes and either have semi- or fully autonomous modes—subject to human

oversight. In some cases, they may be limited in their intelligent capabilities; that is, systems are constrained

in the tasks for which they are used, the types of targets they attack, and the circumstances in which they

may be employed.145,146,147 For example, they could be authorized to defend against incoming missiles or

mortar shells in cases where timing is not sufficient for human decisionmaking.148 They can be sensitive to

variations in the environment, such as poor weather conditions or a cluttered landscape, meaning that

systems employing these technologies cannot be used safely in all circumstances.149 Greater use of AI-

enabled autonomous weapon systems affords the prospect of enhanced speed and precision while

potentially enhancing soldier protection.150

• Current state-of-the-art intelligent autonomous weapon systems have some ability to independently

decide how to execute orders, but they are more limited in their capacity to discern threats. For

example, loitering attack munitions—cruise-missile-like devices that are launched into a general

area and whose mission is to loiter, search for targets according to programmed targeting criteria,

and attack them if found—can select and engage targets without human intervention.151,*

Improvements in image classification and object recognition in the short term are likely to improve

the accuracy and the discrimination of these systems, but adversary countermeasures are also likely

to improve.

• Over the medium term, AI advancements probably will fall short of resolving the tension between

developing machines whose behavior is predictable enough that they can be safely deployed yet

flexible enough that they can handle fluid situations.152,153 The availability of training and test data

is a key hurdle, as these systems will need to be trained and tested on data related to identified

mission scenarios; for many target types and operational situations, however, generating

sufficiently large or comprehensive datasets will prove difficult.154 In parallel, AI technology

advancements, such as self-organizing adaptive swarm capabilities (to include loitering electronic

or conventional weapons) are possible with sufficient and sustained DOD investment.155,156

• The debate over how much independence should be given to autonomous weapons systems is likely

to intensify over the medium and long terms, suggesting that policy decisions will likely play a

greater role in shaping autonomous weapon capabilities than technological advancements. As the

Center for Naval Analyses’ Andrew Ilachinski explains, as autonomous systems increase in

complexity, “we can expect a commensurate decrease in our ability to both predict and control such

systems.”157 This is a case where U.S. policy choices will need to account both for the continuing

evolution of relevant technologies and for the choices made by foreign competitors.

7. Command and control: “AlphaWar” is probably decades away

AI applied to command and control (Figure 10).158,159 AI applied in this area centers on the concept of

using machine learning and natural language processing to enhance decisionmaking at the operational and

strategic levels of warfare. Joint Doctrine defines command as “the art of motivating and directing people

and organizations into action to accomplish missions,” and control as “manag[ing] and direct[ing] forces

* The only publicly known operational loitering attack munition is the Israeli Defense Forces’ Harpy, a “fire-and-forget” anti-radar weapon that searches for enemy radars over a designated area and dive-bombs to destroy the radar if found. Three additional Israeli loitering attack systems in operation can find, track, and attack targets in complete autonomy once launched, but they also have human-in-the-loop modes: the Orbiter 1K ‘Kingfisher,’ a compact unmanned aerial system launched via catapult that can detect and destroy a moving or stationary target; the Harop, an anti-radiation drone that can autonomously home in on radio emissions; and the Harpy NG, an anti-radiation drone optimized for the suppression of enemy air defenses. [Reference: A. Ilachinski, Center for Naval Analyses | AI, Robots, and Swarms: Issues, Questions, and Recommended Studies | January 2017 | pp. 50–55.]

Page 29: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 29

UNCLASSIFIED

and functions consistent with a commander’s authority.”160 The goal of integrating AI into these functions

would be to enable commanders to more rapidly generate courses of action, evaluate the impact of potential

decisions on other plans or operations, and envision the sequence of events by which to achieve a desired

end state. As one proposal for an AI-enabled command-and-control system postulates, on the modern

battlefield, “the only invariant is constant change, particularly the situation and goals. Under uncertain and

time-critical conditions, it is important for commanders to have the ability to rapidly understand the

unfolding trajectory of the operation and generate options quickly.”161 The need for such a capability is

likely to become even more acute as AI advancements enable greater autonomy—and, by extension, faster

processing—of adversaries’ information, logistics, and weapons systems.

Figure 10. Potential utility of increased AI use in command and control over the short, medium, and

long terms, compared with current levels of DOD and commercial investment and China’s evident

level of interest. Information on China is derived primarily from Elsa B. Kania, Battlefield Singularity.162

Scope of AI’s potential for command and control. DOD has pursued the development of AI-enabled

military decision support systems for more than a decade, with limited success.163,* The key obstacle has

been the inherent complexity of military decisionmaking, which is vastly more complicated than games in

which machines have outperformed humans thus far. The progression from a point in which a machine

proved capable of defeating a human world champion in chess to the more complex game of Go took nearly

* Between 2004 and 2008, the Defense Advanced Research Projects Agency (DARPA) pursued a project called Real-time Adversarial Intelligence and Decision-making that proposed courses of action for tactical leaders. In 2007, DARPA released a requirement for Deep Green, a tool to help battlefield commanders with decisionmaking during mission execution. In 2009, the University of Iowa’s David Ezra Sidran attempted to create a computer program called the Tactical Inference GenERator that could evaluate a battlefield situation and classify it in the same way as a group of subject matter experts. [Reference: Maj. S. Banks, School of Advanced Military Studies, United States Army Command and General Staff College | Lifting Off of the Digital Plateau with Military Decision Support Systems | 2013 | http://www.dtic.mil/dtic/tr/fulltext/u2/a583735.pdf | pp. 43–48.]

Page 30: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 30

UNCLASSIFIED

two decades.164,* In comparison, the state and action spaces involved in military decisionmaking are

probably many orders of magnitude more complex than they are for Go, as are the goals and risks. Humans

continue to outperform AI-enabled machines in more complex games such as StarCraft because it features

more complicated multiplayer dynamics, requires balancing between potentially competing shorter- and

longer-term objectives, and incorporates the fog of war (that is, uncertainty about the state of the world),

which deep-reinforcement-learning algorithms are not yet able to effectively address (see appendix B).

• AI technologies in the short term are well suited to support decisionmaking related to specific,

narrowly defined tasks. For example, machines can help optimize decisions regarding team

composition for specific missions, using inputs such as potential team members’ skills, experience,

personality, strengths, and weaknesses. Current technology may also facilitate corporate-level

decisions, such as on resource optimization.

• The medium term holds the prospect of improved sensor-data fusion to enhance operational

decision support. Obtaining and processing large volumes of high-fidelity data, however, is likely

to be a challenge. To train AlphaGo, Google DeepMind used on the order of 30 million moves from

games played by human experts,165 a volume of training data that would be difficult to match in a

real-time battlefield environment where the possible world states are likely to be large and the

sequence of actions necessary to achieve each state reflects a move/countermove dynamic among

multiple parties.

• Some observers have imagined over the long term the development of what has been termed

“AlphaWar,” a hypothetical militarized version of AlphaGo that would be able in real time to ingest

myriad relevant data, run a deep but rapid analysis, develop options, and prioritize and execute

actions faster and more efficiently than human operators.166 Presumably, such a system would be

able to successfully account for multiple assets concurrently performing hundreds of complex tasks.

We expect the timetable for development of such a system to coincide with the achievement of

pervasive general AI capabilities: a development at least two (and probably more) decades away.

Outlook and Considerations

Our qualitative evaluation shows that narrow AI technologies have significant potential to enable high

degrees of automation in a growing range of DOD applications. A combination of AI and associated

technologies will advance the vision of the RMA. In his 1992 study for the Office of Net Assessment of the

military-technical revolution, Andrew Krepinevich described three technical pathways in which

advancements in microelectronics were providing this foundation: (1) the growing ability of the U.S.

military to gather, process, and disseminate information; (2) the increasing capabilities of stand-off

precision guided munitions; and (3) improvements in the ability of computers to conduct advanced

simulations that could help conceptualize, test, and optimize new organizational and operational

concepts.167 In his view, a key breakthrough would occur when the U.S. military succeeded in fielding and

integrating operational information networks for reconnaissance strike, weapon platforms, and battle

damage assessment. Accordingly, two elements would be critical to success: (1) integrating reconnaissance

strike complexes with information dominance operations, thereby increasing space capabilities, unmanned

systems, and automated detection and engagement; and (2) developing advanced technologies, such as

advanced robotics and directed-energy weapons.168

* In 1997, IBM’s chess-playing computer Deep Blue defeated chess world champion Garry Kasparov, a development that has been heralded as a major milestone in human-versus-machine competition. Nearly two decades later, Google’s AlphaGo AI program in 2016 defeated Go world champion Lee Sedol.

Page 31: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 31

UNCLASSIFIED

In our view, emerging AI technologies could be instrumental in helping DOD fulfill these elements. The

broader integration of deep learning technologies into sensor-to-shooter applications advances the first

element and plays a role—in conjunction with the promise of future deep-reinforcement-learning

technologies—in enabling more autonomous platforms that are essential to eventually fulfilling the second

element. With success, the gap between warning and response times, the distinction between “front lines”

and “rear lines,” and the temporal distinction between “early” and “late” phase operations would each most

likely diminish.169 In this context, creating information gaps in which one side is able to rapidly gain

situational awareness and develop strike options while denying those to the adversary is integral to

operational success. AI technologies are beginning to demonstrate the ability to help DOD towards this end

of enabling the creation of information gaps across the spectrum of RMA-like capabilities. This includes,

for instance, providing real-time information to reconnaissance strike; enhancing information systems such

as electronic warfare, cyber, and ISR capabilities; facilitating rapid discovery of adversary mobile

platforms; informing center-of-gravity analyses for targeting and other purposes; automating for advantage

for information dominance; and engaging effectively in space.

AI forecast: foggy with a chance of technology breakthroughs

Any effort to credibly forecast the pace, scope, and depth of developments in AI must successfully navigate

an important disconnect between the actual and the aspirational. On the one hand, expectations for this

technical area are rising, corporate investment is increasing, and real-world applications for narrow AI are

growing. On the other hand, the available technologies are in their infancy, many companies in this sector

are not yet profitable, and overly optimistic expectations for general AI-like capabilities could stymie

progress. In this context, some have argued that AI is “poised to affect not only businesses but also the

everyday lives of people around the world.”170,171

All told, the $26–39 billion and growing annual investment made by industry in AI for a range of

commercial applications is expected to lead to prospective multitrillion dollar market opportunities spread

across multiple market sectors.172 Not surprisingly, those leading the charge in AI adoption are proactive,

investing heavily in AI, and more digitally inclined. In this context, DOD deciding how important AI is to

its future warfighting strategies, operational approaches, and back-office administration is key. While it can

potentially benefit in several areas, the common adoption pitfall identified by Salesforce.com’s Marco

Casalaina should be avoided: “‘Hey, we want to use AI’, without really thinking about why, or what it can

do for them.”173

In DOD’s case, this means choosing where to invest, what to leverage, and how best to adapt an industrial

warfighting structure to the emerging digital landscape. To be sure, this challenge would exist in the absence

of AI; exploiting the art-of-the-possible for cybersecurity, electronic warfare, and other activities is only

made more urgent with the rise of AI and related technologies. Similarly, while AI may help the department

achieve cost-savings through significant administrative efficiencies, this is true if—and only if—DOD is

ultimately able to capitalize effectively. Efforts to incorporate AI more broadly may help only at the margins

and could even be counterproductive, if it becomes simply another layer in a complicated, process-heavy

environment: it “isn’t worth much if people are feeding them bad data in the first place or don’t know what

to do with information or analysis once it’s provided.”174,175

A complicating factor in determining whether to lean heavily into this technology area or to adopt more of

a wait-and-see approach relates to expectations for progress in specific AI technologies. At once, those

closest to the AI enterprise have underestimated the pace of scope of key technical achievements, such as

with AlphaGo, and overhyped more substantial developments, such as the prospects for realizing artificial

general intelligence. Based on the current state of the art, Google’s DeepMind, for instance, reported in

Page 32: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 32

UNCLASSIFIED

October 2017 that its AlphaGo Zero AI program taught itself—using reinforcement learning—in just three

days what had required its AlphaGo predecessor more than half a year and data from more than 100,000

human amateur and professional games.176 But a “real intelligence,” finds James Somers, “doesn’t break

when you change the problem.”177

History may ultimately record AI’s transformational prospects on par with development of aerospace,

biotechnology, cyber, and nuclear technologies, as Greg Allen and Taniel Chan postulate.178 But while

AlphaWar-style transformational systems are probably two or more decades away, there is no shortage of

plausible—and potentially quite effective—AI-related enhancements to existing or planned warfighting

platforms, tools, or methods. DOD’s ability to cultivate novel technologies and develop associated novel

operational concepts is not in question. But in the current resource-constrained and mission-expansive

landscape, efforts to prioritize efforts and resources is important, as well as tackling the organizational

challenges that often pose roadblocks to military innovation—in particular, the system’s “adoption

capacity.”179 As a general guideline, DOD components acquiring, developing, or using AI could at

minimum prioritize those efforts which accomplish the following:

• Mitigate known risks associated with current approaches or systems. The ability to manage

complex logistics flows is a critical warfighter enabler. In many cases, field commanders have

relied on lengthy truck convoys or aerial resupply missions to enable and sustain operations. In this

context, activities such as the Army’s Leader-Follower Automated Ground Resupply unmanned

ground vehicle program seek to reduce the number and associated vulnerability of personnel

through autonomous vehicles.180,181 In principle, such activities could both diminish operational

risk and provide new combat support options.

• Enable cost-savings over existing approaches. For example, Robert Cardillo, director of the

National Geospatial-Intelligence Agency (NGA), seeks to use AI to automate three-quarters of the

tasks performed by analysts. Anticipating as much as one million times the amount of geospatial

intelligence currently collected in just five years, the current labor-intensive approach clearly will

not effectively scale. An attempt to keep pace with the “rising tide” of data over the next 20 years

would require, in his view, an additional eight million imagery analysts.182,183 This is a case where

AI may prove both cost-effective and capability-enhancing.

• Accelerate development and/or deployment of military capabilities. Teaming with the University

of Missouri, NGA experimented with deep learning to help analysts find surface-to-air missile sites

over a 90,000-square-kilometer area over Chinese territory. Taking just 42 minutes and achieving

the same overall statistical accuracy as human analysts, the AI approach proved greater than

80 times more efficient in its task.184 The Defense Innovation Unit—Experimental has similarly

sought to work with niche commercial entities for military end uses. Among other efforts, they

work with Capella Space, which is developing small-satellite synthetic-aperture radar systems, and

Orbital Analytics, which specializes in tapping machine learning and computer vision for imagery

exploitation purposes.185

• Provide a novel capability or a new solution using existing capabilities. To help meet emerging

operational requirements in the Pacific theater, the Navy and Defense Advanced Research Projects

Agency collaborated to develop a new anti-ship missile. The resulting Long-Range Anti-Ship

Missile is based on an existing missile platform, the Joint Air-to-Surface Standoff Missile—

Extended Range. More significant, however, is its AI-enabled performance: humans designate

specific targets and provide relevant data while the missile executes the assigned mission,

navigating autonomously around enemy radar and defenses and adjusting course in the absence of

human supervision or satellite guidance.186,187 Such tools arguably enhance operational flexibility

and provide new attack options aligned with emerging theater priorities.

Page 33: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 33

UNCLASSIFIED

Based on the current state of the art for AI technologies, their prospective development over the next several

years, and their potential application to DOD enterprise-level and mission functions, we envision the

following five-part approach to incorporating AI and related technologies (table 2 and appendix C):

1. Acquire or adapt existing commercial technologies for routine administrative tasks. Arguably,

this is the low-hanging fruit: application of or limited modification to existing, commercially

developed technologies for “back office” functions. According to the Defense Business Board,

DOD plans to spend approximately $670 billion over the FY16–20 timeframe in administrative

costs associated with logistics and supply chain management; acquisition, procurement, and

financial flow management; health care and medical services; human capital recruitment and

retention; and infrastructure operations and management.188,189 Substantial commercial

investment in AI-enabled logistics-related technologies is likely to produce a range of off-the-

shelf capabilities that DOD could use or adapt, providing utility and cost savings beginning in

and extending beyond the FY16–20 timeframe, as long as DOD investments are directed

appropriately. For instance, in its 2017 survey of 2,500 corporate executives, PwC found that

54 percent agreed that AI solutions already implemented had improved productivity.

Table 2. Possible DOD approaches to AI over the next decade. (cont.)

Area Action Potential payoffs Key challenges

Streamline “back office” functions (e.g., human resources, travel).

1. Acquire or adapt existing commercial AI products and services.

Wider adoption of AI technologies could provide cost savings and operational efficiencies, improving DOD’s tooth-to-tail ratio and potentially:

• Enabling reinvestment in other DOD accounts (e.g., modernization).

• Allowing top-line reductions to DOD without sacrificing combat power.

• Technology adoption: enabling digital footprint/cloud environment uneven or lacking.

• Human capital: structural dislocation prospects for some DOD and contract support personnel.

Improve select combat support functions (e.g., logistics).

2. Acquire or adapt existing commercial AI technologies.

Greater use of AI-infused autonomous vehicles and predictive maintenance technologies could:

• Reduce the number of required support personnel, decreasing operational risk.

• Realize cost-savings, streamline time-phased force flows, and improve operational planning.

• Procurement: probably applies more to new system acquisitions than legacy system upgrades.

• Deployment: Pilot activities may be useful, but use at scale is required to realize substantial benefits.

Provide support to current operations (e.g., ISR).

3. Adapt or partially develop AI tools for specific defense missions.

Rapid processing and exploitation of high-volume data streams could:

• Generate actionable information.

• Make use of information that would otherwise go unexploited.

• Reduce the signal-to-noise ratio, improving analytic focus.

• Sidestep the need for substantial increases in analytic personnel to keep pace with growing information flows.

• Need/capability mismatch: labeled data is the long pole in the tent—but in short supply.

• Contracting: the rapid pace of commercial developments requires greater use of rapid acquisition authorities, prototyping, and warfighting experimentation.

Page 34: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 34

UNCLASSIFIED

Table 2. Possible DOD approaches to AI over the next decade. (cont.)

Area Action Potential payoffs Key challenges

Enhance fielded and developmental combat systems (e.g., electronic warfare or cyber).

4. Develop and deploy new or improved AI products.

A spiral development or modular upgrade approach could help:

• Ensure that offensive and defensive systems effectively pace adversary capability developments in fast-moving technology areas.

• Enable more effective operations through machine-speed performance.

• System integration, test, and evaluation: against a backdrop of rapid and continuing adversary capability developments, timely and routine upgrades require effective partnerships across the value chain, from program offices to system integrators to niche technology providers.

Develop new combat capabilities (e.g., new technologies and new operational concepts).

5. Design, develop, prototype, and experiment with new AI technologies in operational context.

Maturing and capitalizing on narrow AI technologies for DOD missions:

• Opens the door to new military options, including reduced-risk and enhanced-capability force packages.

• Contributes to better understanding of the pace, scope, and implications of key technology developments.

• Helps with capabilities-based planning—determining how an adversary may choose to employ AI-related systems.

• Development: DOD has a limited AI-related talent bench, and its industrial partners field a finite and oversubscribed AI talent pool focused primarily on commercial products and services.

• Prototyping and experimentation: critical to success, yet underused or only selectively applied across the defense establishment.

Looking ahead, those polled sought to tap AI to reduce a range of routine administrative tasks

such as paperwork (82 percent), scheduling (79 percent), accounting (69 percent), and human

resources (60 percent).190 Applied at scale in DOD, such efforts could in principle save billions

over any given five-year budget cycle. At its core, this is primarily a technology adoption

challenge. To succeed in this area, DOD would need to enhance its digital adoption and would

need to plan for localized workforce disruption. In principle, this could enable DOD either to

free up resources for reinvestment in modernization or other activities or to enable top-line

reductions without sacrificing combat power.

2. Acquire or adapt commercial products or services to improve select combat support

functions. DOD could build on pilot initiatives to improve select combat support functions—

most notably with respect to supply chain efficiencies and logistics management. As above,

substantial commercial investment in AI-enabled logistical capabilities could help DOD

achieve this goal at or near current funding levels if DOD channels its resources effectively.

Departmental efforts in this area should also be able to capitalize on commercial development

of autonomous vehicle and related technologies. Adapting key AI technologies for DOD

applications for predictive maintenance or semi-autonomous resupply operations in the near-

to-medium term could help reduce the number of support personnel required, thus decreasing

operational risk, achieving cost savings, streamlining time-phased force deployment flows,

and improving operational planning. Because retrofitting existing ground, naval, or air

platforms would probably not prove cost effective, it might be reasonable to adapt and apply

AI systems to future procurements. While pilot activities are useful, adoption at scale is likely

required to fully realize both significant cost efficiencies and performance enhancements.

Succeeding is primarily a business-model challenge for defense acquisition. While DOD

Page 35: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 35

UNCLASSIFIED

Instruction 5000.02 allows for spiral development and incremental deployment, DOD’s

ability to leverage rapidly developing commercial capabilities will require full use of

available rapid-acquisition mechanisms.191,*

3. Leveraging commercial technologies, conduct limited development of AI tools to provide

enhanced support to current reconnaissance, surveillance, and target acquisition (RSTA)

operations. Our qualitative evaluation of applying AI for sensor-to-shooter situational

awareness applications suggests that AI’s broader incorporation could enhance support to

combat operations where sufficient labeled training data and computational horsepower are

available, but cost savings may not be realized until the medium term and are likely to depend

at least somewhat on continued progress in unsupervised and few-shot learning techniques.

The relative level of DOD investment in enabling technologies such as ISR-related computer

vision, virtual agents, and machine and deep learning for now seems appropriate; commercial

funding for underlying technologies provide a credible foundation for enhanced collection and

exploitation of relevant data, but Project Maven has demonstrated that limited or modest DOD

development of such tools for identified military purposes is required. If successful, a

combination of faster processing and an ability to handle larger volumes of data can lead to

reduced latency in sensor cueing and increased accuracy in data triage, which together have the

potential to generate actionable information and accelerate military operations. At the same

time, greater use of AI could make use of information that would otherwise go unexploited,

reduce to signal-to-noise ratio in data collected and consequently improve analytic focus, and

sidestep the need to substantially grow the workforce to keep pace with the increasing volume

of data available.

4. Develop and deploy AI products designed to enhance fielded and developmental combat

systems such as automated electronic warfare or cyber defense. While commercial

developments can in some cases provide a springboard for cybersecurity, electronic warfare,

or other algorithm-intensive technologies, in many cases DOD will need to customize—and

continuously refresh—its combat platforms and associated payloads. As above, labeled

training data is a key ingredient for more effective and efficient operations. But development,

deployment, and modernization of effective digital combat capabilities also pose acute system

integration, test, and evaluation challenges against a backdrop of continuing adversary

capability developments. In this context, a spiral development or modular upgrade approach

can help ensure that offensive and defensive systems pace or exceed adversary developments

in fast-moving technology areas. Ultimately, timely and sustained modernization of deployed

systems will require both a sustainable resource base and productive partnerships across the

value chain, from program offices, to technology developers, to system integrators, to front-

line units.

5. Design, develop, prototype, and experiment with a combination of new AI technologies and

new operational concepts to develop new or enhanced combat capability. This is at the core of

the third offset strategy, and the most challenging of this five-part approach to effectively

realize. In principle, maturing AI technologies open the door to new military options, including

reduced-risk and enhanced-capability force packages. In practice, as our qualitative evaluation

of AI applied to command and control and, to a lesser extent, wargaming revealed, the potential

payoff in many cases may not be realized until the long term. With wargaming, in particular,

DOD investments may not be achieving the desired near-to-medium-term bang for the buck.

* This instruction allows for development and delivery of software-intensive capabilities in one- to two-year cycles (Model 3, pp. 11–12) and other acquisition measures when schedule considerations outweigh those relating to cost or technical risk (Model 4, p. 13). [Reference: DOD | DOD Instruction 5000.02 | Operation of the Defense Acquisition System | 26 November 2013 | https://www.acq.osd.mil/fo/docs/DSD%205000.02_Memo+Doc.pdf.]

Page 36: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 36

UNCLASSIFIED

DOD could focus its marginal dollar in this area on prototyping and experimentation—

practices that are underutilized or only selectively applied across the defense establishment—

with an eye toward better understanding the full potential of current cutting-edge technologies.

Just as important, both DOD and its industrial partners draw upon a finite and oversubscribed

talent pool. While too early to tell, new entities such as the Defense Digital Service and greater

connectivity to innovation hubs in Silicon Valley, Boston, Austin, and elsewhere may help.

Certainly, more can be done to groom talent within the Service Laboratories, to partner with

relevant pockets of expertise within the Department of Energy-managed National Laboratories,

and to incentivize relevant academic pursuits nationally. Ultimately, measures to augment the

nation’s AI bench depth will prove critical to design, development, and fielding of next-

generation AI-capable combat capabilities.

Among other things, the rapid pace of development in AI, robotics, and related technologies

suggests the need for robust experimentation and, where appropriate, rapid acquisition and

spiral development. Such an approach is consistent with the Defense Innovation Initiative,

which calls for development of advanced technologies, new operational concepts, and

enhanced use of wargaming, as well as with the department’s Better Buying Power 3.0

initiative, which seeks reinvigorated use of prototyping and experimentation to facilitate rapid

fielding of advanced weapon systems and the opportunity to explore novel concepts of

operation.192,193 Experimentation in this context centers on developing and field-testing

prototypes for evaluation in an applied setting. As such, such activities help establish or refine

the requirements and manufacturing processes, can help reduce or eliminate technical or other

acquisition risks, highlight both potentially useful (and counterproductive) operational use

concepts and tactics, and may help identify both unanticipated attack options and system

vulnerabilities.

For example, the Low-Cost Unmanned Aerial Vehicle Swarming Technology (LOCUST,

figure 11) highlights the potential for intelligent systems to “autonomously overwhelm an

adversary.”194 But what does it mean to overwhelm? Experimentation is required to determine,

among other things, the operational circumstances in which such a capability might prove most

effective, the potential utility of mixed swarm payloads, the efficacy of possible defensive

measures, the requirements for and prospective roles of such a capability in Service and Joint

force concepts, and prospects for both risk mitigation and technology improvements. Much the

same would apply to prototype systems such as the U.S. Navy’s Antisubmarine Warfare

Continuous Trail Unmanned Vessel195 or its unmanned undersea “distributed lethality”

concepts;196 the performance of combined manned/unmanned force packages, such as the Air

Force’s loyal wingman concept;197 AI’s ability to effectively perform cybersecurity198 or

electronic warfare199 tasks; the potential use of AI to rapidly triage and identify items of interest

within large volumes of data, such as mobile missile launchers;200 and many other potential use

cases.

Virtually any offensive or defensive capabilities drawing significantly on “big data” may benefit

from AI. While the AWFCT pilot activity focuses in the counterterrorism mission area, such an

approach could also be applied to algorithmic-heavy activities such as space situational

awareness. Focused defensive experimentation for DOD may also be useful. For example, while

Deputy Secretary of Defense Robert Work underscored that “we are not going to design weapons

that choose what target to hit,”201,202 it is not clear that other states may similarly self-constrain.

For example, the Final Experimental Demonstration Object Research (FEDOR) humanoid robot

developed by the Russian Foundation for Advanced Research Projects is reportedly capable of

performing a wide range of human-like tasks—including precision small-arms shooting

Page 37: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 37

UNCLASSIFIED

(Figure 12).203 While Russian Deputy Prime Minister Dmitry Rogozin claims that FEDOR is not

intended as a Terminator-style robot,204 the prospect that U.S. forces could in the future

potentially face capable AI-infused robotic adversaries highlights the utility of tailored

experimentation as a way to understand the military potential of developmental systems.

Figure 11. LOCUST demonstration, April 2016.

Figure 12. FEDOR.

Page 38: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 38

UNCLASSIFIED

The principal risks inherent in this five-part approach to AI are both modest and manageable:

• Risk 1: The commercial market fails to materialize. DOD’s ability to fully harness AI’s potential

is predicated on continued private-sector investment. In turn, commercial investment depends on

development of a sufficiently profitable marketplace for AI-related products and services. In past

AI-related boom-and-bust cycles, commercial entities largely proved unsuccessful in their ability

to expand the market for developmental products. However, recent and continuing advances in

computational capability and data availability are poised to solve the core challenges that stymied

previous attempts to grow the business- and consumer-oriented marketplace for AI products and

services. Based on the available data, AI appears to be in a virtuous commercial cycle. For more

than a decade, AI applications such as Google’s PageRank and Amazon product recommendations

have existed in the consumer space while enterprise processes have benefitted from AI workflow

enhancements. Companies are now developing AI-centric applications, such as the use of computer

vision technologies and deep learning techniques for automated appointment scheduling, predictive

maintenance, inventory tracking, and other uses; AI-enabled applications, such as models that

predict the supply and demand of electricity, are early-stage.205,206 This growing market for AI-

related technologies affords DOD the ability to leverage key commercial developments for national

security applications. Indeed, the department can already leverage existing AI products for the first

two elements of the approach to AI described above.

• Risk 2: Anticipated AI technologies underperform. A related possibility is that AI technologies

could ultimately overpromise and underdeliver. This risk factor primarily relates to expectation

management, which requires a proactive approach to, and sustained DOD senior-leader support for,

AI adoption. AI is not a silver bullet, but rather a force multiplier; many AI technologies will be

relevant for DOD applications. Some envisioned AI tools will probably fail, while others will

almost certainly succeed in unanticipated ways. Narrow AI applications can in theory help DOD

achieve administrative efficiencies and enhance its combat capabilities, but only if DOD

demonstrates the will to evolve its business processes and the operational agility required to wage

algorithmic warfare. In some cases, even “underperforming” AI technologies—presumably, those

which do not live up to their theorized potential—may constitute an improvement on currently

fielded systems. Greater use of experimentation and pilot-scale activities such as Project Maven,

which embodies the third element of the approach to AI described above, can help ground

expectations and demonstrate the technology’s evolving potential.

• Risk 3: DOD fails to fully embrace the technology. AI adoption prospects are highest in more

digitized market sectors, among entities that are proactive in their approach, and among those who

both harbor an intent and demonstrate a willingness to invest in relevant technologies. As a point

of departure, DOD senior leaders have publicly supported development and use of AI-related

technologies, both in the context of the third offset and more broadly; many big-ticket DOD

systems—such as the F-35—are inherently digital combat platforms. In turn, increased DOD

investment in cloud-related capabilities can strengthen the department’s ability to adopt AI-related

technologies for varied enterprise and mission applications.207 While it is possible that one or more

projects or programs will ultimately fall short of leadership intent, and while adoption across the

department is likely to remain uneven, AI benefits from DOD’s continuing digital adoption and

will in turn advance its further adoption prospects. At minimum, there appears sufficient

commitment to developing enhanced electronic warfare, cybersecurity, and other digital

capabilities—the necessary precondition for the fourth element in the described approach to AI.

• Risk 4: Desired AI technologies prove more difficult to field than anticipated. This is a two-part

risk factor, stemming from the possibility that (a) desired AI technologies may not be successfully

developed within desired cost, schedule, or performance parameters and (b) they may not be fully

integrated within the defense posture as a result of organizational, operational, policy, or other

Page 39: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 39

UNCLASSIFIED

considerations. In this context, history suggests an organizational willingness to incorporate AI

technologies in key areas, such as autonomous or semi-autonomous defensive weapon systems

employment. While DOD senior leaders suggest their intent to make greater use of AI and related

technologies in the future, the extent to which such capabilities ultimately transition at scale is a

key question. Consistent with the fifth element of the described approach to AI, prototyping,

experimentation, wargaming, and related activities will be important to help more broadly shape

future operational concepts, use doctrine, theater planning, and defense strategy.

• Risk 5: Others out-innovate DOD, leading to tactical, operational, or strategic surprise. While

many would agree with former Baidu chief scientist Andrew Ng that “the U.S. currently leads in

AI,” it is unclear how enduring that advantage might be.208 In part, because the bulk of AI-related

research is conducted by industry, the prospects for continued U.S. leadership depend on the

maturation of vibrant commercial markets—which draws both talent and investment capital. In this

context, defense innovation prospects increase with an active DOD technology pull, backed by

significant programmatic resources. In part, the prospects for continued U.S. leadership depend on

the extent to which Chinese leadership is able to fulfill its ambitious AI vision.209,210,211,212

Alphabet’s Eric Schmidt calls this a “Sputnik moment” for the United States, while

SparkCognition’s Amir Husain finds that AI is the “next space race.”213,214 But even if

implementation falls short, it is clear that DOD faces serious future competition from Chinese,

Russian, and potentially other modern militaries. Over the next decade, as AI technologies continue

to mature, this emerging competition—grounded in technology change, military systems evolution,

operational innovation, and organizational adaptation—will almost certainly become more acute.

In this respect, modern applications of AI represent both a new dimension to the continuing

revolution in military affairs and a core attribute of DOD’s investments for emerging century

military-technical competitions.

Calibrating Expectations

Advancing the technical state of the art in AI is both timely and consequential, and DOD can benefit from

recent and continuing AI-related developments. To do so, it will need to overcome the development,

adoption, experimentation, integration, procurement, and other challenges identified above. To the extent

that it can do so effectively, DOD stands to realize both administrative efficiencies and novel operational

capabilities. Over the longer term, DOD’s ability to exploit transformational advances in AI and related

technologies will require substantial improvements across the spectrum of necessary doctrine, operations,

training, materiel, leadership and education, personnel, facilities, and policy actions. Failure to effect the

necessary changes will not only affect the fate of the department’s third offset strategy, but will potentially

cede significant warfighting advantages to key U.S. strategic competitors.

For the next several years—and potentially two or more decades—advances in AI are likely to remain more

“weak” than “strong” in nature. However, those concerned with the world-transforming effects of strong

AI systems often paint an apocalyptic picture. Tesla and SpaceX CEO Elon Musk, for example, sees “vastly

more risk” from AI than from North Korean nuclear weapons.215 Similarly, Nobel laureate Stephen

Hawking considers that “the development of full artificial intelligence could spell the end of the human

race.”216 Others, such as former Microsoft head Bill Gates, have developed a more nuanced view. AI

“should be positive if we manage it well,” he observed in 2015, but in a few decades the intelligence might

be “strong enough to be a concern.”217,218 In turn, Google AI chief John Giannandrea finds that, while

machine learning and related AI technologies “will revolutionize many vertical industries,” speculation

over generalized superhuman intelligence is “unwarranted hype.”219

Page 40: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 40

UNCLASSIFIED

Below the apocalyptic threshold, others focus on what they view as inappropriate or counterproductive

applications of AI systems. For example, the Future of Life Institute has issued an open letter warning that

development of lethal autonomous weapons—those able to select and engage targets without human

intervention—is both feasible in the near term and dangerous.220 In this view, a global ban on offensive

autonomous weapons “beyond meaningful human control” is needed to prevent a “military AI arms race.”

Whether or not the U.S. national security community ultimately embraces and aggressively develops a

broad array of military AI applications, it is unlikely that an arms-control-style ban will prove effective. At

minimum, prudent defense planning would need to anticipate less-than-complete compliance and hedge its

bets accordingly. At worst, any such effort to ban weapons with AI technologies could lead to false hopes

of security—a modern-day instantiation of the 1928 Kellogg-Briand pact to outlaw war.221

Taken together, the apocalyptic fears of Elon Musk and the unrealistic hopes of the Future of Life Institute

share a common theme: AI’s revolutionary potential could ultimately lead to undesirable consequences.

For DOD, neither view is constructive, and focusing at length on such views serves primarily to distract

from the central issue at hand: how DOD could position to capitalize effectively on promising AI

technologies. For at least the next decade, and probably longer, the underlying assumptions include the

following:

• Artificial general intelligence is not at all likely, and therefore should not present a material

concern.

• Operationally relevant advances in narrow AI are very likely, with prospective defensive and

offensive mission applications.

• Development of AI technologies reflects a globalized supplier base, which suggests that multiple

actors will have access to militarily relevant AI technologies and that an offense/defense dynamic

will likely materialize among leading military competitors.

• Industry is leading the charge in AI technologies, which suggests a productive DOD focus in both

leveraging commercial developments and focusing on niche military applications.

• AI technologies are probably less well-suited to arms-control- or nonproliferation-style policy

approaches than to deterrent, hedging, or other defensive strategies.

DOD senior leaders face a significant expectation-management challenge: striking an appropriate balance

between overhype and undersell of a rapidly emerging technology area. For at least the next decade, AI

technology developments are better viewed as prospective evolutionary advances to existing virtual

assistants or developmental self-driving vehicles than the revolutionary change that would be represented

by the HAL-9000-style computer of 2001: A Space Odyssey or the swarming autonomous combat systems

of the film Ender’s Game. In some areas, however, AI developments will be able to provide novel solutions

to a number of vexing operational, administrative, and warfighting challenges inherent in the emerging

international security environment.

Page 41: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 41

UNCLASSIFIED

Appendix A: Challenges Associated with Credibly Estimating Commercial

Investment in AI-Related Technologies and Applications

Reliable data on commercial investment in AI and related technologies and applications is important in

formulating judgments about their future prospects. For DOD purposes, such information is useful in

helping to gauge where to leverage corporate innovation in “winning” areas and where to focus investment

to groom desired technologies. Although this information can be instructive, available estimates of

commercial investment in AI vary widely for several reasons, including the following:

• Estimates of macro corporate investment patterns or industry investment trends are available, with

significant error bars. In McKinsey’s estimate, 2016 external investment levels were three times

greater than 2013 levels.

• Market forecasts differ wildly, depending on assumptions made for key variables, such as the pace

and scope of technology maturation or the extent and timing of business, consumer, or government

adoption. McKinsey finds, for instance, that 2025 global market estimates for AI range from a low

of about $644 million to a high of more than $126 billion—a variance of more than two orders of

magnitude.

• Estimates for the development and maturation timeframes for identified AI technologies are

predicated on possible, but not certain, market futures.

• Measurement standards vary.

• Corporate annual reports and other public disclosures rarely detail specific AI investments, and

companies may choose not to disclose the size and scope of their AI-related investments for

competitive reasons.

• Privately owned companies may not be subject to public disclosures.

• Overlap exists in many cases between different AI-enabled applications, e.g., “Internet of Things”

(IoT) devices and intelligent virtual assistants, or robots and autonomous systems.

• Not all estimates of AI-related developments include government investment.

• Not all estimates separate internal and external investment in AI-related technologies or applications.

At the same time, it is possible to anticipate trend lines based on planned spending in AI and related

technology areas. Accenture polled corporate aerospace and defense executives on this topic. While specific

planned investment levels in each area are unavailable, it is reasonable to expect that those polled plan to

spend in each of the seven technical areas identified over the next few years. It is also possible to infer that

they collectively see somewhat greater near-term return prospects for investments in, for example, computer

vision or deep learning than embedded AI solutions or video analytics. Finally, it may be important to

recognize (in Sherlock Holmes style) the dogs not barking, such as swarming technology.

Taken together, the crystal ball is cloudy—both for the pace and scope of technology development and

adoption. But while market prognostication error bars are significant, AI’s promise appears—at long last—

to be finding utility across varied market sectors. In the virtuous cycle that appears to be developing,

accelerating progress in multiple AI technology areas should lead to development of new and growing

markets, attracting new AI-related products and services, and building a vibrant AI ecosystem.

Page 42: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 42

UNCLASSIFIED

This appendix draws upon the following sources:

Defense Science Board | Report of the Defense Science Board Summer Study on Autonomy | June 2016 |

https://www.hsdl.org/?view&did=794641 | p. 11.

Accenture | Accelerating Through Turbulence: Technology Vision for Aerospace and Defense 2017 |

June 2017 | https://www.accenture.com/t00010101T000000__w__/gb-en/_acnmedia/PDF-

53/Accenture-Accelerating-Through-Digital-Turbulence.pdf | p. 4.

J. Bughin, et al. | Artificial Intelligence: The Next Digital Frontier? | McKinsey Global Institute

discussion paper | June 2017 | http://www.mckinsey.com/business-functions/mckinsey-analytics/our-

insights/how-artificial-intelligence-can-deliver-real-value-to-companies | pp. 2, 6, 10.

R. Curran, B. Purcell | TechRadar: Artificial Intelligence Technologies, Q1 2017 | 18 January 2017 |

https://kloudrydermcaasicmforrester.s3.amazonaws.com/mcaas/Reprints/RES129161.pdf | p. 9.

Page 43: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 43

UNCLASSIFIED

Appendix B: Deep Learning and Deep Reinforcement Learning Fuel the

Latest AI Breakthroughs

Over the past several years, some narrow AI technologies have achieved (whether individually or in

combination) what are referred to as “superhuman” results—output exceeds human-level performance for

a given task. For example, computers have for the first time exceeded human expert performance in areas

such as object detection in imagery, speech recognition, and video and board game wins.

The technology most responsible for recent successes in AI is “deep learning” and “deep reinforcement

learning.” In this appendix, we focus on these approaches.

Deep learning and deep reinforcement learning fuel “superhuman” AI performance

Deep learning refers to the latest generation of artificial neural networks (or neural nets). Deep

reinforcement learning refers to the use of deep neural networks to implement the models used in

reinforcement learning that infer the state of the world from input percepts or infer the best action to take

based on the state of the world.

Simplistically, neural nets are universal function approximators that map input percepts 𝑋, where 𝑋 is a

vector of measurements, and output 𝑌, another vector of numbers. For example, in the case of object

recognition in images, 𝑋 are pixel values and 𝑌 are posterior probabilities of objects present in the image.

Neural networks learn these mappings via an optimization procedure that finds the set of network

parameters that minimizes misclassification and other costs over a training set of 𝑋, 𝑌 pairs, where 𝑋s are

measurements of samples and 𝑌s are corresponding labels often provided by human annotation.

Unlike previous generations of neural nets, deep learning networks often consist of many more layers of

neurons (hence, “deep”). These deeper-network topologies have been shown to more efficiently model

complex patterns with exponentially fewer neurons. Until recently, training deeper networks was hard to

do effectively and often resulted in poorly trained networks that were outperformed by shallow networks.

Breakthroughs in optimization algorithms and new neural network activation functions now make it

possible to train networks with many more layers of neurons.

What is especially impressive about deep learning networks compared to previous generations of neural

networks is their breakthrough performance on tasks that were once extremely difficult for computers. In

some cases, deep learning systems surpass the performance of human experts. For example, with respect to

image classification, the AI is given an image and asked to classify it as being an image of one of several

categories—for example, a “cat,” “car,” or “airplane.” The ILSVRC is the standard benchmark for

measuring image classification performance.222 Error! Reference source not found.B1 shows several

example ILSVRC images and predicted categories from Microsoft’s RESNET deep learning system, the

best-performing network in 2016.223

Page 44: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 44

UNCLASSIFIED

Figure B1. “Deep

Residual Learning for

Image Recognition.”

Sample ImageNet

images correctly

classified by the deep

learning network. For

each image, the

ground-truth label and

the top-5 labels

predicted by their

network are listed.

Figure B2. Deep

learning systems

have made

tremendous progress

over the last five

years in image

classification

performance on the

ILSVRC. The best-

available systems

today make roughly

half as many

classification errors as

a human expert.

B2 plots the ILSVRC top-5 error rate of the best AI systems since 2010.224,225,226,227 The dramatic reduction

in error rate beginning in 2012 comes from the development of sophisticated deep learning networks trained

on increasingly larger image datasets. Image classification on ILSVRC is now generally considered to be a

“solved” problem, but it is worth noting that ILSVRC is a narrow-image classification task limited to object

Page 45: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 45

UNCLASSIFIED

classification into 1,000 image categories from relatively well-taken images (for example, centered, with

good lighting, and/or uncluttered).

Figure B1. “Deep

Residual Learning for

Image Recognition.”

Sample ImageNet

images correctly

classified by the deep

learning network. For

each image, the

ground-truth label and

the top-5 labels

predicted by their

network are listed.

Figure B2. Deep

learning systems

have made

tremendous progress

over the last five

years in image

classification

performance on the

ILSVRC. The best-

available systems

today make roughly

half as many

classification errors as

a human expert.

Page 46: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 46

UNCLASSIFIED

The success of deep learning on ILSVRC has propagated to other image recognition tasks. Winning

ILSVRC networks are frequently open-sourced and made available to the research community as a common

good. This encourages other researchers to reuse high-quality deep learning networks. In fact, many

researchers have achieved high performance on their own image recognition applications by reusing or

transferring ILSVRC networks and further training these networks on the targeted application data. One

example of this type of “transfer learning” is in the application of detecting and counting vehicles from

overhead imagery.228 Although these deep learning networks were originally trained on ground-level

ImageNet images, after fine-tuning to overhead imagery these networks attained very high vehicle detection

and discrete counting performance.229

Another challenging task where deep learning has now surpassed human performance is in the recognition

of conversational telephone speech. For this task, given telephone audio of people speaking, the AI is asked

to transcribe the conversation. The Switchboard benchmark is the standard large-scale conversational

telephone speech dataset for measuring speech recognition performance (

Figure B3. In 2017,

deep learning

systems have now

surpassed human

expert performance

on the challenging

task of transcribing

conversational

telephone speech,

as measured by

word error rate on

the Switchboard

benchmark. The best

system today makes

13.5 percent fewer

errors than human

experts.

B3).230 In the early 2000s, the prevailing opinion of researchers in speech recognition was that

conversational telephone speech was so difficult it was not clear whether a computer could outperform

humans. Nevertheless, advances in deep learning and the availability of thousands of hours of labeled

training data has made it possible to achieve a 5.1 percent word error rate on Switchboard—13.5 percent

better than human expert performance.231

Page 47: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 47

UNCLASSIFIED

Figure B3. In 2017,

deep learning

systems have now

surpassed human

expert performance

on the challenging

task of transcribing

conversational

telephone speech,

as measured by

word error rate on

the Switchboard

benchmark. The best

system today makes

13.5 percent fewer

errors than human

experts.

Superhuman performance is not guaranteed with deep learning or deep reinforcement learning

Not enough labeled training data

Large amounts of labeled training data are not always possible to obtain. Crowd-sourcing is the standard

approach to obtaining vast quantities of labels. By paying small amounts of money to large numbers of

human annotators, companies can quickly assemble large labeled training sets to train their deep learning

and deep reinforcement learning systems. Crowd-sourcing also mitigates the potential problems associated

with training networks on mislabeled data. When multiple people annotate the same datum and when many

instances of related data are labeled, it becomes less likely that labeling errors will systematically dominate

a training set and lead the network training astray.

Unfortunately, not all applications can be crowd-sourced. If the sophistication or expertise required to label

a dataset is high, as in the case of medical imagery, simulation outputs, and other scientific data, fewer

people will be qualified to annotate the data. In other applications, preserving the confidentiality of data is

paramount, so again fewer human eyeballs are available for labeling the data. In these cases, deep learning

and deep reinforcement learning will not be able to learn from large amounts of labeled data to achieve

superhuman performance.

Not enough gameplay training data

A key ingredient for deep reinforcement learning success is the availability of substantial amounts of

gameplay data. Unlike the standard classification or regression problem, reinforcement learning does not

always provide a reward signal (that is, a label) per input measurement. For example, when playing Go, the

only accurate reward signal is ultimately a victory or a loss; but this reward is not realized until each player

has made multiple moves resulting in a sequence of board configurations. The task of the deep neural networks

is to estimate the efficacy of all actions to take given the state of the system. To do so, training data in the

form of board configurations and best action pairs are needed. The amount of data required grows as the

number of possible board configurations and possible actions increase. Even in relatively simple games, the

combinations of possible states and possible actions easily exceed millions, and the training data required for

superior performance can exceed tens or even hundreds of millions of examples.

Page 48: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 48

UNCLASSIFIED

This much training data can be acquired in two ways. First, collecting gameplay data from actual games

played by humans can result in high-quality training if human performance levels are desired. Games that

are played online are excellent sources of such data. Other sources of gameplay data can be collected by

tracking what humans do in certain world-states. A fitting example of this comes from the car companies

actively developing autonomous vehicles. Tesla, for example, collects millions of miles of human driving

data each month from all its cars on the road. Using such data, companies can train neural networks to take

actions that mimic what humans would do in any situation. The challenge of this approach is that obtaining

training data for every possible situation is very difficult because the possible world-states could be quite

large and the sequence of actions necessary to observe such a world-state may be exceedingly rare. What

training data might be relevant to a proverbial “black swan” event?

The second way to collect data is to watch the computer play the game itself many times. This approach

has two advantages over the collection of human gameplay data. Because the computer can play the game

an almost infinite number of times, rare states and black swan events will eventually materialize, and

effective actions will be observed in these situations. Additionally, the computer often tries actions that

may be counterintuitive for human players. In several of the matches between AlphaGo and the human

experts, AlphaGo made moves that surprised its human competitors and resulted in victory. With this type

of data collection, it is possible to not only match human performance levels, but significantly exceed them.

In the context of defense applications, such performance could be devastating. The main challenge in

realizing this data collection approach is building a realistic gameplay engine. For board games such as

Chess and Go, this is a solved problem, but for autonomous driving and other real-world “games,” a game

simulation may not adequately represent all the possible situations that arise in the real world. Ultimately,

effective use of AI requires humans to trust algorithmic responses in desired application areas; adoption

practices at scale will vary accordingly.

Page 49: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 49

UNCLASSIFIED

Not enough computational horsepower

To achieve superhuman results, large training data and commensurately large neural network models are

required. Training such neural networks in reasonable timeframes requires substantial computational

horsepower. It is not uncommon for large-image recognition networks to take several weeks to train on a

single GPU. Such long turnaround times are unacceptable for developing new neural networks. Developers

must explore the massive neural network design space of different sizes, shapes, connectivity, and

optimization parameters to achieve good results. Reducing experimental turnaround time is paramount, and

computation horsepower is the key enabler.

In deep reinforcement learning, additional computation resources are required for game-world simulations.

This is not a meaningful issue for simple games, but for complex real-world simulations, high-performance

computing is required. National Laboratories charged with ensuring the efficacy of the nuclear stockpile

routinely simulate the real-world physics of particle interactions using the world’s largest supercomputers.

While not all applications require such high-end capabilities, to the extent that potential applications of

reinforcement learning require such high-fidelity simulations, commensurate computational capabilities

will also be required.

Deep reinforcement learning faces additional challenges prior to wide-scale adoption

Deep reinforcement learning faces all the challenges of deep learning, and a few others. First, deep

reinforcement learning models take longer to train due to the higher complexity of the models. Since the

reward signals are not as direct and readily available as in the classification and regression cases, models

tend to take more training data and more training iterations to achieve satisfactory performance. This may

result in development and application cycles that are simply too long to be effective. In the worst case, they

may be so slow that by the time one can train a competent AI, the underlying game has materially changed;

the world-state evolved more quickly than the AI’s sense-making abilities.

Second, obtaining the kind of training data described above can be difficult or costly. On the one hand,

there may not be adequate systems in place to collect, store, and sort vast quantities of human gameplay

data. On the other, a relevant game may not be popular enough or have enough humans playing the game,

which suggests a data shortfall. For applications that require a great deal of realism, developing a faithful

game environment may prove impossible; some level of abstraction or simulation may be required. In such

cases, the lack of training data severely hinders the chance of successfully developing deep reinforcement

learning systems with acceptable performance quality.

Watch this space: AI conquering StarCraftII will herald further deep reinforcement learning success

In August 2017, Google and Activision established the StarCraftII AI challenge.232 In this context, the

former is providing algorithm building blocks for deep reinforcement learning while the latter is releasing

an open-source version of StarCraftII and training data comprising millions of actual online matches

between human players. In the same way that the ILSVRC paved the way toward superhuman AI vision

performance, the StarCraftII challenge has the potential to accelerate reinforcement learning research. It is

difficult to predict if or when the community will develop an AI with superhuman StarCraft performance,

but success will carry significant implications for AI applications to defense missions. Among other things,

StarCraft may provide a useful abstraction for the logistical planning and management tasks associated with

some battlefield contingencies.

Page 50: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 50

UNCLASSIFIED

Appendix C: Candidate Near-Term Implementation Measures to Strengthen

DOD’s AI Posture

Table 2 on page 33 outlines a five-part, decade-scale approach to strengthening DOD’s AI posture. In rough

order of difficulty:

1. Acquiring or adapting existing commercial AI products and services to streamline DOD’s

extensive “back office” functions.

2. Acquiring or adapting existing commercial AI products and services to improve select combat

support and combat service support functions.

3. Adapting or conducting limited development of existing AI tools for specific defense missions.

4. Developing and deploying new or improved AI products to enhance fielded and developmental

combat systems.

5. Designing, developing, prototyping, and experimenting with new or prospective AI

technologies in operational context to develop new combat capabilities.

To put such an approach in play, table C1 provides a menu of possible near-term implementation steps that

one or more DOD components could undertake. The specific tasks identified gravitate more toward the

low-hanging fruit than to a comprehensive menu of options, and the option set could be revisited as DOD

components make further headway. The offices of primary responsibility (OPRs) listed below are in many

cases not the sole DOD stakeholder, but rather could serve as a reasonable entry point for broader

department-wide consideration.

Table C1. Candidate near-term implementation measures. (cont.)

Area Illustrative task Suggested OPR

1. Acquiring or adapting existing commercial AI products and services to streamline DOD’s extensive “back office”

functions.

• Evaluate the extent to which existing commercial products could help streamline key DOD cost centers (e.g., human resources, medical records management, travel).

• Explore the cost and feasibility of contracted, service-based AI options.

• Undertake pilot activity, whether product- or service-based, to streamline operations in a single cost center.

• Augment DOD’s secure and reliable access to cloud-related infrastructure or other relevant high-performance computing architectures.

• Use other transaction authority or develop other contractual mechanisms that enable routine (~quarterly) update, refresh, or spiral acquisition cycles.

CMO or USD/P&R, with CIO, DDS, and Service business operations organizations.

2. Acquiring or adapting existing commercial AI products and services to improve select combat support and combat service support functions.

• Build toward force-wide deployment of successfully demonstrated ground-based AI-infused leader/follower autonomous systems (e.g., logistics or medevac).

• Accelerate development of manned/unmanned air and naval leader/follower capabilities.

• Leverage commercial products to enhance supply chain management and operational supply/resupply efficiencies.

USD/A&S, with TRANSCOM, DIU-x, DLA, and Service acquisition and logistics

components.

Page 51: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 51

UNCLASSIFIED

Table C1. Candidate near-term implementation measures. (cont.)

Area Illustrative task Suggested OPR

3. Adapting or conducting limited development of existing AI tools for specific defense missions.

• Build on Project Maven, extending ISR-related algorithm developments to other operational theaters, to additional warfighting challenges, and to other platforms.

• Conduct pilot activity using AI techniques to enhance the nation’s space situational awareness posture.

• Conduct pilot activity using AI techniques to enhance the nation’s ability to locate and identify deployed foreign undersea vehicles.

• Conduct pilot activity using AI techniques to enhance the nation’s ability to rapidly locate and identify foreign mobile missile systems.

USDI and USD/A&S, with USD/R&E, NGA, and Service rapid development offices; operational elements in

their assigned areas.

4. Developing and deploying new or improved AI products to enhance fielded and developmental combat systems.

• Increase the quantity and availability of labeled training data in priority areas (e.g., electronic or cyber warfare).

• Explore Bayesian or other AI techniques to enhance AI’s ability to effectively navigate data-poor contexts.

• Develop test and evaluation methodologies appropriate for systems designed to operate in “big data” environments and suitable for rapid technology refresh rates.

USD/R&E, with USD/A&S and Service research, development, and acquisition components.

5. Designing, developing, prototyping, and experimenting with new or prospective AI technologies in operational context to develop new combat

capabilities.

• Increase emphasis on AI-related experimentation and wargaming to develop new operating concepts and novel technical approaches.

• Leverage commercial AI technology developments where possible (e.g., image classification, speech/language processing); focus DOD resources where industry does not emphasize (e.g., swarming technologies).

• Accelerate DOD prototyping of AI-infused systems to capitalize on rapid commercial technology development (~60-90-day technology refresh rate).

• Conduct a decade-scale U.S.-Chinese-Russian AI net evaluation.

USD/P, VCJCS, and USD/R&E, with Service research, doctrine development, and operational experimentation elements.

Page 52: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 52

UNCLASSIFIED

References

1. A.F. Krepinevich, Jr., Center for Strategic and Budgetary Assessments | The Military-Technical Revolution: A Preliminary Assessment | 2002 | http://csbaonline.org/uploads/documents/2002.10.02-Military-Technical-Revolution.pdf | p. 3.

2. A.F. Krepinevich, Jr., Center for Strategic and Budgetary Assessments | The Military-Technical Revolution: A Preliminary Assessment | 2002 | http://csbaonline.org/uploads/documents/2002.10.02-Military-Technical-Revolution.pdf | p. 3.

3. F.P.B. Osinga | Science, Strategy and War: The Strategic Theory of John Boyd | Abingdon, U.K.: Routledge | 2006.

4. P. Scharre and M.C. Horowitz, Center for a New American Security | An Introduction to Autonomy in Weapon Systems | Working paper | February 2015 | https://s3.amazonaws.com/files.cnas.org/ documents/Ethical-Autonomy-Working-Paper_021015_v02.pdf | pp. 3, 21–23.

5. H.M. Roth and R. Moyes, Future of Life Institute | Lethal Autonomous Weapons, Artificial Intelligence and Meaningful Human Control | Project briefing | https://futureoflife.org/wp-content/uploads/2017/01/Heather-Roff.pdf?x33688 | pp. 10–13.

6. Deputy Secretary of Defense memorandum | Establishment of an Algorithmic Warfare Cross-Functional Team (Project Maven) | 26 April 2017 | https://www.govexec.com/media/gbc/docs/ pdfs_edit/establishment_of_the_awcft_project_maven.pdf.

7. S.J. Freedberg, Jr. | Artificial Intelligence Will Help Hunt Daesh By December | Breaking Defense | 13 July 2017 | http://breakingdefense.com/2017/07/artificial-intelligence-will-help-hunt-daesh-by-december.

8. Deputy Secretary of Defense | Speech delivered to the Center for a New American Security Defense Forum, Washington, D.C. | 14 December 2015 | https://www.defense.gov/News/Speeches/Speech-View/Article/634214/cnas-defense-forum.

9. IDC press release | Worldwide Spending on Cognitive and Artificial Intelligence Systems Forecast to Reach $12.5 Billion This Year, According to New IDC Spending Guide | 3 April 2017 | http://www.idc.com/getdoc.jsp?containerId=prUS42439617.

10. Magazine article | Gil Press | Forrester Predicts Investment in Artificial Intelligence Will Grow 300% in 2017 | Forbes | 1 November 2016 | https://www.forbes.com/sites/gilpress/2016/11/01/ forrester-predicts-investment-in-artificial-intelligence-will-grow-300-in-2017/#5a14a5c35509.

11. B. Ma, S. Nahal, and F. Tran | Thematic Investing: Robot Revolution—Global Robot and AI Primer | Bank of America Merrill Lynch | 3 November 2015 | https://www.bofaml.com/content/dam/ boamlimages/documents/PDFs/robotics_and_ai_condensed_primer.pdf.

12. J. Manyika, et al., McKinsey Global Institute | Disruptive Technologies: Advances that will transform life, business, and the global economy | May 2013 | http://www.mckinsey.com/business-functions/digital-mckinsey/our-insights/disruptive-technologies| pp. 4–5, 40–85.

13. World Economic Forum | The Global Risks Report 2017, 12th ed. | http://www3.weforum.org/ docs/GRR17_Report_web.pdf.

14. L.T. Wood | Russia to develop missile featuring artificial intelligence | Washington Times | 21 July 2017 | http://www.washingtontimes.com/news/2017/jul/21/russia-develop-missile-featuring-artificial-intell.

15. P. Tucker | Russian Weapons Maker to Build AI-Directed Guns | Defense One | 14 July 2017 | http://www.defenseone.com/technology/2017/07/russian-weapons-maker-build-ai-guns/139452.

16. V. Kashin and M. Raska | Countering the U.S. Third Offset Strategy: Russian Perspectives, Responses and Challenges | RSiS Policy Report | January 2017 | https://www.rsis.edu.sg/wp-content/uploads/2017/01/PR170124_Countering-the-U.S.-Third-Offset-Strategy.pdf.

Page 53: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 53

UNCLASSIFIED

17. E.B. Kania | Chinese Advances in Unmanned Systems and the Military Applications of Artificial Intelligence—the PLA’s Trajectory Toward Unmanned, ‘Intelligentized’ Warfare | Testimony before the U.S.-China Economic and Security Review Commission | 23 February 2017 | https://www.uscc.gov/sites/default/files/Kania_Testimony.pdf.

18. Jiang Jie | China Expected to Overtake West In Future Air Operations With Big Data, AI: Expert | People’s Daily Online | 3 July 2017 | Reprinted at http://www.defense-aerospace.com/articles-view/release/3/184989/china-expects-to-overtake-west-in-future-air-operations-with-big-data%2C-ai.html.

19. J. McDonald | China announces goal of AI leadership by 2030 | Washington Post | 21 July 2017 | https://www.washingtonpost.com/world/asia_pacific/china-announces-goal-of-ai-leadership-by-2030/2017/07/21/c9c98984-6dd0-11e7-abbc-a53480672286_story.html?utm_term=.e10355152670.

20. S. Pham | China wants to build a $150 billion AI industry | CNN | 21 July 2017 | http://money.cnn.com/ 2017/07/21/technology/china-artificial-intelligence-future/index.html.

21. Pan Yue | China Sets Up National Laboratory to Develop Brain-Like Artificial Intelligence | 15 May 2017 | https://www.chinamoneynetwork.com/2017/05/15/china-sets-up-national-laboratory-to-develop-brain-like-artificial-intelligence.

22. S. Russell and P. Norvig | Artificial Intelligence: A Modern Approach, 3rd ed. | 2010 | http://web.cecs.pdx.edu/~mperkows/CLASS_479/2017_ZZ_00/02__GOOD_Russel=Norvig=Artificial%20Intelligence%20A%20Modern%20Approach%20(3rd%20Edition).pdf) | p. viii.

23. Govini | Department of Defense: Artificial Intelligence, Big Data and Cloud Taxonomy | December 2017 | http://www.govini.com/research-item/dod-artificial-intelligence-and-big-data-taxonomy/ | p. 1.

24. J. Bughin, et al. | Artificial Intelligence: The Next Digital Frontier? | McKinsey Global Institute discussion paper | June 2017 | http://www.mckinsey.com/business-functions/mckinsey-analytics/our-insights/how-artificial-intelligence-can-deliver-real-value-to-companies | p. 10.

25. J. Bughin, et al. | Artificial Intelligence: The Next Digital Frontier? | McKinsey Global Institute discussion paper | June 2017 | http://www.mckinsey.com/business-functions/mckinsey-analytics/our-insights/how-artificial-intelligence-can-deliver-real-value-to-companies | p. 5.

26. J. Bughin, et al. | Artificial Intelligence: The Next Digital Frontier? | McKinsey Global Institute discussion paper | June 2017 | http://www.mckinsey.com/business-functions/mckinsey-analytics/our-insights/how-artificial-intelligence-can-deliver-real-value-to-companies | p. 12.

27. CB Insights | A ranking of the 100 most promising private artificial intelligence companies in the world | https://www.cbinsights.com/research-ai-100.

28. Shivon Zilis | The Current State of Machine Intelligence 3.0 | www.shivonsilis.com.

29. R. Curran, B. Purcell | TechRadar: Artificial Intelligence Technologies, Q1 2017 | 18 January 2017 | https://kloudrydermcaasicmforrester.s3.amazonaws.com/mcaas/Reprints/RES129161.pdf | p. 9.

30. Magazine article | Gil Press | Top 10 Artificial Intelligence Technologies | Forbes | 23 January 2017 | https://www.forbes.com/sites/gilpress/2017/01/23/top-10-hot-artificial-intelligence-ai-technologies/#2d70fb0c1928.

31. A. Ilachinksi, Center for Naval Analyses | AI, Robots, and Swarms: Issues, Questions, and Recommended Studies | 2017 | https://www.cna.org/CNA_files/PDF/DRM-2017-U-014796-Final.pdf | p. 3.

32. M. Weisgerber | The Pentagon’s New Algorithmic Warfare Task Force Gets Its First Mission: Hunt ISIS | Defense One | 14 May 2017 | http://www.defenseone.com/technology/2017/05/ pentagons-new-algorithmic-warfare-cell-gets-its-first-mission-hunt-isis/137833/.

33. S.J. Freedberg, Jr. | Artificial Intelligence Will Help Hunt Daesh By December | Breaking Defense | 13 July 2017 | http://breakingdefense.com/2017/07/artificial-intelligence-will-help-hunt-daesh-by-december.

Page 54: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 54

UNCLASSIFIED

34. P. Tucker | The Next Big War Will Turn on AI, Says US Secret-Weapons Czar | Defense One | 28 March 2017 | http://www.defenseone.com/technology/2017/03/next-big-war-will-turn-ai-says-pentagons-secret-weapons-czar/136537.

35. Jie Hu, Li Shen, and Gang Sun | Squeeze-and-Excitation Networks | arXiv | 5 September 2017 | https://arxiv.org/abs/1709.01507.

36. W. Xiong, et al. | The Microsoft 2017 Conversational Speech Recognition System | Microsoft Research Technical Report MSR-TR-2017-39 | 2017 | https://www.microsoft.com/en-us/research/wp-content/uploads/2017/08/ms_swbd17-2.pdf.

37. V. Mnih, et al. | Human-level control through deep reinforcement learning | Nature | Vol. 518 | 26 February 2015 | pp. 529–533.

38. The Future of Go Summit | 23–27 May, Wuzhen, China | http://events.google.com/alphago2017.

39. David Silver, et al. | Mastering the game of Go with deep neural networks and tree search | Nature | Vol. 529 | January 28, 2016 | pp. 484–489.

40. T. Hastie, R. Tibshirani, and J. Friedman | The Elements of Statistical Learning | 2d ed. | 2009 | https://web.stanford.edu/~hastie/Papers/ESLII.pdf.

41. D. Hull | The Tesla Advantage: 1.3 Million Miles of Data | Bloomberg | 20 December 2016 | https://www.bloomberg.com/news/articles/2016-12-20/the-tesla-advantage-1-3-billion-miles-of-data.

42. O. Russakovsky, et al. | ImageNet Large-Scale Visual Recognition Challenge | arXiv | 1 December 2014 | https://arxiv.org/abs/1409.0575.

43. P. Goyal, et al. | Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour | arXiv | 8 June 2017 | https://arxiv.org/abs/1706.02677.

44. B. Van Essen, et al., LLNL | LBANN: Livermore Big Artificial Neural Network HPC Toolkit | LLNL-CONF-677443 | 23 September 2015.

45. Caffe | http://caffe.berkeleyvision.org/; Torch, http://torch.ch/; Theano, http://deeplearning.net/software/theano/; Keras, https://keras.io/; TensorFlow, https://www.tensorflow.org/; and MXNet, http://mxnet.incubator.apache.org/.

46. P. Baldi | Autoencoders, Unsupervised Learning, and Deep Architectures | Journal of Machine Learning Research | Vol. 27 | 2012 | http://proceedings.mlr.press/v27/baldi12a/baldi12a.pdf | pp. 37–50.

47. C. Doersch, A. Gupta, and A.A. Efros | Unsupervised Visual Representation Learning by Context Prediction | arXiv | 16 January 2016 | https://arxiv.org/abs/1505.05192.

48. Attributed to M. Francis | Unmanned air systems: challenge and opportunity | Journal of Aircraft | Vol. 49, No. 6 | Nov–Dec 2012. Reproduced in A. Ilachinksi, Center for Naval Analyses | AI, Robots, and Swarms: Issues, Questions, and Recommended Studies | 2017 | https://www.cna.org/CNA_files/PDF/DRM-2017-U-014796-Final.pdf.

49. V. Mnih, et al. | Human-level control through deep reinforcement learning | Nature | Vol. 518 | 26 February 2015 | pp. 529–533.

50. V. Mnih, et al. | Human-level control through deep reinforcement learning | Nature | Vol. 518 | 26 February 2015 | pp. 529–533.

51. David Silver, et al. | Mastering the game of Go with deep neural networks and tree search | Nature | Vol. 529 | January 28, 2016 | pp. 484–489.

52. The Future of Go Summit | 23–27 May, Wuzhen, China | http://events.google.com/alphago2017.

53. YouTube | SSCAIT 2016: Man vs Machine Matches | Published 25 March 2016 | https://www.youtube.com/watch?v=ztNYOnx_YQo.

54. I.J. Goodfellow, J. Shlens, and C. Szegedy | Explaining and Harnessing Adversarial Examples | arXiv | 20 March 2015 | https://arxiv.org/abs/1412.6572.

Page 55: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 55

UNCLASSIFIED

55. N. Papernot, et al. | Practical Black-Box Attacks against Machine Learning | arXiv | 19 March 2017 | https://arxiv.org/abs/1602.02697

56. S.-M. Moosavi-Dezfooli, et al. | Universal adversarial perturbations | arXiv | 9 March 2017 | https://arxiv.org/abs/1610.08401.

57. N. Papernot, et al. | Distillation as a Defense to Adversarial Perturbations against Deep Neural Networks | arXiv | 14 March 2016 | https://arxiv.org/abs/1511.04508.

58. M.N. Gibbs | Bayesian Gaussian Processes for Classification and Regression | Doctoral dissertation, University of Cambridge | 1997 | http://citeseerx.ist.psu.edu/viewdoc/download?doi= 10.1.1.147.1130&rep=rep1&type=pdf); Radford M. Neal, Bayesian Learning for Neural Networks (Springer Science & Business Media, 2012).

59. C.K.I. Williams and C.E. Rasmussen | Gaussian processes for regression | Proceedings of the 8th International Conference on Neural Information Processing Systems | Cambridge: MIT Press | 1995 | pp. 514–520.

60. F. Anselmi, et al. | Deep Convolutional Networks are Hierarchical Kernel Machines | ArXiv | August 2015 | https://arxiv.org/abs/1508.01084.

61. A. Daniely, R. Frostig, and Y. Singer | Toward Deeper Understanding of Neural Networks: The Power of Initialization and a Dual View on Expressivity | ArXiv | February 2016 | https://arxiv.org/abs/1602.05897.

62. C.A. Micchelli, Yueshang Xu, and Haizhang Zhang | Universal kernels | Journal of Machine Learning Research | Vol. 7 | December 2006 | pp. 2651–2667| http://jmlr.csail.mit.edu/papers/volume7/ micchelli06a/micchelli06a.pdf.

63. Jaehoon Lee, et al. | Deep Neural Networks as Gaussian Processes | ArXiv | October 2017 | https://arxiv.org/abs/1711.00165.

64. Jaehoon Lee, et al. | Deep Neural Networks as Gaussian Processes | ArXiv | October 2017 | https://arxiv.org/abs/1711.00165.

65. K. Grosse, et al. | How Wrong Am I? Studying Adversarial Examples and their Impact on Uncertainty in Gaussian Process Machine Learning Models | ArXiv | November 2017 | https://arxiv.org/abs/1711.06598.

66. Y. Lecun, Y. Bengio, and G. Hinton | Deep learning | Nature | Vol. 521 | May 2015 | pp. 436–444 | https://www.nature.com/articles/nature14539.

67. Govini | Department of Defense: Artificial Intelligence, Big Data and Cloud Taxonomy | December 2017 | http://www.govini.com/research-item/dod-artificial-intelligence-and-big-data-taxonomy/ | p. 12.

68. E.B. Kania, Center for a New American Security | Battlefield Singularity: Artificial Intelligence, Military Revolution, and China’s Future Military Power | November 2017 | https://s3.amazonaws.com/files.cnas.org/documents/Battlefield-Singularity-November-2017.pdf?mtime=20171129235804.

69. A.J. Tellis, et al. | Measuring National Power in the Postindustrial Age | RAND | 2000 | https://www.rand.org/pubs/monograph_reports/MR1110.html | p. 137.

70. E.B. Kania, Center for a New American Security | Battlefield Singularity: Artificial Intelligence, Military Revolution, and China’s Future Military Power | November 2017 | https://s3.amazonaws.com/files.cnas.org/documents/Battlefield-Singularity-November-2017.pdf?mtime=20171129235804.

71. C. Prine | Robots poised to take over wide range of military jobs | San Diego Union-Tribune | 20 February 2017 | http://www.sandiegouniontribune.com/military/sd-me-robots-military-20170130-story.html.

Page 56: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 56

UNCLASSIFIED

72. J. Golson | The Army wants to use this giant drone to resupply soldiers | The Verge | 19 January | https://www.theverge.com/2017/1/18/14311912/hoverbike-drone-us-army-jtarv-resupply-autonomous.

73. R.O. Work and S. Brimley, Center for a New American Security | 20YY: Preparing for War in the Robotic Age | January 2014 | p. 25.

74. ONR on Track to Demo Autonomous Cargo Resupply | Aviation Week | http://aviationweek.com/awin/onr-track-demo-autonomous-cargo-resupply.

75. U.S. Army Capabilities Integration Center | Robotic and Autonomous Systems Strategy | March 2017 | http://www.arcic.army.mil/App_Documents/RAS_Strategy.pdf | pp. 5–6.

76. Lt. Gen. M.H. Stevenson | A Vision of Army Logistics with 20/20 Hindsight | Army Sustainment | March–April 2011 | http://www.alu.army.mil/alog/issues/marapr11/2020_hindsight.html.

77. U.S. Army Capabilities Integration Center | Robotic and Autonomous Systems Strategy | March 2017 | http://www.arcic.army.mil/App_Documents/RAS_Strategy.pdf | p. 6.

78. V. Boulanin and M. Verbruggen, Stockholm International Peace Research Institute | Mapping the Development of Autonomy in Weapon Systems | November 2017 | p. 107.

79. ONR on Track to Demo Autonomous Cargo Resupply | Aviation Week | http://aviationweek.com/awin/onr-track-demo-autonomous-cargo-resupply.

80. B. Jensen and R. Kendall | Waze for War: How the Army Can Integrate Artificial Intelligence | WarOnTheRocks | 2 September 2016 | https://warontherocks.com/2016/09/waze-for-war-how-the-army-can-integrate-artificial-intelligence.

81. Govini | Department of Defense: Artificial Intelligence, Big Data and Cloud Taxonomy | December 2017 | http://www.govini.com/research-item/dod-artificial-intelligence-and-big-data-taxonomy/ | p. 12.

82. Top 5 Vendors in the Intelligent Video Analytics Market from 2017 to 2021: Technavio | Business Wire | 8 August 2017 | http://www.businesswire.com/news/home/20170808005693/en/.

83. E.B. Kania, Center for a New American Security | Battlefield Singularity: Artificial Intelligence, Military Revolution, and China’s Future Military Power | November 2017 | https://s3.amazonaws.com/files.cnas.org/documents/Battlefield-Singularity-November-2017.pdf?mtime=20171129235804.

84. B. Rosenberg | Army grapples with sensor overload | Defense Systems | 7 April 2010 | https://defensesystems.com/articles/2010/04/06/cover-story-sensor-overload.asp.

85. E.B. Kania, Center for a New American Security | Battlefield Singularity: Artificial Intelligence, Military Revolution, and China’s Future Military Power | November 2017 | https://s3.amazonaws.com/files.cnas.org/documents/Battlefield-Singularity-November-2017.pdf?mtime=20171129235804.

86. The Future of Military Tech | Defense One | June 2017 | http://www.defenseone.com/assets/future-military-tech/portal/ | p. 3.

87. J. Doubleday | DOD deploys first artificial intelligence algorithms to process surveillance feeds in Middle East | Inside Defense | 18 December 2017 | https://insidedefense.com/daily-news/dod-deploys-first-artificial-intelligence-algorithms-process-surveillance-feeds-middle.

88. J. Somers | Is AI Riding a One-Trick Pony? | MIT Technology Review | 29 September 2017 | https://www.technologyreview.com/s/608911/is-ai-riding-a-one-trick-pony.

89. World Economic Forum | The Global Risks Report 2017, 12th ed. | http://www3.weforum.org/ docs/GRR17_Report_web.pdf | p. 48.

Page 57: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 57

UNCLASSIFIED

90. G. Allen and T. Chan, Belfer Center for Science and International Affairs | Artificial Intelligence and National Security | July 2017 | https://www.belfercenter.org/sites/default/files/files/publication/ AI%20NatSec%20-%20final.pdf | p. 31.

91. M. Bickert and B. Fishman | Hard Questions: How We Counter Terrorism | Facebook Newsroom | 15 June 2017 | https://newsroom.fb.com/news/2017/06/how-we-counter-terrorism.

92. V. Boulanin and M. Verbruggen, Stockholm International Peace Research Institute | Mapping the Development of Autonomy in Weapon Systems | November 2017 | pp. 27–28.

93. V. Boulanin and M. Verbruggen, Stockholm International Peace Research Institute | Mapping the Development of Autonomy in Weapon Systems | November 2017 | pp. 27–28.

94. Govini | Department of Defense: Artificial Intelligence, Big Data and Cloud Taxonomy | December 2017 | http://www.govini.com/research-item/dod-artificial-intelligence-and-big-data-taxonomy/ | p. 12.

95. E.B. Kania, Center for a New American Security | Battlefield Singularity: Artificial Intelligence, Military Revolution, and China’s Future Military Power | November 2017 | https://s3.amazonaws.com/files.cnas.org/documents/Battlefield-Singularity-November-2017.pdf?mtime=20171129235804.

96. Col. (ret) L.G. Shattuck | Transitioning to Autonomy: A Human Systems Integration Perspective | Naval Postgraduate School workshop.

97. Hong Cheng | Autonomous Intelligent Vehicles: Theory, Algorithms, and Implementation | Berlin: Springer | 2011 | p. 3.

98. Office of the Under Secretary of Defense for Acquisition, Technology and Logistics | Report of the Defense Science Board Summer Study on Autonomy | June 2016 | https://www.hsdl.org/?view&did=794641 | p. 11.

99. R. Martinage, Center for Strategic and Budgetary Assessments | Toward a New Offset Strategy: Exploiting U.S. Long-Term Advantages to Restore U.S. Global Power Projection Capability | 2014 | http://csbaonline.org/uploads/documents/Offset-Strategy-Web.pdf | p. 41.

100. M.L. Cummings | Artificial Intelligence and the Future of Warfare | Royal Institute of International Affairs research paper | Chatham House | January 2017 | https://www.chathamhouse.org/ sites/files/chathamhouse/publications/research/2017-01-26-artificial-intelligence-future-warfare-cummings-final.pdf | p. 1.

101. P. Scharre, Center for a New American Security | Robotics on the Battlefield, Part II: The Coming Swarm | October 2014 | https://s3.amazonaws.com/files.cnas.org/documents/ CNAS_TheComingSwarm_Scharre.pdf?mtime=20160906082059 | p. 42.

102. P. Scharre, Center for a New American Security | Robotics on the Battlefield, Part II: The Coming Swarm | October 2014 | https://s3.amazonaws.com/files.cnas.org/documents/ CNAS_TheComingSwarm_Scharre.pdf?mtime=20160906082059 | p. 10.

103. V. Boulanin and M. Verbruggen, Stockholm International Peace Research Institute | Mapping the Development of Autonomy in Weapon Systems | November 2017 | p. 30.

104. P. Scharre, Center for a New American Security | Robotics on the Battlefield, Part II: The Coming Swarm | October 2014 | https://s3.amazonaws.com/files.cnas.org/documents/ CNAS_TheComingSwarm_Scharre.pdf?mtime=20160906082059 | p. 26.

105. P. Scharre, Center for a New American Security | Robotics on the Battlefield, Part I: Range, Persistence, and Daring | May 2014 | https://s3.amazonaws.com/files.cnas.org/documents/ CNAS_RoboticsOnTheBattlefield_Scharre.pdf?mtime=20160906081925 | pp. 5, 10.

106. D. Gonzales and S. Harting | Designing Unmanned Systems with Greater Autonomy | RAND | 2014 | https://www.rand.org/content/dam/rand/pubs/research_reports/RR600/RR626/RAND_RR626.pdf | p. 23.

Page 58: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 58

UNCLASSIFIED

107. P. Scharre, Center for a New American Security | Robotics on the Battlefield, Part II: The Coming Swarm | October 2014 | https://s3.amazonaws.com/files.cnas.org/documents/ CNAS_TheComingSwarm_Scharre.pdf?mtime=20160906082059 | p. 10.

108. E.B. Kania, Center for a New American Security | Battlefield Singularity: Artificial Intelligence, Military Revolution, and China’s Future Military Power | November 2017 | https://s3.amazonaws.com/files.cnas.org/documents/Battlefield-Singularity-November-2017.pdf?mtime=20171129235804.

109. A. Ilachinski, Center for Naval Analyses | AI, Robots, and Swarms: Issues, Questions, and Recommended Studies | January 2017 | pp. 139–145.

110. Office of the Under Secretary of Defense for Acquisition, Technology, and Logistics | Report 11-S-3613 | Unmanned Systems Integrated Roadmap: FY2011-2036 | p. 27 | http://www.acq.osd.mil/ sts/docs/Unmanned%20Systems%20Integrated%20Roadmap%20FY2011-2036.pdf.

111. V. Boulanin and M. Verbruggen, Stockholm International Peace Research Institute | Mapping the Development of Autonomy in Weapon Systems | November 2017 | p. 14.

112. A. Ilachinski, Center for Naval Analyses | AI, Robots, and Swarms: Issues, Questions, and Recommended Studies | January 2017 | p. 139.

113. P. Scharre, Center for a New American Security | Robotics on the Battlefield, Part II: The Coming Swarm | October 2014 | https://s3.amazonaws.com/files.cnas.org/documents/ CNAS_TheComingSwarm_Scharre.pdf?mtime=20160906082059 | p. 15.

114. V. Boulanin and M. Verbruggen, Stockholm International Peace Research Institute | Mapping the Development of Autonomy in Weapon Systems | November 2017 | p. 30.

115. P. Scharre, Center for a New American Security | Robotics on the Battlefield, Part II: The Coming Swarm | October 2014 | https://s3.amazonaws.com/files.cnas.org/documents/ CNAS_TheComingSwarm_Scharre.pdf?mtime=20160906082059 | pp. 10, 15, 22.

116. W.C. Marra and S.K. McNeil | Understanding ‘The Loop:’ Regulating the Next Generation of War Machines | Harvard Journal of Law and Public Policy | Vol. 36, No. 3 | 2013 | p. 48 | https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2043131.

117. Office of the Under Secretary of Defense for Acquisition, Technology and Logistics | Report of the Defense Science Board Summer Study on Autonomy | June 2016 | https://www.hsdl.org/?view&did=794641 | pp. 12–13.

118. Govini | Department of Defense: Artificial Intelligence, Big Data and Cloud Taxonomy | December 2017 | http://www.govini.com/research-item/dod-artificial-intelligence-and-big-data-taxonomy/ | p. 12.

119. E.B. Kania, Center for a New American Security | Battlefield Singularity: Artificial Intelligence, Military Revolution, and China’s Future Military Power | November 2017 | https://s3.amazonaws.com/files.cnas.org/documents/Battlefield-Singularity-November-2017.pdf?mtime=20171129235804.

120. E. Jones | The Rise of Artificial Intelligence in Cyber Defense | Entrepreneur | 9 September 2016 | https://www.entrepreneur.com/article/281040.

121. Council of Insurance Agents and Brokers | Artificial Intelligence in Cybersecurity: Funding history, market breakdown, patents, forward looking trends | https://www.cbinsights.com/reports/CB-Insights_AI-in-Cybersecurity-Webinar.pdf.

122. G. Ollmann | How Artificial Intelligence Will Solve the Security Skills Shortage | 28 December 2016 | http://www.darkreading.com/operations/how-artificial-intelligence-will-solve-the-security-skills-shortage/a/d-id/1327756.

123. Defense Science Board | 21st Century Military Operations in a Complex Electromagnetic Environment | July 2015 | https://www.acq.osd.mil/dsb/reports/2010s/DSB_SS13--EW_Study.pdf.

Page 59: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 59

UNCLASSIFIED

124. C. Bing | The tech behind the DARPA Grand Challenge winner will now be used by the Pentagon | CyberScoop | 11 August 2017 | https://www.cyberscoop.com/mayhem-darpa-cyber-grand-challenge-dod-voltron/.

125. D. Song | (Artificial) Intelligence: What questions should DOD be asking? | Office of Net Assessment summer study | July 2016 | p. 44.

126. S.J. Feedberg, Jr. | Do Young Humans + Artificial Intelligence = Cybersecurity? | Breaking Defense | 13 November 2017 | https://breakingdefense.com/2017/11/do-young-humans-artificial-intelligence-cybersecurity/.

127. G.I. Seffers | Smarter AI for Electronic Warfare | The Cyber Edge | 1 November 2017 | https://www.afcea.org/content/smarter-ai-electronic-warfare.

128. D. Song | (Artificial) Intelligence: What questions should DOD be asking? | Office of Net Assessment summer study | July 2016 | p. 44.

129. S.J. Feedberg, Jr. | Do Young Humans + Artificial Intelligence = Cybersecurity? | Breaking Defense | 13 November 2017 | https://breakingdefense.com/2017/11/do-young-humans-artificial-intelligence-cybersecurity/.

130. G. Allen and T. Chan, Belfer Center for Science and International Affairs | Artificial Intelligence and National Security | July 2017 | https://www.belfercenter.org/sites/default/files/files/publication/ AI%20NatSec%20-%20final.pdf | p. 23.

131. E.B. Kania, Center for a New American Security | Battlefield Singularity: Artificial Intelligence, Military Revolution, and China’s Future Military Power | November 2017 | https://s3.amazonaws.com/files.cnas.org/documents/Battlefield-Singularity-November-2017.pdf?mtime=20171129235804.

132. Govini | Department of Defense: Artificial Intelligence, Big Data and Cloud Taxonomy | December 2017 | http://www.govini.com/research-item/dod-artificial-intelligence-and-big-data-taxonomy/ | p. 12.

133. E.B. Kania, Center for a New American Security | Battlefield Singularity: Artificial Intelligence, Military Revolution, and China’s Future Military Power | November 2017 | https://s3.amazonaws.com/files.cnas.org/documents/Battlefield-Singularity-November-2017.pdf?mtime=20171129235804.

134. Maj. S. Banks, School of Advanced Military Studies, United States Army Command and General Staff College | Lifting Off of the Digital Plateau with Military Decision Support Systems | 2013 | http://www.dtic.mil/get-tr-doc/pdf?AD=ADA583735 | p. 1.

135. E.B. Kania, Center for a New American Security | Battlefield Singularity: Artificial Intelligence, Military Revolution, and China’s Future Military Power | November 2017 | https://s3.amazonaws.com/files.cnas.org/documents/Battlefield-Singularity-November-2017.pdf?mtime=20171129235804.

136. Govini | Department of Defense: Artificial Intelligence, Big Data and Cloud Taxonomy | December 2017 | http://www.govini.com/research-item/dod-artificial-intelligence-and-big-data-taxonomy/ | p. 12.

137. E.B. Kania, Center for a New American Security | Battlefield Singularity: Artificial Intelligence, Military Revolution, and China’s Future Military Power | November 2017 | https://s3.amazonaws.com/files.cnas.org/documents/Battlefield-Singularity-November-2017.pdf?mtime=20171129235804.

138. DOD | Directive 3000.09 | Autonomy in Weapon Systems | Last revised 8 May 2017 | http://www.esd.whs.mil/Portals/54/Documents/DD/issuances/dodd/300009p.pdf.

139. S. Pham | Elon Must backs call for global ban on killer robots | CNN | 21 August 2017 | http://money.cnn.com/2017/08/21/technology/elon-musk-killer-robot-un-ban/index.html.

Page 60: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 60

UNCLASSIFIED

140. A. Petroff | Elon Musk says Mark Zuckerberg’s understanding of AI is ‘limited’ | CNN | 25 July 2017 | http://money.cnn.com/2017/07/25/technology/elon-musk-mark-zuckerberg-ai-artificial-intelligence/index.html.

141. G.P. Noone and D.C. Noone | The Debate Over Autonomous Weapons Systems | Case Western Reserve Journal of International Law | Vol. 47, Iss. 1 | Spring 2015 | p. 25 | http://scholarlycommons.law.case.edu/cgi/viewcontent.cgi?article=1005&context=jil).

142. M.C. Haas | Autonomous Weapon Systems: The Military’s Smartest Toys? | The National Interest | 20 November 2014 | http://nationalinterest.org/feature/autonomous-weapon-systems-the-militarys-smartest-toys-11708.

143. J. Rabkin and J. Yoo | Striking Power | Encounter Books | 2017.

144. E.B. Kania, Center for a New American Security | Battlefield Singularity: Artificial Intelligence, Military Revolution, and China’s Future Military Power | November 2017 | https://s3.amazonaws.com/files.cnas.org/documents/Battlefield-Singularity-November-2017.pdf?mtime=20171129235804.

145. M. Worcestor, Institut für Strategie- Politik- Sicherheits- und Wirtschaftsberatung | Autonomous Warfare – A Revolution in Military Affairs | Strategy Series | No. 340 | April 2015 | https://www.files.ethz.ch/isn/190160/340_Worcester.pdf | p. 4.

146. A. Ilachinski, Center for Naval Analyses | AI, Robots, and Swarms: Issues, Questions, and Recommended Studies | January 2017 | p. 210.

147. V. Boulanin and M. Verbruggen, Stockholm International Peace Research Institute | Mapping the Development of Autonomy in Weapon Systems | November 2017 | p. 2.

148. F. Sauer, Arms Control Association | Stopping ‘Killer Robots’ | October 2016 | https://www.armscontrol.org/ACT/2016_10/Features/Stopping-Killer-Robots-Why-Now-Is-the-Time-to-Ban-Autonomous-Weapons-Systems.

149. V. Boulanin and M. Verbruggen, Stockholm International Peace Research Institute | Mapping the Development of Autonomy in Weapon Systems | November 2017 | p. 25.

150. Rosenberg and Markoff | The Pentagon’s ‘Terminator Conundrum’ |

151. A. Ilachinski, Center for Naval Analyses | AI, Robots, and Swarms: Issues, Questions, and Recommended Studies | January 2017 | p. 209.

152. M. Rosenberg and J. Markoff | The Pentagon’s ‘Terminator Conundrum’ | The New York Times | 25 October 2016 | https://www.nytimes.com/2016/10/26/us/pentagon-artificial-intelligence-terminator.html.

153. R. Baldwin | The robots of war: AI and the future of combat | Engadget | 18 August 2016 | https://www.engadget.com/2016/08/18/robots-of-war-ai-and-the-future-of-combat/.

154. V. Boulanin and M. Verbruggen, Stockholm International Peace Research Institute | Mapping the Development of Autonomy in Weapon Systems | November 2017 | p. 25.

155. P. Scharre, Center for a New American Security | Robotics on the Battlefield, Part II: The Coming Swarm | October 2014 | https://s3.amazonaws.com/files.cnas.org/documents/ CNAS_TheComingSwarm_Scharre.pdf?mtime=20160906082059 | p. 15.

156. World Economic Forum | The Global Risks Report 2017, 12th ed. | http://www3.weforum.org/ docs/GRR17_Report_web.pdf | p. 48.

157. A. Ilachinski, Center for Naval Analyses | AI, Robots, and Swarms: Issues, Questions, and Recommended Studies | January 2017 | p. vii.

158. Govini | Department of Defense: Artificial Intelligence, Big Data and Cloud Taxonomy | December 2017 | http://www.govini.com/research-item/dod-artificial-intelligence-and-big-data-taxonomy/ | p. 12.

Page 61: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 61

UNCLASSIFIED

159. E.B. Kania, Center for a New American Security | Battlefield Singularity: Artificial Intelligence, Military Revolution, and China’s Future Military Power | November 2017 | https://s3.amazonaws.com/files.cnas.org/documents/Battlefield-Singularity-November-2017.pdf?mtime=20171129235804.

160. Joint Publication 1 | Doctrine for the Armed Forces of the United States | 25 March 2013 | http://www.dtic.mil/doctrine/new_pubs/jp1.pdf | p. I-18.

161. Col. J.R. Surdu and K. Kittka | Deep Green: Commander’s tool for COA’s Concept | http://www.bucksurdu.com/Professional/Documents/11260-CCCT-08-DeepGreen.pdf | p. 2.

162. E.B. Kania, Center for a New American Security | Battlefield Singularity: Artificial Intelligence, Military Revolution, and China’s Future Military Power | November 2017 | https://s3.amazonaws.com/files.cnas.org/documents/Battlefield-Singularity-November-2017.pdf?mtime=20171129235804.

163. Maj. S. Banks, School of Advanced Military Studies, United States Army Command and General Staff College | Lifting Off of the Digital Plateau with Military Decision Support Systems | 2013 | http://www.dtic.mil/get-tr-doc/pdf?AD=ADA583735 | p. 43.

164. Google DeepMind’s AlphaGo: How it works | Tastehit | 16 March 2016 | https://www.tastehit.com/blog/google-deepmind-alphago-how-it-works/.

165. D. Muoio | Why Go is so much harder for AI to beat than chess | 10 March 2016 | http://www.businessinsider.com/why-google-ai-game-go-is-harder-than-chess-2016-3.

166. Military Implications of AlphaGo | 30 May 2017 | http://actionablethought.com/military-implications-of-alphago/.

167. A.F. Krepinevich, Jr., Center for Strategic and Budgetary Assessments | The Military-Technical Revolution: A Preliminary Assessment | 2002 | http://csbaonline.org/uploads/documents/2002.10.02-Military-Technical-Revolution.pdf | pp. 11–14.

168. A.F. Krepinevich, Jr., Center for Strategic and Budgetary Assessments | The Military-Technical Revolution: A Preliminary Assessment | 2002 | http://csbaonline.org/uploads/documents/2002.10.02-Military-Technical-Revolution.pdf | pp. 11–18.

169. A.F. Krepinevich, Jr., Center for Strategic and Budgetary Assessments | The Military-Technical Revolution: A Preliminary Assessment | 2002 | http://csbaonline.org/uploads/documents/2002.10.02-Military-Technical-Revolution.pdf | p. 6.

170. R. Safian | When Artificial Intelligence Meets Actual Life | Fast Company | 23 October 2017 | https://www.fastcompany.com/40473767/when-artificial-intelligence-meets-actual-life?utm_source=postup&utm_medium=email&utm_campaign=Fast%20Company%20Daily&position=3&partner=newsletter&campaign_date=10232017.

171. What AI Can Really Do Right Now | Popular Mechanics | November 2017 | http://www.popularmechanics.com/technology/robots/a28380/everything-to-know-about-ai/.

172. J. Bughin, et al. | Artificial Intelligence: The Next Digital Frontier? | McKinsey Global Institute discussion paper | June 2017 | http://www.mckinsey.com/business-functions/mckinsey-analytics/our-insights/how-artificial-intelligence-can-deliver-real-value-to-companies | pp. 10, 24.

173. H. McCracken | How to Stop Worrying and Love the Great AI War of 2018 | Fast Company | 10 October 2017 | https://www.fastcompany.com/40474564/how-to-stop-worrying-and-love-the-great-ai-war-of-2018.

174. J. Ross | The Fatal Flaw of AI Implementation | MIT Sloan Review | 14 July 2017 | http://sloanreview.mit.edu/article/the-fatal-flaw-of-ai-implementation/.

175. J. Bughin and E. Hazan | Five Management Strategies for Getting the Most From AI | MIT Sloan Review | 19 September 2017 | http://sloanreview.mit.edu/article/five-management-strategies-for-getting-the-most-from-ai/.

Page 62: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 62

UNCLASSIFIED

176. C. Clark | ‘Landmark Event’ In Artificial Intelligence: DeepMind Trains Itself | Breaking Defense | 19 October 2017 | https://breakingdefense.com/2017/10/landmark-event-in-artificial-intelligence-deepmind-trains-itself/ and https://deepmind.com/blog/alphago-zero-learning-scratch/.

177. J. Somers | Is AI Riding a One-Trick Pony? | MIT Technology Review | 29 September 2017 | https://www.technologyreview.com/s/608911/is-ai-riding-a-one-trick-pony.

178. G. Allen and T. Chan, Belfer Center for Science and International Affairs | Artificial Intelligence and National Security | July 2017 | https://www.belfercenter.org/sites/default/files/files/publication/ AI%20NatSec%20-%20final.pdf | 12-41 and 71-110.

179. M.C. Horowitz | The Diffusion of Military Power: Causes and Consequences for International Politics | Princeton, NJ: Princeton University Press | 2010 | chapter 2.

180. S.J. Freedberg Jr. | Armed Robots: US Lags Rhetoric, Russia | Breaking Defense | 18 October 2017 | https://breakingdefense.com/2017/10/armed-robots-us-lags-rhetoric-russia/.

181. S. Snow | The army is developing robot resupply vehicles | Army Times | 20 September 2017 | https://www.armytimes.com/news/your-army/2017/09/20/the-army-is-developing-robot-resupply-vehicles/.

182. C. Clark | Cardillo: 1 Million Times More GEOINT Data in 5 Years | Breaking Defense | 5 June 2017 | https://breakingdefense.com/2017/06/cardillo-1-million-times-more-geoint-data-in-5-years/).

183. J. McLaughlin | Artificial Intelligence Will Put Spies Out of Work, Too | Foreign Policy | 9 June 2017 | http://foreignpolicy.com/2017/06/09/artificial-intelligence-will-put-spies-out-of-work-too/.

184. S. Erwin | With commercial satellite imagery, computer learns to quickly find missile sites in China | Space News | 19 October 2017 | http://spacenews.com/with-commercial-satellite-imagery-computer-learns-to-quickly-find-missile-sites-in-china/?utm_source=RC+Defense+Morning+Recon&utm_ campaign=79b5d537c7-EMAIL_CAMPAIGN_2017_10_20&utm_medium=email&utm_term= 0_694f73a8dc-79b5d537c7-84044145.

185. D. Warne | DIUx likely to continue radar initiative in spite of setback | Space News | 13 October 2017 | http://spacenews.com/diux-likely-to-continue-radar-initiative-in-spite-of-setback/.

186. K. Mizokami | The Navy’s New AI Missile Sinks Ships the Smart Way | Popular Mechanics | 25 February 2016 | http://www.popularmechanics.com/military/weapons/a19624/the-navys-new-missile-sinks-ships-the-smart-way/).

187. S.J. Freedberg Jr. | Navy Warships Get New Heavy Missile: 2,500-Lb LRASM | Breaking Defense | 26 July 2017 | https://breakingdefense.com/2017/07/navy-warships-get-new-heavy-missile-2500-lb-lrasm/.

188. Defense Business Board | Transforming the Department of Defense’s Core Business Processes for Revolutionary Change | DBB Report 15-1 | http://www.dtic.mil/dtic/tr/fulltext/u2/a618526.pdf) | p. 19.

189. Defense Business Board | Focusing a Transition: Challenges Facing the New Administration | DBB Report 16-5 | http://dbb.defense.gov/Portals/35/Documents/Reports/ 2016/DBB%20Transition%20Report%202016%20-%2020160920.pdf | pp. 53–61.

190. PwC | Bot.Me: A revolutionary partnership | April 2017 | http://www.pwc.in/assets/pdfs/consulting/ digital-enablement-advisory1/pwc-botme-booklet.pdf | p. 16.

191. DOD Instruction 5000.02 | Operation of the Defense Acquisition System | 7 January 2015 | http://www.esd.whs.mil/Portals/54/Documents/DD/issuances/dodi/500002_dodi_2015.pdf.

192. Secretary of Defense Chuck Hagel | Memorandum establishing the Defense Innovation Initiative | 15 November 2014 | http://archive.defense.gov/pubs/OSD013411-14.pdf | p. 2.

193. Under Secretary of Defense for Acquisition, Technology, and Logistics Frank Kendall | Implementation directive for Better Buying Power 3.0 | 9 April 2015 | http://www.acq.osd.mil/fo/docs/betterBuyingPower3.0(9Apr15).pdf | pp. 12–13.

Page 63: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 63

UNCLASSIFIED

194. Office of Naval Research | LOCUST demonstration video | April 2016 | https://www.youtube.com/watch?v=8FukTsKmXOo.

195. D.B. Larter | Meet the Navy’s new sub-hunting drone ship | Navy Times | 8 April 2016 | https://www.navytimes.com/news/your-navy/2016/04/08/meet-the-navy-s-new-sub-hunting-drone-ship/.

196. M. Pomerleau | DOD plans to invest $600M in unmanned underwater vehicles | Defense Systems | 4 February 2016 | https://defensesystems.com/articles/2016/02/04/dod-navy-uuv-investments.aspx.

197. D. Axe | U.S. Air Force Sends Robotic F-16s Into Mock Combat | National Interest | 16 May 2017 | http://nationalinterest.org/blog/the-buzz/us-air-force-sends-robotic-f-16s-mock-combat-20684.

198. A. Ng | Stop Cyberattacks. Just add robots | CNET | 1 September 2017 | https://www.cnet.com/ news/cyberattacks-artificial-intelligence-ai-hackers-defcon-black-hat/.

199. E. Tegler | Why DARPA Needs AI to Defeat Enemy Radar | Popular Mechanics | 13 September 2016 | http://www.popularmechanics.com/military/research/news/a22834/darpa-ai-defeat-enemy-radar/.

200. M. Weisgerber | The Increasingly Automated Hunt for Mobile Missile Launchers | Defense One | 28 April 2016 | https://www.defenseone.com/technology/2016/04/increasingly-automated-hunt-mobile-missile-launchers/127864/.

201. Deputy Secretary of Defense Robert Work quoted in Johns Hopkins Applied Physics Laboratory | The Future of Humans & Machines: Partnership, Fusion, or Fear? | Event summary | Summer 2017 | p. 5.

202. C. Clark | VCJCS Selva Says US Must Not Let Robots Decide Who Dies; Supports LRSO | Breaking Defense | 18 July 2017 | https://breakingdefense.com/2017/07/vcjcs-selva-us-must-not-let-robots-decide-who-dies-supports-lrso/.

203. YouTube | This Russian robot shoots guns | Published 25 April 2017 | https://www.youtube.com/watch?v=HTPIED6jUdU.

204. A. Sulleyman | Robot Being Trained to Shoot Guns is ‘Not a Terminator,’ Insists Russian Deputy Prime Minister | The Independent | 14 April 2017 | http://www.independent.co.uk/life-style/gadgets-and-tech/news/terminator-robot-fedor-guns-russia-shooting-dmitry-rogozin-a7684406.html.

205. I. Nguyen | AI adoption is limited by incurred risk, not potential benefit | Venture Beat | 25 November 2017 | https://venturebeat.com/2017/11/25/ai-adoption-is-limited-by-incurred-risk-not-potential-benefit/.

206. R. Brown | Where is AI Headed in 2018? | 3 December 2017 | https://blogs.nvidia.com/ blog/2017/12/03/ai-headed-2018/.

207. J. Doubleday | Pentagon plans multibillion-dollar bet on single cloud | Inside Defense | 8 March 2018| https://insidedefense.com/daily-news/pentagon-plans-multibillion-dollar-bet-single-cloud.

208. W. Knight | The U.S. Leads in Artificial Intelligence, but for How Long? | MIT Technology Review | 6 December 2017 | https://www.technologyreview.com/s/609610/the-us-leads-in-artificial-intelligence-but-for-how-long/.

209. E.B. Kania, Center for a New American Security | Battlefield Singularity: Artificial Intelligence, Military Revolution, and China’s Future Military Power | November 2017 | https://s3.amazonaws.com/files.cnas.org/documents/Battlefield-Singularity-November-2017.pdf?mtime=20171129235804.

210. P. Mozur and K. Bradsher | China’s A.I. Advances Help Its Tech Industry, and state Security | New York Times | 3 December 2017 | https://www.nytimes.com/2017/12/03/business/china-artificial-intelligence.htm.

Page 64: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 64

UNCLASSIFIED

211. China recruits Baidu, Alibaba and Tencent to AI ‘national team’ | South China Morning Post | 21 November 2017 | http://www.scmp.com/tech/china-tech/article/2120913/china-recruits-baidu-alibaba-and-tencent-ai-national-team.

212. H. Sender | China gains in race to develop AI-enabled weapons | Nikkei Asian Review | 29 November 2017 | https://asia.nikkei.com/Features/Cover-story/China-gains-in-race-to-develop-AI-enabled-weapons.

213. C. Clark | Our Artificial Intelligence ‘Sputnik Moment’ is Now: Eric Schmidt & Bob Work | Breaking Defense | 1 November 2017 | https://breakingdefense.com/2017/11/our-artificial-intelligence-sputnik-moment-is-now-eric-schmidt-bob-work/.

214. J.R. Allen and A. Husain | The Next Space Race Is Artificial Intelligence. And the United States is Losing | Foreign Policy | 3 November 2017 | https://breakingdefense.com/2017/11/our-artificial-intelligence-sputnik-moment-is-now-eric-schmidt-bob-work/.

215. S. Gibbs | Elon Musk: AI ‘vastly more risky than North Korea’ | The Guardian | 14 August 2017 | https://www.theguardian.com/technology/2017/aug/14/elon-musk-ai-vastly-more-risky-north-korea.

216. R. Cellan-Jones | Stephen Hawking warns artificial intelligence could end mankind | BBC News | 2 December 2014 | http://www.bbc.com/news/technology-30290540.

217. P. Holley | Bill Gates on dangers of artificial intelligence: ‘I don’t understand why some people are not concerned’ | The Washington Post | 29 January 2015 | https://www.washingtonpost.com/news/the-switch/wp/2015/01/28/bill-gates-on-dangers-of-artificial-intelligence-dont-understand-why-some-people-are-not-concerned/?utm_term=.e5effe7af7ff).

218. C. Clifford | Bill Gates: I do not agree with Elon Musk about A.I. ‘We shouldn’t panic about it’ | CNBC | 25 September 2017 | https://www.cnbc.com/2017/09/25/bill-gates-disagrees-with-elon-musk-we-shouldnt-panic-about-a-i.html.

219. C. Clifford | Head of A.I. at Google slams the kind of ‘A.I. apocalypse’ fear-mongering that Elon Musk has been doing | CNBC | 21 September 2017 | https://www.cnbc.com/2017/09/21/head-of-google-a-i-slams-fear-mongering-about-the-future-of-a-i.html.

220. Future of Life Institute | Autonomous Weapons: An Open Letter from AI & Robotics Researchers | 28 July 2017 | https://futureoflife.org/open-letter-autonomous-weapons/.

221. S.M. Walt | There’s Still No Reason to Think the Kellogg-Briand Pact Accomplished Anything | Foreign Policy | 29 September 2017 | http://foreignpolicy.com/2017/09/29/theres-still-no-reason-to-think-the-kellogg-briand-pact-accomplished-anything/.

222. O. Russakovsky, et al. | ImageNet Large-Scale Visual Recognition Challenge | arXiv | 1 December 2014 | https://arxiv.org/abs/1409.0575.

223. K. He, et al. | Deep Residual Learning for Image Recognition | arXiv | 10 December 2015 | https://arxiv.org/abs/1512.03385.

224. A. Krizhevsky, I. Sutskever, and G.E. Hinton | ImageNet Classification with Deep Convolutional Neural Networks | Proceedings of the 25th International Conference on Neural Information Processing Systems | Vol. I | December 2012 | https://dl.acm.org/citation.cfm?id=2999257.

225. C. Szegedy, et al. | Going deeper with Convolutions | IEEE Conference on Computer Vision and Pattern Recognition | 2015 | http://ieeexplore.ieee.org/document/7298594/?reload=true.

226. K. He, et al. | Deep Residual Learning for Image Recognition | arXiv | 10 December 2015 | https://arxiv.org/abs/1512.03385.

227. Jie Hu, Li Shen, and Gang Sun | Squeeze-and-Excitation Networks | arXiv | 5 September 2017 | https://arxiv.org/abs/1709.01507.

228. T.N. Mundhenk, et al., | A Large Contextual Dataset for Classification, Detection, and Counting of Cars with deep learning | arXiv | 14 September 2016 | https://arxiv.org/abs/1609.04453.

Page 65: Coming of Age: Artificial Intelligence and the Continuing Revolution … · 2019-02-12 · Coming of Age: Artificial Intelligence and the Continuing Revolution in Military Affairs

UNCLASSIFIED

LLNL-MI-750860

Page 65

UNCLASSIFIED

229. T.N. Mundhenk, et al., | A Large Contextual Dataset for Classification, Detection, and Counting of Cars with deep learning | arXiv | 14 September 2016 | https://arxiv.org/abs/1609.04453.

230. J.J. Godfrey, E.C. Holliman, and J. McDaniel | Switchboard: Telephone speech corpus for research and development | IEEE | 1992 | http://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=225858.

231. W. Xiong, et al. | The Microsoft 2017 Conversational Speech Recognition System | Microsoft Research Technical Report MSR-TR-2017-39 | 2017 | https://www.microsoft.com/en-us/research/wp-content/uploads/2017/08/ms_swbd17-2.pdf.

232. O. Vinyals, S. Gaffney, and T. Ewalds, Deepmind | DeepMind and Blizzard open StarCraft II as an AI research environment | 9 August 2017 | https://deepmind.com/blog/deepmind-and-blizzard-open-starcraft-ii-ai-research-environment/.