[Senate Hearing 116-502]
[From the U.S. Government Publishing Office]
S. Hrg. 116-502
ARTIFICIAL INTELLIGENCE INITIATIVES WITHIN THE DEPARTMENT OF DEFENSE
=======================================================================
HEARING
before the
SUBCOMMITTEE ON
EMERGING THREATS AND CAPABILITIES
of the
COMMITTEE ON ARMED SERVICES
UNITED STATES SENATE
ONE HUNDRED SIXTEENTH CONGRESS
FIRST SESSION
__________
MARCH 12, 2019
__________
Printed for the use of the Committee on Armed Services
GRAPHIC(S) NOT AVAILABLE IN TIFF FORMAT]
Available via: http://www.govinfo.gov
______
U.S. GOVERNMENT PUBLISHING OFFICE
46-003 PDF WASHINGTON : 2021
COMMITTEE ON ARMED SERVICES
JAMES M. INHOFE, Oklahoma, JACK REED, Rhode Island
Chairman JEANNE SHAHEEN, New Hampshire
ROGER F. WICKER, Mississippi KIRSTEN E. GILLIBRAND, New York
DEB FISCHER, Nebraska RICHARD BLUMENTHAL, Connecticut
TOM COTTON, Arkansas MAZIE K. HIRONO, Hawaii
MIKE ROUNDS, South Dakota TIM KAINE, Virginia
JONI ERNST, Iowa ANGUS S. KING, Jr., Maine
THOM TILLIS, North Carolina MARTIN HEINRICH, New Mexico
DAN SULLIVAN, Alaska ELIZABETH WARREN, Massachusetts
DAVID PERDUE, Georgia GARY C. PETERS, Michigan
KEVIN CRAMER, North Dakota JOE MANCHIN, West Virginia
MARTHA McSALLY, Arizona TAMMY DUCKWORTH, Illinois
RICK SCOTT, Florida DOUG JONES, Alabama
MARSHA BLACKBURN, Tennessee
JOSH HAWLEY, Missouri
John Bonsell, Staff Director
Elizabeth L. King, Minority Staff
Director
Subcommittee on Emerging Threats and Capabilities
JONI ERNST, Iowa, Chairman GARY C. PETERS, Michigan
DEB FISCHER, Nebraska JEANNE SHAHEEN, New Hampshire
KEVIN CRAMER, North Dakota MAZIE K. HIRONO, Hawaii
MARSHA BLACKBURN, Tennessee MARTIN HEINRICH, New Mexico
JOSH HAWLEY, Missouri
(ii)
C O N T E N T S
March 12, 2019
Page
Artificial Intelligence Initiatives Within the Department of 1
Defense.
Statements of Members of the Subcommittee
Statement of Senator Joni Ernst.................................. 1
Statement of Senator Gary C. Peters.............................. 2
Witness Statements
Highnam, Peter T., Deputy Director, Defense Advanced Research 3
Projects Agency.
Brown, Michael A., Director, Defense Innovation Unit............. 9
Shanahan, Lieutenant General John N.T., USAF, Director, Joint 15
Artificial Intelligence Center, Office of the Department of
Defense Chief Information Officer.
Questions for the Record......................................... 36
(iii)
ARTIFICIAL INTELLIGENCE INITIATIVES WITHIN THE DEPARTMENT OF DEFENSE
----------
TUESDAY, MARCH 12, 2019
United States Senate,
Subcommittee on Emerging
Threats and Capabilities,
Committee on Armed Services,
Washington, DC.
The Subcommittee met, pursuant to notice, at 2:37 p.m. in
room SR-232A, Russell Senate Office Building, Senator Joni
Ernst (Chairman of the Subcommittee) presiding.
Members present: Senators Ernst, Shaheen, Heinrich, and
Peters.
OPENING STATEMENT OF SENATOR JONI ERNST
Senator Ernst. The Subcommittee on Emerging Threats and
Capabilities meets today to receive testimony on Department of
Defense artificial intelligence (AI) initiatives.
I want to thank you, gentlemen, for being here today.
I do apologize. We have a vote series going on right now,
so, at some point, Senator Peters and I may have to switch on
and off. We'll run down and vote accordingly.
But, I do want to thank you for being here. I'd love to
welcome you. We have a very distinguished panel joining us
today, ladies and gentlemen.
As highlighted in the NDS [National Defense Strategy]
Commission Report, the U.S. must stay ahead in several emerging
technologies in order to maintain or regain a warfighting
advantage. These technologies include hypersonics, directed
energy, artificial intelligence, 5G, and quantum computing.
Russia and China are aggressively developing these capabilities
and, in some cases, have already surpassed, or will soon
surpass, our technologies. Without action, the U.S. may find
itself at a technological disadvantage in future conflicts.
Over the next few months, this Subcommittee will focus our
efforts to ensure the Department is well positioned to outpace
our adversaries and maintain a strategic advantage in these key
technologies. Today, we will focus on one technology of
particular importance, which is artificial intelligence.
The recently released Department of Defense Artificial
Intelligence Strategy makes clear AI is poised to transform
every industry and is expected to impact every corner of the
Department, spanning operations, training, sustainment, force
protection, recruiting, healthcare, and many others. AI has the
ability to provide powerful new capabilities to our warfighters
that we are only beginning to imagine. When applied to back-
office functions and operations within the Department, AI will
be critical in boosting efficiency and increasing the
effectiveness of limited resources. With such broad potential
impacts, it is important that the Department move quickly to
adopt these capabilities so that we don't lose the
technological edge.
Our adversaries understand the critical importance of AI.
Last year, the Chinese government released a strategy detailing
its plan to take the lead in AI by 2030. Less than 2 months
later, Vladimir Putin publicly announced Russia's intent to
pursue AI technologies, stating, ``Whoever becomes the leader
in this field will rule the world.'' Both of these countries
are investing heavily in military applications of AI to achieve
a warfighting advantage.
The United States is also investing heavily in AI
applications. Over the last year, the Department has initiated
several important efforts to accelerate the adoption of AI,
including the establishment of the Joint AI Center, or JAIC,
the development of a DARPA [Defense Advanced Research Projects
Agency] AI Next Campaign, and the release of the DOD
[Department of Defense] AI Strategy. These efforts demonstrate
the extent of the AI transformation already underway within the
Department of Defense and the priority that is being placed on
growing these capabilities. However, with these important
efforts comes the challenge of coordinating hundreds of
disparate AI efforts across multiple offices and organizations.
I look to our witnesses to help the committee better understand
how AI can be adopted more rapidly, how coordinating the
initiatives already underway within the Department can help us
harness this powerful technology, and where we must invest in
future research to ensure we maintain a long-term advantage.
Again, I thank our witnesses for being with us today. I
look forward to their testimony.
I would turn to my Ranking Member, Mr. Peters. Senator
Peters, thank you. New Ranking Member joining us. This is our
first subcommittee hearing of the Congress. Senator Peters, I
welcome you onboard. Thank you very much.
STATEMENT OF SENATOR GARY C. PETERS
Senator Peters. Well, Madam Chairwoman, thank you. It's an
honor to be with you. I've enjoyed working with you over the
years, and we'll continue to do that in the next 2 years in
this Congress.
Also, I'd like to thank the gentlemen for you being here
today, for your testimony.
You know, artificial intelligence is already impacting our
daily lives through commercial products and services, from
applications as simple as Alexa to very complex systems, like
self-driving automobiles. AI has huge implications for our
national security, as well. For example, it'll enable new
capabilities in intelligence analysis, autonomous systems, as
well as cybersecurity. At the same time, AI can create new
threats in these and other areas that can be exploited by our
adversaries. I hope, in this hearing, that we can hear from our
witnesses on the Pentagon's efforts to position itself in the
rapidly changing world of AI, from the more near-term
activities of the Joint AI Center to the long-term, high-risk,
high-payoff research efforts of DARPA and the Defense
Innovation Unit (DIU).
The current AI systems that exist today only exist because
of decades of research in computer science, control systems,
microelectronics, and other fields. There are many amazing
commercial capabilities available today, yet AI is still
relatively primitive to what we all think it can be. We need to
continue to invest in fields like computer science and
electronics, but also research in areas such as our
understanding of how machines and people learn and can work
together to make sure that that promise becomes a reality.
I would like to understand how the DOD is working to move
AI capabilities quickly into fielded systems, as well as how
the Pentagon is developing a long-range strategy on R&D
[research and development] of new capabilities.
I hope we can discuss how we are engaging with the best
minds in defense and the commercial industry, Silicon Valley,
government labs, and universities to address many of these
challenges.
I would also like to learn about efforts to ensure that the
DOD and the Nation has the expert workforce that we will need
within government to stay at the leading edge of this
technology.
Finally, I'd like to recognize--General Shanahan is a
distinguished graduate of the University of Michigan, earning
his commission there through the ROTC [Reserve Officer Training
Corps]. The University of Michigan is one of many academic
institutions in Michigan that prioritizes artificial
intelligence research, particularly for the development and
testing of autonomous vehicle systems.
General Shanahan, I know that U of M [University of
Michigan] would love to host you back on campus to see the work
that they're doing on AI and on autonomy and its relevance to
your work. I hope that you and--as well as our other members of
the panel, are able to take a trip to Michigan sometime soon.
I thank the Chair again for holding this hearing. I
certainly look forward to our discussion.
Senator Ernst. Yes. Thank you very much, Senator Peters.
We will start with Dr. Peter T. Highnam. Dr. Highnam became
the Deputy Director of the Defense Advanced Research Projects
Agency, what we know of as DARPA, in February of 2018. Prior to
coming to DARPA, Dr. Highnam was the Director of Research at
the National Geospatial-Intelligence Agency, on assignment from
the Office of the Director of National Intelligence (ODNI).
Prior to that assignment, he also served 6 years at ODNI's
Intelligence Advanced Research Projects Activity, initially as
an Office Director and then as Director.
Dr. Highnam, we welcome you. You may start your opening
remarks. Thank you very much.
STATEMENT OF PETER T. HIGHNAM, DEPUTY DIRECTOR, DEFENSE
ADVANCED RESEARCH PROJECTS AGENCY
Dr. Highnam. Thank you, Chairwoman Ernst, Ranking Member
Peters. I'm pleased to be here to represent the Defense
Advanced Research Projects Agency and share with the
Subcommittee DARPA's work to advance AI technologies.
I'm going to begin with a little bit about DARPA's history
in this field. In 1960, shortly after DARPA was created, ARPA
[Advanced Research Projects Agency] was created. One of the
first information technology offices that we had gave this
quote, ``It seems reasonable to envision bringing computing
machines effectively into processes of thinking that must go on
in realtime, time that moves too fast to permit using computers
in conventional ways. To think in interaction with a computer
in the same way that you think with a colleague whose
competence supplements your own will require much tighter
coupling between man and machine than is possible today.''
Back then, when computers were large, were room-size, when
they were being used for computing missile trajectories and so
on, this man saw what was possible, saw the insights that were
becoming available, and saw the push that we're still working
on, which is changing computers from tools to partners. That
actually is the history of AI investments by DARPA for the last
60 years.
That is also quintessential DARPA. This man had, you know,
one foot in defense, seeing the mission and seeing what was
needed, and one foot in the technology side, and framed the
problem using use cases, knew what had to be done, and started
driving. It's unlikely that he thought that there would be 6
decades of investments and hard work that followed that to get
to where we are today.
I'd like to say that, after 60 years of pushing, AI is an
overnight success.
[Laughter.]
Dr. Highnam. Really, within the last 10 years, when you
think about the kind of technologies. A lot of transitions and
successes over the decades, much accomplished, and much still
to do.
DARPA describes the investments in AI using a waves
construct. The first wave at the beginning, for the first 20
years or so, are normally known as describe. This is where
knowledge was encoded in rules, ``If A, then B.'' If you look
inside the tax--if you do personal taxes today, there's a rule-
based system inside there that was what, 40 years ago, would
have been called AI technologies, now is just computer science,
or IT [information technology]. That's the price of success.
It's no longer AI. It's just commonplace.
Then, beginning in the mid-1970s, the technology--science
had put in place to begin what became machine learning. The
theory was put down, but it--only in the last 10-15 years, we
now have the compute cycles, we have the data availability.
That's when the current wave, the second wave, of machine
learning really took place and really came into being.
Now we're looking past that, at DARPA, into what comes
next. We have two waves of technology. One was descriptive, one
was recognizing situations, classifying, and so on. Now we have
to be able to explain, to really build the trust between these
systems and the people who are using them and working with them
in realtime, in difficult, stressful situations, but building
the trust so that they really can become partners. This is the
role of explanation.
It's a great time to be at DARPA, because we're now on the
brink of a lot of really exciting things. That's the genesis of
the current initiative, the $2 billion investment that we've
said we're now making in AI technologies.
That said, there's a brittleness to the current
technologies. The tools are immature, still. We don't have an
engineering discipline behind AI technologies. There are
issues, that I'm sure everyone will talk to you about, about
missing data volume, missing data quality, provenance, and so
on, the training, second-wave systems. These systems tend to
have unexpected failure modes.
In front of you, there should be an example of the
brittleness of AI. These are drawn from the academia mixture.
You may have seen these before. In the first picture, on the
left, there's a panda, which you and I look at with all the
history that we have of looking at these critters. On the right
is also a picture that looks like a panda, as well, to us. The
difference is that, in the digital representation, a certain
amount of ``noise'' represented by the middle picture was laid
on top of it, and a highly trained second-wave classification
system, machine-learning system, went from classifying that
picture as a panda to now as a gibbon, with high certainty. The
fragility of these methods--these are very literal methods.
There is no semantics, there is no intelligence.
The second example is perhaps of more concern. This shows a
stop sign in a physical situation. Think autonomous vehicles.
To you and I, again, it's a stop sign. It has a certain shape.
To a trained system, to a highly trained system, it's no
longer--when you put that little white sticker--or that yellow
sticker onto the stop sign, it's now classified as a speed
limit sign. You can think in terms of autonomous vehicles, the
brittleness and fragility of the systems. You can also think in
terms of adversarial endeavors. It takes camouflage and
deception to a whole new level.
Very important to point those things out.
Today we have autonomy. We have a lot of work successes in
cyber, from first- and second-wave technologies. We have novel
hardware, high-performance hardware, low-energy hardware coming
into place. Yes, then we have a lot of tools, and hundreds of
thousands of people are being trained and really wanting to use
machine learning. We have to go to the next step, this
commonsense reasoning, being able to explain where this
inference came from. We have to get there. Otherwise, trust
won't come into place.
What we've done is to talk about, in our new initiative,
robust AI, dealing with adversarial AI, both unintentional and
intentional, high-performance, in terms of compute cycles and
minimizing energy, and delivering radically new capabilities.
This is the genesis of the AI Next Campaign, creating systems
capable of reasoning, regenerative, contextual, and explanatory
models. We already have over 20 programs running in AI, new
programs--research programs started. We have over 80 programs
in the agency. About one-third of the programs in the entire
agency now are either creating AI technologies or aggressive
users of those technologies.
Last, to your point about workforce, we really had to get
more people engaged. Typically, we put out a call for
proposals--research proposals, people apply, and, 6 to 9 months
later, if selected, they're on contract. We have something
called AI Exploration, by which we are driving the research
community to explore this--the space of the third wave
aggressively. We post a topic, and we award contracts within 90
days of posting the topic. We've now done this six times. We've
invested, so far, on the order of $45 million in this. There's
tremendous uptake from the research community, these
opportunities. All unclassified, all fundamental work.
From 60 years ago to now, I don't think Mr. Licklider, at
the time, would have anticipated that the Department of Defense
would have an AI strategy, such a huge success in recognition,
and that the President would sign an AI executive order. Who
would have thought?
Game-changing capabilities for the Defense Department and
the world, from 60 years of investment, much accomplished, and
much to do.
Thank you.
[The prepared statement of Dr. Peter Highnam follows:]
Prepared Statement by Dr. Peter Highnam
darpa's seminal role in the field of artificial intelligence
Seventy years ago, when early electronic computers ran on vacuum
tubes and filled entire rooms, researchers already were striving to
enable machines to think as people do. Only a few years after its start
in 1958, DARPA began playing a central role in realizing this ambition
by laying some of the groundwork for the field of artificial
intelligence (AI). Early work in AI emphasized handcrafted knowledge,
and computer scientists constructed so-called expert systems that
captured the rules that the system could then apply to situations of
interest. Such ``first wave'' AI technologies were quite successful--
tax preparation software is a good example of an expert system--but the
need to handcraft rules is costly and time-consuming and therefore
limits the applicability of rules-based AI technologies.
The past few years have seen an explosion of interest in a sub-
field of AI dubbed ``machine learning'' that applies statistical and
probabilistic methods to large data sets to create generalized
representations that can be applied to future samples. Foremost among
these approaches are deep learning (artificial) neural networks trained
to perform a variety of classification and prediction tasks when
adequate historical data is available. Therein lies the rub, however,
as the task of collecting, labelling, and vetting data on which to
train such ``second wave'' AI techniques is prohibitively costly and
time-consuming.
DARPA envisions a future in which machines are more than just tools
that execute human-programmed rules or generalize from human-curated
data sets. Rather, the machines DARPA envisions will function more as
colleagues than as tools. Towards this end, DARPA is focusing its
investments on a ``third wave'' of AI technologies that brings forth
machines that can reason in context. Incorporating these technologies
in military systems that collaborate with warfighters will facilitate
better decisions in complex, time-critical, battlefield environments;
enable a shared understanding of massive, incomplete, and contradictory
information; and empower unmanned systems to perform critical missions
safely and with high degrees of autonomy.
Today, DARPA is funding more than 24 programs exploring ways to
advance the state of the art in AI, pushing beyond second wave machine
learning towards contextual reasoning capabilities. This is in addition
to more than 55 active programs that are leveraging machine learning or
AI technologies in some capacity-from managing the electromagnetic
spectrum to detecting and patching cyber vulnerabilities.
This level of investment has been years in the making and will
define scientific and technical exploration, as well as resulting
military capabilities, for decades to come.
current programs
DARPA's Lifelong Learning Machines (L2M) program is exploring ways
to enable machines to learn while doing without catastrophic
forgetting. Such a capability would enable systems to improve on the
fly, recover from surprises, and keep them from drifting out of sync
with the world. First announced in 2017, L2M research teams are
developing complete systems and their components, as well as exploring
learning mechanisms in biological organisms with the goal of
translating them into computational processes. Discoveries in both
technical areas are expected to generate new methodologies that will
allow AI systems to learn and improve during tasks, apply previous
skills and knowledge to new situations, incorporate innate system
limits, and enhance safety in automated assignments. While the program
is still in its early stages, L2M researchers already have identified
and solved challenges associated with building and training a self-
reproducing neural network.
DARPA is also currently running a program called Explainable AI or
XAI to develop new machine-learning architectures that can produce
accurate explanations of their decisions in a form that makes sense to
humans. As AI algorithms become more widely used, reasonable self-
explanation will help users understand how these systems work, and how
much to trust them in various situations. XAI specifically aims to
create a suite of machine learning techniques that produce explainable
models--while maintaining a high level of prediction accuracy so human
users understand, appropriately trust, and effectively manage the
emerging generation of artificially intelligent partners. Enabling
computing systems in this manner is critical because sensor,
information, and communication systems generate data at rates beyond
what humans can assimilate, understand, and act upon.
The real breakthrough for artificial intelligence, however, will
not come until researchers figure out a way for machines to learn or
otherwise acquire common sense. Without common sense, AI systems will
be powerful but limited tools that require human inputs to function.
With common sense, an AI could become a partner in problem solving.
Common sense knowledge is so pervasive in our lives that it can be hard
to recognize. For example, in conflict and warzone situations, people
tend to make snap decisions about the cause of the problem and ignore
evidence that does not support their point of view. To act as a valued
partner in such situations, the AI system will need sufficient common
sense to know when to speak and what to say, which will require that it
have a good idea of what each person knows. Interrupting to state the
obvious would quickly result in its deactivation, particularly under
stressful conditions.
In order to find answers to the common sense problem, DARPA
launched in October of last year the Machine Common Sense (MCS)
program, which will explore recent advances in cognitive understanding,
natural language processing, deep learning, and other areas of AI
research. MCS is pursuing two approaches for developing and evaluating
different machine common sense services. The first approach seeks to
create computational models that learn from experience and mimic the
core domains of cognition as defined by developmental psychology. This
includes the domains of objects (intuitive physics), places (spatial
navigation), and agents (intentional actors). Researchers will develop
systems that think and learn as humans do in the very early stages of
development, leveraging advances in the field of cognitive development
to provide empirical and theoretical guidance.
To assess the progress and success of the first strategy's
computational models, researchers will explore developmental psychology
research studies and literature to create evaluation criteria. DARPA
will use the resulting set of cognitive development milestones to
determine how well the models are able to learn against three levels of
performance: prediction/expectation, experience learning, and problem
solving.
The second MCS approach will construct a common sense knowledge
repository capable of answering natural language and image-based
queries about common sense phenomena by reading from the Web. DARPA
expects that researchers will use a combination of manual construction,
information extraction, machine learning, crowdsourcing techniques, and
other computational approaches to develop the repository. The resulting
capability will be measured against the Allen Institute for Artificial
Intelligence (AI2) Common sense benchmark tests, which are constructed
through an extensive crowdsourcing process to represent and measure the
broad common sense knowledge of an average adult.
ai next campaign
DARPA announced in September 2018, a multi-year investment of more
than $2 billion in new and existing programs called the ``AI Next''
campaign. Campaign key areas include providing robust foundations for
second wave technologies, aggressively applying second wave AI
technologies into appropriate systems, and exploring and creating third
wave AI science and technologies.
AI Next builds on DARPA's five decades of AI technology creation to
define and to shape the future, always with the Department's hardest
problems in mind. Accordingly, DARPA will create powerful capabilities
for the DoD by attending specifically to the following areas:
New Capabilities: AI technologies are applied routinely to enable
DARPA R&D projects, including more than 60 ongoing programs, such as
the Electronic Resurgence Initiative, and other programs related to
real-time analysis of sophisticated cyber attacks, detection of
fraudulent imagery, construction of dynamic kill-chains for all-domain
warfare, human language technologies, multi-modality automatic target
recognition, biomedical advances, and control of prosthetic limbs.
DARPA will advance AI technologies to enable automation of critical
Department business processes. One such process is the lengthy
accreditation of software systems prior to operational deployment.
Automating this accreditation process with known AI and other
technologies now appears possible.
Robust AI: AI technologies have demonstrated great value to
missions as diverse as space-based imagery analysis, cyber attack
warning, supply chain logistics and analysis of microbiologic systems.
At the same time, the failure modes of AI technologies are poorly
understood. DARPA is working to address this shortfall, with focused
R&D, both analytic and empirical. DARPA's success is essential for the
Department to deploy AI technologies, particularly to the tactical
edge, where reliable performance is required.
Adversarial AI: The most powerful AI tool today is machine
learning. Machine learning systems are easily duped by changes to
inputs that would never fool a human. The data used to train such
systems can be corrupted, and the software itself is vulnerable to
cyber attack. These areas, and more, must be addressed at scale as more
AI-enabled systems are operationally deployed.
High Performance AI: Computer performance increases over the last
decade have enabled the success of machine learning, in combination
with large data sets, and software libraries. More performance at lower
electrical power is essential to allow both data center and tactical
deployments. DARPA has demonstrated analog processing of AI algorithms
with 1000 times speedup and 1000 times power efficiency over state-of-
the-art digital processors, and is researching AI-specific hardware
designs. DARPA is also attacking the current inefficiency of machine
learning, by researching methods to drastically reduce requirements for
labeled training data.
Next Generation AI: The machine learning algorithms that enable
face recognition and self-driving vehicles were invented over 20 years
ago. DARPA has taken the lead in pioneering research to develop the
next generation of AI algorithms, which will transform computers from
tools into problem-solving partners. DARPA research aims to enable AI
systems to explain their actions, and to acquire and reason with common
sense knowledge. DARPA R&D produced the first AI successes, such as
expert systems and search, and more recently has advanced machine
learning tools and hardware.
In addition to new and ongoing DARPA research, a key component of
the AI Next campaign will be DARPA's Artificial Intelligence
Exploration (AIE) program, first announced in July 2018. AIE
constitutes a series of high-risk, high payoff projects where
researchers work to establish the feasibility of new AI concepts within
18 months of award. Leveraging streamlined contracting procedures and
funding mechanisms enables these efforts to move from proposal to
project kick-off within 3 months of an opportunity announcement.
conclusion
Over its 60-year history, DARPA has made significant investments in
the creation and advancement of artificial intelligence technologies
that have produced game-changing capabilities for the Department of
Defense and beyond. DARPA's AI Next effort is simply a continuing part
of its historic investment in the exploration and advancement of AI
technologies.
Current R&D investment around the world is largely focused on
second wave AI or machine learning, which is very good in finding
patterns in voice and imagery and has many commercial applications. The
difference is, in the United States, DARPA is aggressively pursuing
programs that will make second wave AI more robust for defense and
security applications, all while helping realize the third wave of AI,
or contextual reasoning. DARPA has unique access to the United States'
world-class science and technology community, comprised of leading
universities, government labs, and industry partners--this mix cannot
be found or replicated anywhere else in the world. Marshalling those
unique resources, the Agency's third wave research efforts will forge
new theories and methods that will make it possible for machines to
adapt contextually to changing situations, advancing computers from
tools to true collaborative partners. Going forward, the agency will be
fearless about exploring these new technologies and their
capabilities--DARPA's core function--pushing critical frontiers ahead
of our nation's adversaries.
Senator Ernst. Very good. Thank you so much.
Mr. Michael Brown is the Director of the Defense Innovation
Unit, DIU, at the U.S. Department of Defense. DIU fields
leading-edge capability to the military, using commercial
technologies faster and more cost-effectively than traditional
acquisition methods. Prior to that, Mr. Brown served as a White
House Presidential Innovation Fellow at the Defense Department.
He has also worked as CEO [Chief Executive Officer] of Semantec
Corporation and as CEO of Quantum Corporation.
Thank you, Mr. Brown, for joining us today. It's good to
have you here again. If you would, please go ahead with your
opening statements.
STATEMENT OF MICHAEL A. BROWN, DIRECTOR, DEFENSE INNOVATION
UNIT
Mr. Brown. Thank you, Chairman Ernst, Ranking Member
Peters, and Members of the Subcommittee. Thank you for inviting
me here today to discuss DIU's efforts in AI.
As you said, about 6 months ago, I joined DIU as the
Director, and, having led a number of technology companies,
most recently Semantec, I've witnessed how new technology like
AI can fundamentally redefine how we live and work, and how we
fight wars.
Before joining DIU as the Director, as you mentioned, I
worked as a Fellow, responding to the Secretary of Defense's
request to understand China's investments in early-stage
technology firms, many of which were AI-focused, and its
technology transfer implications for national security.
As you mentioned in opening remarks, China and Russia have
already recognized the enormous commercial and military
potential of AI, and are investing heavily, with aims to become
dominant. By 2025, China aims to achieve major breakthroughs in
AI and increase its domestic market to reach $60 billion. To
achieve this target, the Chinese government leverages civil-
military fusion, where, by law, every commercial AI innovation
is immediately transferred to the Chinese military. China also
leverages United States talent and resources by establishing
research institutes in the United States, investing in AI-
related startups in the United States, recruiting talent in the
United States, and building academic partnerships.
Russia, as you mentioned, with Vladimir Putin's comments,
is similarly focused on building its AI capacity, but is behind
the United States and China, in terms of overall investment,
research, and startups.
In the face of great-power competition, DIU is working
alongside--with the rest of DOD to maintain our technological
edge, not only in AI, but other dual-use technologies, as well.
Accessing mature AI-driven technologies from the commercial
sector is an essential component of the Defense Department's
Artificial Intelligence Strategy and a paradigm shift from
defense industrial base to a national security innovation base
prescribed by the National Defense Strategy.
DIU's AI portfolio focuses on understanding, tracking, and
vetting commercial companies' abilities to solve high-impact
problems identified by our military leadership. AI projects
today include work with the Air Force, Army, Navy, and
components, as well as Joint Chiefs of Staff.
As a foundational technology, the DIU AI portfolio
specifically prioritizes projects that address three major
impact areas where AI has proven to excel commercially. Here
are three examples:
First, computer vision. Adding automation to object
recognition and infrastructure assessment, DIU is prototyping
computer-vision algorithms in humanitarian assistance and
disaster recovery use cases.
Second, large dataset analytics and predictions, making
sense of massive datasets and patterns more efficiently and
cost-effectively than human analysts. For example, DIU is
prototyping predictive maintenance applications for Air Force
and Army platforms, with the potential to save the Department
billions of dollars.
Third, strategic reasoning, mapping probabilistic chains of
events and developing alternative strategies to inform top-down
planning in environments characterized by uncertainty, missing
information, and speculation. DIU is prototyping an application
that leverages AI to provide insights to high-level strategic
questions.
With these projects, DIU engages across the Department on
AI and makes its commercial knowledge and relationships with
potential vendors available to any of the services, service
labs, and components. We already have in place a strategic
partnership with JAIC, which we've agreed upon with General
Shanahan. Simply stated, DIU will prototype commercially
successful AI applications and measure their relevance to
mission imperatives. If successful, we transition those to JAIC
so they can be scaled and integrated into their national
mission initiatives. We look forward to working closely
together with JAIC.
DIU also works with the Defense Innovation Board and will
work with the newly established Congressional National Security
Commission on AI to leverage the best practices and learnings
from the commercial software industry executives who
participate on that board.
Cultural divides and ethical differences are often blamed
for the lack of closer cooperation between DOD and Silicon
Valley, but, more often than not, the true deterrent is
misaligned economics. Enabling DOD to be a better customer for
early-stage companies will not only help DOD acquire the best
commercial technology faster and cheaper, but will also provide
access to the ideas of sought-after AI talent that DOD may not
be able to attract. The more we collaborate with the private
sector on mutually-beneficial projects, the more opportunities
we'll have to engage in an open dialogue about the applications
and principles for the use of AI.
DIU plans to continue its focus on AI as a key technology
portfolio, solving DOD problems with commercial AI solutions to
bring the Department new capabilities and encourage
nontraditional technology firms to work with DOD as part of the
national security innovation base, will be a priority.
Thank you.
[The prepared statement of Mr. Brown follows:]
Prepared Statement by Michael Brown
introduction
Chairman Ernst, Ranking Member Peters, and distinguished Members of
the Subcommittee on Emerging Technologies and Threats, thank you for
inviting me to appear before you today to discuss the Defense
Innovation Unit (DIU) and our efforts in artificial intelligence (AI)
alongside my colleagues at the Defense Advanced Research Projects
Agency (DARPA) and the newly formed Joint Artificial Intelligence
Center (JAIC).
AI is fundamentally redefining how we live, work, and fight wars.
Within the Department of Defense (DOD), AI has the potential to
transform how the Department operates at all levels, from business to
the battlefield. In the face of competition from China and Russia, DOD
aims to maintain its technological edge through establishing a more
decentralized, experimental procurement approach: cultivating a leading
AI workforce, engaging academic, commercial, and international allies
and partners, and developing ethical and lawful guidelines for AI use.
\1\
---------------------------------------------------------------------------
\1\ United States Department of Defense, Summary of the 2018
Department of Defense Artificial Intelligence Strategy: Harnessing AI
to Advance Our Security and Prosperity, (February 12, 2019), https://
media.defense.gov/2019/Feb/12/2002088963/-1/-1/1/SUMMARY-OF-DOD-AI-
STRATEGY.PDF.
---------------------------------------------------------------------------
China and Russia have recognized the enormous commercial and
military potential of AI and are investing heavily to become dominant
in the field. In its 13th Five-Year Plan (2016-2020) and subsequent
industrial plans, the Chinese Government has outlined a comprehensive,
whole-of-government strategy to become the global leader in AI. \2\ In
July 2017, the State Council released the Next-Generation Artificial
Intelligence Development Plan that laid out a 2020 target for Chinese
AI technology and applications to match international developments and
create a $22.3 billion (Renminbi [RMB] 150 billion) \3\ domestic
market. \4\ By 2025, China will aim to achieve major breakthroughs in
AI and increase its domestic market to reach $59.6 billion (RMB 400
billion). \5\ To achieve these targets, China's National Development
and Reform Commission (China's industrial policy-making agency) funded
the creation of a national AI laboratory, and Chinese local governments
have pledged more than than $7 billion in AI funding. \6\ In addition,
Chinese firms and the Chinese Government are leveraging United States
talent and ecosystems through the establishment of research institutes
in the United States, investment in U.S. AI-related startups and firms,
recruitment of U.S.-based talent, and commercial and academic
partnerships. \7\ Russia is similarly focused on building its AI
capacity but is behind the United States and China in terms of overall
AI investment, research, and startups. \8\
---------------------------------------------------------------------------
\2\ Central Committee of the Communist Party of China Central
Compilation and Translation Press, The 13th Five-Year Plan for Economic
and Social Development of the People's Republic of China (2016-2020),
(March 17, 2016), 64.
\3\ For this testimony, the exchange rate is: $1 = 6.72 RMB.
\4\ PRC State Council, Xinyidai rengongzineng fazhan guihua de
tongzi [Next-Generation Artificial Intelligence Development Plan], PRC
State Council-2017-35 (July 20, 2017).
\5\ Ibid.
\6\ ``2017 Annual Report to Congress,'' (United States-China
Economic and Security Review Commission, November 15, 2017), 525-527;
Michael Brown and Pavneet Singh, ``China's Technology Transfer
Strategy: How Chinese Investments in Emerging Technology Enable a
Strategic Competitor to Access the Crown Jewels of United States
Innovation,'' (Defense Innovation Unit, January 2018).
\7\ Ibid.
\8\ Alina Polyakova, ``Weapons of the Weak: Russia and AI-driven
Asymmetric Warfare,'' (Brookings Institution, November 14, 2018);
``Artificial Intelligence -A Strategy for European Startups:
Recommendations for Policymakers,'' (Asgard and Roland Berger, May 14,
2018).
---------------------------------------------------------------------------
Underscoring the potential magnitude of AI's impact on the whole of
society, the breadth of its applications, and the urgency of this
emerging technology race, President Trump signed the executive order,
Maintaining American Leadership in Artificial Intelligence, on February
11, 2019, launching the American AI Initiative. This was immediately
followed by the release of DOD's first-ever AI strategy. \9\ These
documents emphasize the essential role of research and development
(R&D) across the Federal Government, business, and academia to maintain
U.S. leadership in AI, bolster national security, and safeguard the
values shared by the United States, its allies, and partners.
---------------------------------------------------------------------------
\9\ Maintaining American Leadership in Artificial Intelligence,
Exec. Order No. 13859, 84 Fed. Reg. 3967 (February 11, 2019). U.S.
Department of Defense, Summary of the 2018 Department of Defense
Artificial Intelligence Strategy: Harnessing AI to Advance Our Security
and Prosperity.
---------------------------------------------------------------------------
To increase intergovernmental coordination, DIU will engage with
DARPA and JAIC, among other DOD entities focused on AI, as well as make
its commercial knowledge and relationships with potential vendors
available to any of the Services and Service Labs. For example, DIU
will be working with the Services and Defense Agencies as DOD customers
for the projects it undertakes. AI projects today include work with the
Air Force, Army, Navy, and components as well as the Joint Chiefs of
Staff. DIU also works with the Defense Innovation Board and the newly
established Congressional Commission on AI to leverage the best
practices and learnings from the commercial software industry
executives who participate on the Board.
In particular, we anticipate a close partnership with JAIC, the
outlines of which DIU has already agreed upon with Lieutenant General
Jack Shanahan. As JAIC matures, we anticipate that DIU will be at the
leading edge of the Department's National Mission Initiatives (NMIs),
proving that commercial technology can be applied to critical national
security challenges via accelerated prototypes that lay the groundwork
for future scaling through JAIC. DIU looks to bring in key elements of
AI development pursued by the commercial sector, which relies heavily
on continuous feedback loops, vigorous experimentation using data, and
iterative development, all to achieve the measurable outcome, mission
impact.
reinvigorating outreach to commercial technology companies: defense
innovation unit
DIU is focused on accelerating commercial technology into the hands
of men and women in uniform. Its staff is comprised of Active Duty
military from every service, civilians, and individuals with extensive
private sector experience and deep ties into venture capital and
startup communities. DIU partners with the Services, Combatant
Commands, and component organizations to seek out and rapidly prototype
advanced commercial solutions--spanning AI, autonomy, cyber, human
systems, and space--to address military challenges ranging from the
tactical level to the defense enterprise. Within OUSD(R&E) and the
broader DOD, DIU is unique in its focus on developing and fielding
commercial hardware, software, and methodologies within an
approximately 24-month timeframe.
Accessing R&D and mature AI-driven technologies advanced by the
commercial sector is an essential component of the strategic approach
defined by the 2018 Department of Defense Artificial Intelligence
Strategy and the paradigm shift from ``defense industrial base'' to
``national security innovation base'' prescribed by the 2018 National
Defense Strategy. Senior leaders in the Department understand that DOD
no longer holds a monopoly on emerging technologies like AI that will
sway strategic, deterrent, and battlefield advantage in future wars.
\10\ U.S. businesses began outspending the Federal Government in R&D in
the 1980s, and now, industry-funded R&D represents approximately 67
percent of total U.S. investments. \11\
---------------------------------------------------------------------------
\10\ Ben FitzGerald, Alexandra Sander, Jacqueline Parziale,
``Future Foundry: A New Strategic Approach to Military-Technical
Advantage,'' (Center for a New American Security, December 2016).
\11\ 11 DIU exists, in part, in response to the growing disparity
between federal and commercial R&D, leading to a global technology
landscape in which commercial companies are leading the development of
some of the world's most advanced technologies: ``In 1960, the United
States accounted for 69 percent of global R&D, with U.S. defense-
related R&D alone accounting for more than one-third of global R&D. The
Federal Government funded approximately twice as much R&D as U.S.
business. However, from 1960 to 2016, the U.S. share of global R&D fell
to 28 percent, and the Federal Government's share of total U.S. R&D
fell from 65 percent to 24 percent, while business's share more than
doubled from 33 percent to 67 percent. As a result of these global,
national, and federal trends, federal defense R&D's share of total
global R&D fell to 3.7 percent in 2016.'' Moshe Schwartz and Heidi M.
Peters, ``Department of Defense Use of Other Transaction Authority:
Background, Analysis, and Issues for Congress,'' Report no. R45521
(Congressional Research Service), 43.
---------------------------------------------------------------------------
Moreover, venture capital funding for AI-related companies reached
record highs in 2018, increasing 72 percent from 2017 totaling $9.3
billion. \12\ With offices in Silicon Valley, Boston, Austin, and
Washington, D.C., DIU is embedded in the core innovation ecosystems
where these deals are taking place, AI startups are thriving, and top
tech companies and universities are conducting groundbreaking research.
\13\ DIU's location not only facilitates deeper ties with leading edge
companies but allows the Department to establish a closer relationship
with venture firms as they scout the horizon for their next big bets
and take into consideration clear demand signals from DOD.
---------------------------------------------------------------------------
\12\ ``MoneyTree Report: Q4 2018,'' (PricewaterhouseCoopers and CB
Insights, 2019), https://www.pwc.com/us/en/moneytree-report/moneytree-
report-q4-2018.pdf.
\13\ The top five states for AI investment in 2018, in order, were
California, Massachusetts, New York, Texas, and Washington. Ibid.
---------------------------------------------------------------------------
DIU seeks to lower barriers to entry into the defense market by
more closely matching commercial terms and contracting speeds via its
Commercial Solutions Opening (CSO) solicitation process, which
leverages Other Transaction (OT) authority. Traditional acquisition
pathways overburden technology companies operating with little or no
prior DOD contracting experience and runways that are often shorter
than the typical time to award a contract under the Federal Acquisition
Regulation. Shaping the DOD into a better customer through new
processes allows the Department to acquire the best commercial
technology faster and cheaper than the traditional system. Furthermore,
new acquisition pathways create more opportunities for national
security service, making DOD a more competitive employer of AI and
other sought-after tech talent through commercial contracts.
While cultural divides and ethical differences are often blamed for
the lack of closer cooperation between DOD and Silicon Valley, more
often than not, the true deterrent is misaligned economics. \14\ Since
DIU opened its first competitive solicitation using the CSO process in
June 2016, there has been no shortage of top-performing companies
interested in working alongside our DOD partners to solve some of the
toughest military challenges. DIU has awarded contracts to 103 of these
companies, 43 of which are first-time, non-traditional DOD contractors.
\15\
---------------------------------------------------------------------------
\14\ Rachel Olney, ``The Rift Between Silicon Valley and the
Pentagon is Economic, not Moral,'' War on the Rocks, January 28, 2019,
https://warontherocks.com/2019/01/the-rift-between-silicon-valley-and-
the-pentagon-iseconomic-not-moral/.
\15\ The 2018 OT Guide defines non-traditional defense contractor
as ``an entity that is not currently performing and has not performed,
for at least the one-year period preceding the solicitation of sources
by DOD for the procurement or transaction, any contract or subcontract
for the DOD that is subject to full coverage under the cost accounting
standards prescribed pursuant to section 1502 of title 41 and the
regulations implementing such section (see 10 U.S.C. 2302(9).'' Defense
Acquisition University, Other Transactions (OT) Guide, (November 2018),
https://aaf.dau.mil/ot-guide/.
---------------------------------------------------------------------------
diu's ai strategy & projects
Commercial AI companies are active across a wide range of sectors
and the opportunities for dual-use applications within DOD are vast.
The DIU AI portfolio focuses on understanding, tracking, and vetting
these commercial companies' ability to solve high-impact problems
identified by our military leadership and DOD partners. The portfolio
team combines depth of commercial AI, machine learning, and data
science experience from the commercial sector with military operators.
As a foundational technology, AI-driven solutions appear across a
number of DIU projects administered by other portfolio teams, however,
the AI portfolio specifically prioritizes projects that address three
major impact areas where AI is proven to excel:
1. Computer vision: AI and machine learning adds automation to
object recognition and infrastructure assessment; for example, DIU is
prototyping computer vision algorithms in humanitarian assistance and
disaster recovery scenarios.
2. Large dataset analytics and predictions: AI and machine
learning can help make sense of massive datasets and patterns more
efficiently and cost-effectively than human analysts; for example, DIU
is prototyping predictive maintenance applications for Air Force and
Army platforms.
3. Strategic reasoning: AI and machine learning has the capacity
to inform top-down planning in environments characterized by
uncertainty, missing information, and speculation; for example, DIU is
prototyping an application that leverages AI to reason about high-level
strategic questions, map probabilistic chains of events, and develop
alternative strategies.
Furthermore, DIU has a strategic partnership with JAIC wherein
DIU's prototype AI applications ``pull'' on commercial capabilities,
prove and measure their applicability to mission imperatives, and (if
successful) are transitioned to JAIC to be scaled and integrated into
their NMI. In previous testimony before the House Armed Services
Subcommittee on Emerging Threats and Capabilities, Dr. Lisa Porter,
Deputy Under Secretary of Defense for Research and Engineering,
discussed the need to rigorously assess AI performance against
quantitative metrics tied to specific mission needs. \16\ DIU's
partnership with JAIC aims to institutionalize the rigor Dr. Porter
spoke of----the AI portfolio's prototype projects are designed to drive
metrics, establish benchmarks, and contribute infrastructure towards a
common foundation as described by the 2018 Department of Defense
Artificial Intelligence Strategy. \17\
---------------------------------------------------------------------------
\16\ Dr. Lisa Porter, Deputy Under Secretary of Defense for
Research and Engineering, testimony to the Subcommittee on Emerging
Threats and Technologies, House Armed Services Committee, December 11,
2018, https://docs.house.gov/meetings/AS/AS26/20181211/108795/HHRG-115-
AS26-Wstate-PorterL-20181211.pdf.
\17\ U.S. Department of Defense, Summary of the 2018 Department of
Defense Artificial Intelligence Strategy: Harnessing AI to Advance Our
Security and Prosperity.
---------------------------------------------------------------------------
Following are three specific use cases and projects which employ AI
technology:
Applying Computer Vision to Humanitarian Assistance and Disaster Relief
In 2018, DIU hosted the xView Challenge to test computer vision and
the use of algorithms to automatically identify objects from images.
The competition attracted more than 4,000 submissions from 100
participants from around the world including companies, universities,
and individuals. The top performing algorithms were 300 percent more
accurate than the government produced baseline, which helps advance
computer vision proficiency across four core elements of overhead
imagery analysis. The winning algorithm was then used to automate post-
disaster assessments in the wake of Hurricane Florence, assisting
emergency personnel to quickly identify flooded areas and impassable
roads. This use of AI holds the potential to automate post-disaster
assessments and accelerate search and rescue efforts on a global scale.
Scaling Predictive Maintenance to Improve Readiness and Cut Costs
DIU's predictive maintenance prototype project provides a specific
example of the synergy that we plan to foster between OUSD(R&E) and
JAIC. DIU identified a leading commercial airline industry supplier of
predictive maintenance solutions and launched a six-month prototype for
E-3 Sentry aircraft maintenance. The prototype began with testing
predictions at the part and sub-part level against historical actuals
to establish the robustness of the AI and its relevance to operational
decision-making. This methodology effectively assesses the accuracy of
the AI predictions, how much they matter, and in which areas the most
impact can be expected (as defined by cost and/or platform
availability). Early results of Air Force applications indicate a
potential 28 percent decrease in unscheduled maintenance on the E-3
across six sub-systems and more than 32 percent reduction on the C-5
across ten sub-systems. DIU is partnering with JAIC to scale this
solution across multiple aircraft platforms, as well as ground vehicles
beginning with DIU's complementary predictive maintenance project
focusing on the Army's Bradley Fighting Vehicle. This is one of DIU's
highest priority projects for fiscal year 2019 given its enormous
potential for impact on readiness and reducing costs.
Automating Cyber Vulnerability Detection & Remediation
DOD's current vulnerability discovery process for weapons systems
software lacks the capability to scale because it relies on time and
labor-intensive human search and analysis. According to an October 2018
GAO report, $1.66 trillion of weapon system development is at risk due
to the scale of unmitigated cyber vulnerabilities. \18\ One of the
tools to address these vulnerabilities is DIU's Project VOLTRON, which
is an active prototype project that has demonstrated artificially
intelligent detection of previously unknown vulnerabilities in
classified weapons systems. The project seeks to demonstrate autonomous
exploitation and patching; development of an application programming
interface (API) for extensibility; and integration into DOD software
development environments. This would give the DOD an end-to-end
capability that goes from writing software free of vulnerabilities to
remediating vulnerabilities in compiled mission software for which
source code is not available. The products from Project VOLTRON help
make DOD owned systems more resilient to cyber attacks and inform
program offices of configuration errors faster and with less errors
than humans. An initial capability demonstration of the commercial
technologies leveraged by VOLTRON yielded previously undiscovered bugs
within the first few minutes of testing against representative aircraft
software provided by a defense contractor. In addition, previously
unknown vulnerabilities have already been discovered in currently
fielded aircraft systems. Integration into software development
pipelines will ensure that most vulnerabilities can be found and
remediated before future systems go into production and/or deployment.
---------------------------------------------------------------------------
\18\ United States Government Accountability Office, Weapon Systems
Cybersecurity: DOD Just Beginning to Grapple with Scale of
Vulnerabilities, GAO-19-128 (October 2018), https://www.gao.gov/assets/
700/694913.pdf.
---------------------------------------------------------------------------
Tremendous Opportunity for DOD/Commercial Collaboration
Commercial industry is breaking ground on AI applications
supporting a wide range of business areas and there is a tremendous
opportunity to re-establish and grow the ties between the user
communities in DOD, commercial entrepreneurs, and partners in
universities and labs dedicated to performing the basic research that
provides a foundation for future advances. While DIU has found the vast
majority of high-tech companies focused on AI to be willing and
enthusiastic partners, there is work yet to be done to provide and
encourage an open dialogue with the private sector and researchers
about applications and principles of use for this powerful tool. DIU
will continue to solve DOD problems with commercial AI solutions to
bring the Department new capabilities and encourage non-traditional
technology firms to work with DOD to grow the national security
innovation base.
Senator Ernst. Thank you very much, Mr. Brown.
Last, certainly not least, we have Lieutenant General John
N.T. ``Jack'' Shanahan. General Shanahan is the Director, Joint
Artificial Intelligence Center, Office of the Department of
Defense Chief Information Officer (CIO) at the Pentagon.
General Shanahan is responsible for accelerating the delivery
of AI-enabled capabilities, scaling the departmentwide impact
of AI, and synchronizing AI activities to expand joint force
advantages.
General, please go ahead.
STATEMENT LIEUTENANT GENERAL JOHN N.T. SHANAHAN, USAF,
DIRECTOR, JOINT ARTIFICIAL INTELLIGENCE CENTER, OFFICE OF THE
DEPARTMENT OF DEFENSE CHIEF INFORMATION OFFICER
General Shanahan. Good afternoon, Madam Chairwoman, Ranking
Member, distinguished Members of the Subcommittee. Thank you
for the opportunity to testify before the Subcommittee today on
the Department's artificial intelligence committees.
I'm honored to serve as the first Director of DOD's Joint
AI Center, or the JAIC. I've been in this position for just
over 2 months. Previously, I served in the Under Secretary of
Defense for Intelligence, where, for 2 years, I was the
Director of the Algorithmic Warfare Cross-functional Team, also
known as Project Maven.
Artificial intelligence, or AI, is rapidly changing an
ever-expanding range of businesses and industries. It offers
the opportunity to transform every corner of the Department
from multidomain operations at the edge to back-office business
functions. As described in the 2019 National Defense Strategy,
or NDS, it is also poised to change the character of warfare.
Thoughtful, responsible, and human-centered adoption of AI in
the DOD will strengthen our national security and transform the
speed and agility of our operations.
Last June, then-Deputy Secretary of Defense Shanahan
directed the DOD Chief Information Officer, Mr. Dana Deasy, to
establish the Joint AI Center to accelerate the delivery and
adoption of AI-enabled capabilities, scale the departmentwide
impact of AI, and synchronize the Department's AI activities.
In parallel, DOD submitted its first AI Strategy to the
Congress as an annex to the NDS. Last month, the Department
released an unclassified summary of DOD's AI Strategy, doing so
on the heels of the President's signature of the executive
order on AI. The JAIC's missions and functions nest well under
the principles and objectives outlined in the AI executive
order.
JAIC's formation also dovetails section 238 of the fiscal
year 2019 NDAA. Additionally, JAIC will benefit from, and help
bring to fruition, recommendations of the new National Security
Commission (NSC) on AI. I was privileged to talk with the
members of the Commission yesterday, when they met for the
first time. I know Senator Heinrich was also there.
The JAIC is the focal point of the DOD AI Strategy and was
established to provide a common vision, mission, and focus to
drive departmentwide AI capability delivery. I want to
highlight three primary themes for our approach:
First, delivering AI-enabled capabilities at speed. JAIC is
collaborating with teams across DOD to identify, prioritize,
and select mission needs, and then execute a series of cross-
functional use cases to demonstrate value and spur momentum. We
need early demonstrable wins to show practical results and the
art of the possible, followed by scaling across the enterprise.
Projects fall into two main categories: national mission
initiatives, or NMIs, and component mission initiatives, or
CMIs. NMIs are driven and executed by the JAIC as broad, joint,
crosscutting AI challenges; whereas, CMIs are component-led,
but are able to make use of JAIC's common tools, libraries,
best practices, and more.
Our emphasis on rapid, iterative delivery of AI complements
the Department's ongoing work at the other end of the AI
spectrum, in fundamental research and development, as you heard
from Dr. Highnam. Our first two NMIs are predictive maintenance
with the Special Operations Command and U.S. Army H-60
helicopter use case and humanitarian assistance and disaster
relief, in which we will field AI capabilities in support of
natural events, such as wildfires and hurricanes. We are also
getting a headstart on a planned fiscal year 2020 cyberspace
NMI designed to use AI-enabled capabilities to improve event
detection, network mapping, and compromised-account
identification.
At the same time, we are now in the early problem-framing
stage for another proposed NMI in fiscal year 2020 that will be
more oriented on the NDS, National Defense Strategy, in
operations against peer competitors. We are also in initial
discussions with the Military Services, components, and
combatant commands on the applicability of AI to help with
solutions in areas as diverse as talent management, suicide
prevention, preventive medicine, and information operations,
among others.
The second theme is scale. As I know firsthand from Project
Maven, scaling AI across the enterprise is hard, but it's also
the only way we will realize the full benefits of AI in the
Department. JAIC's early projects serve a dual purpose, to
deliver new capabilities to end users as well as to
incrementally develop the common foundation that is essential
for scaling AI's impact across DOD. We will put this foundation
in place in a way that aligns with DOD enterprise cloud
adoption.
The third theme is talent. We built the initial JAIC team
with representatives detailed from across each of the services
and other components. Today, we have 30 people, growing to over
50 within the next 5 months. We do not receive our permanent
manpower until fiscal year 2020.
For the JAIC to succeed, we must attract and cultivate a
select group of mission-driven, world-class AI talent, to
include enticing experts from the tech industry to serve with
us. The success of human-centered AI and human-machine teaming
within DOD requires growing and sustaining an AI-ready force,
one that is conversant in the language of AI, willing and able
to operate with a new kind of speed and agility.
In closing, the JAIC is now up and running, and we're open
for business. Thank you for your strong support in driving
momentum in this critical area. I look forward to continuing to
work with Congress as we advance the adoption of AI across the
Department and use the JAIC to accelerate our progress.
Thank you for the opportunity to testify this afternoon. I
look forward to your questions.
[The prepared statement of General Shanahan follows:]
Prepared Statement by Lieutenant General John ``Jack'' N.T. Shanahan
introduction
Good afternoon Madam Chairwoman, Ranking Member, and distinguished
Members of the Subcommittee. Thank you for this opportunity to testify
before the Subcommittee today on the Department's Artificial
Intelligence (AI) Initiatives.
I am Lieutenant General Jack Shanahan, the Director of the Joint
Artificial Intelligence Center or JAIC. I have been in my current
position for a little over two months. Previously, I served in the
Under Secretary of Defense for Intelligence as the Director of the
Algorithmic Warfare Cross-Functional Team or Project Maven, the
Department's pathfinder project to integrate AI capabilities to
augment, accelerate, and automate collection from a variety of manned
and unmanned intelligence platforms and sensors.
AI is rapidly changing an ever-expanding range of business and
industry. As described in the 2018 National Defense Strategy (NDS), AI
is also poised to change the character of warfare. Structurally, we
know AI has the potential to be an enabling layer across nearly
everything--meaning countless applications in industry and everyday
life, while offering the opportunity to positively transform every
corner of the Department. We envision innovative concepts that change
the way we plan and fight, including improvements in the way we
perceive our environment, maintain our equipment, train our men and
women, defend our networks, operate our back offices, provide
humanitarian aid and respond to disasters; and more. By harnessing the
power of AI in defense, we will better support and protect American
servicemembers, safeguard our citizens, defend our allies, and improve
the effectiveness, affordability, and speed of our operations.
Other nations, particularly strategic competitors such as China and
Russia, are making significant investments in AI for military purposes.
These investments threaten to erode our technological and operational
advantages and destabilize the free and open international order. The
Department of Defense, together with our allies and partners, must
adopt AI to maintain its strategic position, prevail on future
battlefields, and safeguard this order.
Per the NDS, the Department will accelerate the delivery and
adoption of AI to expand our military advantages and create a force fit
for our time. AI will enhance operational effectiveness, improve
readiness, and increase efficiency in the general business practices of
the Department. We will make a concerted effort to move AI technologies
in a direction that improves our odds of long-term security, peace, and
stability through vigorous dialogue and multilateral cooperation on the
ethical, safe, and lawful use of AI for national security and
establishing new norms for responsible behavior, consistent with the
law. The Department's AI transformation will ensure that we maintain
the ability to execute our vital mission of protecting the security of
our nation, deterring war, and preserving peace.
establishment of jaic
Last June, then-Deputy Secretary of Defense Patrick Shanahan
directed Mr. Dana Deasy, the Department's Chief Information Officer, to
establish the Joint AI Center. This new organization is tasked to
accelerate the delivery of AI-enabled capabilities, scale the
Department-wide impact of AI, and synchronize the Department's AI
activities. In parallel, the Department submitted its first AI Strategy
to Congress, an annex to the NDS that captures the integrated set of
decisions we are making now to harness AI to advance our security and
prosperity. Last month, the Department released an unclassified summary
of the classified DOD AI strategy, in support of the President's
Executive Order on AI (Maintaining American Leadership in Artificial
Intelligence) that calls for greater AI investment, harmonization of
standards, and training and workforce development initiatives. The
JAIC's missions and functions nest well under the principles and
objectives outlined in the AI Executive Order.
The founding of JAIC supports implementation of section 238 of the
fiscal year 2019 National Defense Authorization Act, this provision
directed a joint approach to coordinate the efforts of the Department
to develop, mature, and transition AI technologies into operational
use. The Department views the requirements of section 238 as a
strategic opportunity to improve its posture for AI. In fact, we used
elements from the language in section 238 to help frame the JAIC's
roles, missions, and functions. In December 2018, JAIC commissioned a
team from the RAND Corporation to support our analysis. The RAND team
built its analytical framework, completed initial DOD-wide data
collection, and is currently building interview protocols and contact
lists for engaging with industry.
As part of this, I will now touch on how we are partnering with the
Under Secretary of Defense (USD) Research & Engineering (R&E), the role
of the Military Services, the Department's initial focus areas for AI
delivery, and how JAIC is supporting whole-of-government efforts in AI.
As the focal point of the DOD AI Strategy, the JAIC was established
to provide a common vision, mission, and focus to drive Department-wide
AI capability delivery. JAIC will operate across the full AI delivery
lifecycle, emphasizing near-term prototyping, execution, and
operational adoption to meet current needs. JAIC's work will complement
the AI efforts of USD(R&E), which are focused on foundational research,
longer-term technology creation, and innovative concepts. Both JAIC and
USD(R&E) will need to collaborate effectively and succeed individually
for the DOD to implement its ambitious AI strategy.
The JAIC communicates a consistent message about transforming DOD
through AI. This refers to the transformation that happens when you
field technology on operationally-relevant timelines, enable frontline
men and women to experiment with it based on their own creativity, and
ultimately generate new ways of working that solve our most critical
challenges and enhance our military strength. As we move to rapidly
incorporate AI, those men and women in America's military will remain
our enduring source of strength. We will use AI-enabled information,
tools, and systems to empower and augment, not replace, those who
serve.
To derive maximum value from AI application throughout the
Department, JAIC will operate across an end-to-end lifecycle of problem
identification, prototyping, integration, scaling, transition, and
sustainment. Emphasizing commerciality to the maximum extent
practicable, JAIC will partner with the Services and other components
across the Joint Force to systematically identify, prioritize, and
select new AI mission initiatives. Then JAIC will stand up cross-
functional teams that will rapidly execute a sequence of use cases that
demonstrate value and spur momentum. We need early, demonstrable wins
that show practical results and the art of the possible. Then, we must
scale these capabilities across the enterprise. To do this, JAIC is
engaging with leading commercial and academic partners for prototypes,
and employing standardized processes with respect to areas such as data
management, testing and evaluation, assessment of delivered
capabilities, and program protection and cybersecurity. Our approach
has been directly informed by the Department's AI pathfinder activity,
Project Maven, which successfully identified and is beginning to
address key challenges with integrating AI into operations. This
program put in place an initial set of data, tools, and infrastructure
for AI delivery, as well as initial templates for contracting and
acquisition, testing and evaluation, operational assessment, and
program protection.
JAIC's early projects serve a dual purpose: to deliver new AI-
enabled capabilities to end users, and to help incrementally develop a
common foundation that is essential for scaling AI's impact across DOD.
This foundation includes shared data, reusable tools, frameworks,
libraries, and standards, and AI cloud and edge services. JAIC will
work with teams throughout the Department to ensure that they can
leverage this foundation to accelerate their progress in a manner that
aligns with DOD enterprise cloud adoption. Our enterprise approach for
AI and enterprise cloud adoption as outlined in the DOD-wide Cloud
Strategy are mutually reinforcing, mutually dependent undertakings.
Finally, JAIC will provide ongoing support to the efforts of the
Services and other organizations to ensure continuous improvement,
assessment, and sustainment of AI systems and solutions across the
enterprise.
The AI capability delivery efforts that will go through this
lifecycle will fall into two categories: National Mission Initiatives
(NMI) and Component Mission Initiatives (CMI). As outlined in the DOD
AI Strategy, a NMI is a pressing operational or business reform joint
challenge, typically identified from the National Defense Strategy's
key operational problems or nominated by a mission owner, and requiring
multi-service innovation, coordination, and the parallel introduction
of new technology and new operating concepts. NMIs are typically driven
by JAIC and are executed by cross-functional teams that comprise both
JAIC personnel as well as subject matter specialists from across the
Department on a rotational basis. Execution of these projects will be
essential for putting in place our initial common foundation.
The second project category is a Component Mission Initiative
(CMI), which is a component-level challenge that can be solved through
AI. JAIC will work closely with individual components on CMIs to help
identify, shape, and accelerate their Component-specific AI deployments
through funding support; usage of common foundational tools, libraries,
cloud infrastructure; application of best practices; partnerships with
industry and academia; and so on. The Component will be responsible for
identifying and implementing the organizational structure required to
accomplish its project in coordination and partnership with the JAIC.
We will form teams to work with the Services, Components, and
Combatant Commands on potential CMIs. Based on initial conversations
with all of these stakeholders, I fully expect that we will see rapid
growth in the number of CMIs in Fiscal Year 2020. We are in early
discussions with the Services, Components, and Combatant Commands on
the applicability of AI to help with solutions in areas as diverse as
talent management, suicide prevention, preventive medicine,
installation and force protection, information operations, operational
war planning, and modeling and simulation. Additionally, we intend to
identify smart automation initiatives that could provide near-term
dividends in terms of increased effectiveness and efficiency for back-
office functions.
All of the Services are increasing their levels of investment in
AI-related capabilities in near term. The JAIC is already forming
strong partnerships with the Services and key Components. For example,
the Army established a new AI Task Force that is working closely with
the JAIC on predictive maintenance. We are actively engaged in an
effort to apply data-driven insights to equipment availability at U.S.
Special Operations Command and in the U.S. Air Force in partnership
with Defense Innovation Unit (DIU). We are partnering with U.S. Cyber
Command and the National Security Agency to shape a new cyberspace-
related mission initiative. These early efforts will better define how
we make use of common approaches to data, tools, libraries,
architectures, development approaches, and more.
JAIC's focus on near-term AI implementation and adoption
complements efforts within the USD (R&E). Organizations such as the
Defense Advanced Research Projects Agency (DARPA) are focused on the
future or next wave of AI research and longer-term technology creation.
When it comes to research for the future versus the ability to apply it
now at scale, DOD needs the best of both, and they feed one another--
USD(R&E) will feed JAIC with updates on leading-edge AI technologies
and concepts, and JAIC will provide R&E insights from operational
fielding, user feedback, and data. There is a distinct and shared
vision of an enterprise approach promulgated by USD(R&E) and DOD CIO.
JAIC is already working with DIU, DARPA, and the Strategic Capabilities
Office to improve integration and enhance unity of effort on current
and future AI projects.
Further examples of early NMI's include:
Perception. Improve the speed, completeness, and accuracy
of Intelligence, Surveillance, Reconnaissance (ISR) Processing,
Exploitation, and Dissemination (PED). Project Maven's efforts are
included here. -Predictive Maintenance (PMx). Provide computational
tools to decision makers to help them better forecast, diagnose, and
manage maintenance issues to increase availability, improve operational
effectiveness, and ensure safety, at reduced cost.
Humanitarian Assistance/Disaster Relief (HA/DR). Reduce
the time associated with search and discovery, resource allocation
decisions, and executing rescue and relief operations to save lives and
livelihood during disaster operations.
Cyber Sensemaking. Detect and deter advanced adversarial
cyber actors who infiltrate and operate within the DOD Information
Network (DODIN) to increase DODIN security, safeguard sensitive
information, and allow warfighters and engineers to focus on strategic
analysis and response.
We selected these initiatives to deliver mission impact at speed,
demonstrate the proof of concept for the JAIC operational model, enable
rapid learning and iterative process refinement, and build out our
library of reusable tools while validating our enterprise cloud
architecture. These efforts will benefit us by growing more AI
credibility and expertise within the JAIC that will return to the
Services and Components to help accelerate and sustain their own AI
projects.
For the predictive/preventive maintenance NMI, we are starting with
Army and Army Special Operations helicopters (H-60s). There is
sufficient data available to train algorithms, there will be defined
return on investment criteria, and this project helps address the
Secretary's direction to the Services to improve their maintenance
readiness rates. We anticipate moving to other airframes and vehicles,
to include working with DIU to scale the promising results they have
demonstrated using AI for predictive maintenance on other Air Force and
Army platforms.
For the humanitarian assistance and disaster relief (HA/DR) NMI, we
are already applying lessons learned and reusable tools from Project
Maven to field AI capabilities in support of federal responses to
events such as wildfires and hurricanes--where DOD plays a supporting
role. One of the most important benefits of this NMI is that it is an
inspiring, societally-beneficial, life-saving mission that is not only
whole-of-government but whole-of-society. It brings in interagency,
state and local governments, non-governmental organizations, allied and
partner nations, and more. It offers a unique opportunity to combine
DOD efforts with industry and academia in a new type of public-private
endeavor to operationalize AI to solve our most challenging problems.
Doing this at scale to address disasters on an integrated basis creates
the potential to both save lives and livelihood as well as advance
common tools, lessons, and partnerships for the benefit of many DOD
missions.
We are also in the early problem-framing stage for another
substantial NMI in Fiscal Year 2020 that will be much more oriented on
the National Defense Strategy and operations against peer and near-peer
competitors. At the same time we will be seeking cutting-edge
technologies within commercial industry and in DOD organizations such
as DARPA that are ready for operational fielding across the Department.
While its primary focus is delivery initiatives such as these, JAIC
has an important role in synchronizing DOD AI activities. This avoids
duplication and excess cost, fosters sharing of lessons, and
establishes a new enterprise approach for translating AI into decisions
and impact at scale across the Joint Force. Under the DOD CIO's
authorities and as delineated in the JAIC establishment memo, JAIC will
coordinate all DOD AI-related projects above $15 million annually. This
does not mean that JAIC will control the execution of these projects or
the funding for Service-and Component-level AI initiatives. It does
mean that we will start to ensure, for example, that they begin to
leverage common tools and libraries, manage data using best practices,
reflect a common governance framework, adhere to rigorous testing and
evaluation methodologies, share lessons learned, and comply with
architectural principles and standards that enable scale. Over time,
when properly resourced, JAIC will assume a greater role with regard to
Component AI programs.
JAIC will be a key resource for whole-of-government efforts in AI,
particularly as we explore as a nation the opportunities and challenges
associated not merely with fundamental AI research, but also with
translating the technology into decisions and impact in operations. To
underscore our focus on ethics, humanitarian considerations, and both
short-term and long-term AI safety, JAIC is working closely with the
Defense Innovation Board (DIB) to foster a broad dialogue and provide
input into the development of AI principles for defense. We are
offering our perspective on crucial policy and research and development
associated with operationalizing AI today in our engagements with the
important work of the National Security Council Staff and the National
Science and Technology Council Select Committee on AI. This remains a
larger Administration priority. On February 11, 2019, President Trump
signed an executive order launching the American AI Initiative, a whole
of government strategy for ensuring American leadership in this
important field. I want to emphasize the importance of our partnerships
with Congress in all areas, but with a particular focus on AI. The
establishment of the National Security Commission on Artificial
Intelligence in the National Defense Authorization Act for Fiscal Year
2019 is one key example of this partnership, to which JAIC will serve
as the DOD liaison element.
The ingredients for JAIC's success include: enterprise cloud
adoption; world-class AI talent, particularly in areas that are scarce
within DOD today such as data science and data engineering, machine and
reinforcement learning, and product management; a workforce that is
taking steps to become broadly AI-ready; strong partnerships with the
Services, Combatant Commands, and other key components; a tight two-way
integration with the critical work of USD(R&E); and energetic, combined
problem-solving enabled by bonds of trust with AI leaders in industry
and academia. The final ingredient for success in cultivating and
sustaining an ``AI Ready'' force for the future is culture:
specifically, the need to become a more data-centric, computer science-
literate, force conversant in the language of AI, and willing and able
to operate with a new kind of speed and agility. Finally, an unwavering
commitment to ethics and principles. These are the table stakes in AI.
DOD's legacy culture and processes are particularly apparent in the
challenges we encounter launching what can only be described as a
startup within the Department of Defense. As we do so, we are
incorporating lessons learned from other Department activities that
resembled startups in how they responded to urgent, compelling
requirements across the Department--such as the Intelligence,
Surveillance and Reconnaissance Task Force, Joint Improvised Explosive
Device Defeat Organization, and Project Maven. As we learned with
Project Maven, there is no substitute for simply embarking on an AI
project to gain critical hands-on experience, but we also acknowledge
the importance of implementing more systemic AI education and training
programs across the entire Department, at all levels. The Defense
Innovation Board has been particularly helpful in charting a path
forward in this area.
All of this requires striking the right balance between top-down
pressure and bottom-up innovation. Adding funding and people will not
by themselves spark the necessary level of institutional change, at
least not until we have a broader and deeper foundation of people--
especially within all of the military Services--who understand how to
operationalize and accelerate the AI pipeline.
AI will change the character of warfare, which in turn will drive
the need for wholesale changes to doctrine, concept development, and
tactics, techniques, and procedures. There will be a need for much more
experimentation, at every level and in every domain. New operating
concepts will depend on a greater understanding of what AI can (and
cannot) help achieve. We need to accelerate fielding AI capabilities
across the joint force, and as we do so, we must validate, refine, and
adapt operating concepts. This includes thinking about entirely new
concepts centered on human-machine teaming, as well as the cognitive
consequences of the widespread fielding of AI capabilities.
The Joint AI Center will play a critical role in transforming the
Department by delivering capability at speed to address key missions;
establishing a common foundation for scaling AI's impact across the
Joint Force; and facilitating AI plans, policies, and standards,
including those that ensure we lead the world in the development of AI
solutions that are robust, resilient, ethical, and secure. We will
attract and cultivate the expertise of a world-class AI team and an AI-
ready workforce.
The speed and scale of technological change required is daunting.
However, the Department must embrace it if we are to reap the benefits
of continued security and prosperity for the future. Our sustained,
systemic approach accompanies a palpable sense of urgency. Ultimately,
this needs to extend across our entire department, government, and
society.
I look forward to continuing to work with Congress in an ongoing
dialogue on our progress in AI adoption, and the ways in which JAIC is
being used to accelerate that progress. Thank you for the opportunity
to testify this afternoon, and I look forward to your questions.
Senator Ernst. Absolutely.
Again, thank you, to our witnesses, for being here today.
I'll go ahead and start with the questioning here, and
then, when Senator Peters--oh, here he comes--when he returns--
I'll go ahead and start with my questions, and then, Senator,
I'll turn it over to you.
Again, thank you very much. This is a very interesting
topic, and I think we can learn a lot from the discussion
today.
To all of our witnesses here, if you could share, how are
the AI efforts in R&D coordinated among DARPA, JAIC, and the
services? General Shanahan, you had mentioned the
synchronization of AI activities, and you had mentioned R&D. If
you could all share, how do you synchronize that information?
How do you share that information? What are the best techniques
in doing that?
Dr. Highnam, if we could start with you, please.
Dr. Highnam. Whenever DARPA starts a research program,
there's a development of use cases. We seek to understand: If
we succeed in that program, who cares, who benefits? That means
that our program managers are out, talking inside the services
all the time; in fact, across the national defense
establishment, writ large. There's the natural inbuilt
connection before we even start, before we even agree to start
a high-risk activity. That's true whether it's hypersonics or
quantum or AI. This is normal business.
Now, in fields like AI technologies, which are software
tools with a lot of tail to them--sustainment, deployment
tails--I personally, as an R&D guy, am really happy to now have
the JAIC sent up as a partner to take on that 6-4 and on, that
engineering, deployment, sustainment tail, because I expect it
will make transitions into practice a lot--not simpler, but
more straightforward. I fully recognize just how much hard work
General Shanahan and his team are going to have to do to make
that end of the business happen.
Senator Ernst. Absolutely. Thank you.
Mr. Brown. Senator Ernst, the most important area for us to
collaborate with is JAIC. As I'm sure you recognize, DARPA has
a different timeframe in mind that we all benefit from, being
longer term. DIU's timeframe is 24 months or less, so we aim to
get commercial companies on contract within 60 days, and then a
prototype fielded within 2 years. In software, we're trying to
go faster than that, a year to 18 months.
In coordinating projects, our strategy with JAIC, which I'm
very pleased to be in partnership with General Shanahan, is,
we'll go out and look at what's successful commercially,
including vendors, and then, if we prototype something
successfully, we're the trial, and we want to scale it. Then we
start working with JAIC for what's the infrastructure we need
and how would we make that available to all of the services.
For example, we're working together now to get a vendor that we
have worked with on prototype to get a production contract that
will be with JAIC so any of the services can take advantage of
that.
For our project-base work, we also coordinate with the Vice
Chairman of the Joint Chiefs. We have a quarterly meeting with
General Selva, not just on AI, but our other projects, to make
sure we're doing things that make sense to joint forces. Then
we have monthly meetings with each of the Assistant Secretaries
for Acquisition--so Army, Navy, Air Force, Dr. Jette, Dr.
Roper, Hondo Geurts--to make sure that what we're working on
makes sense with their priorities. The last thing we want to be
doing is a lot of independent projects that don't have
leverage----
Senator Ernst. Right.
Mr. Brown.--across the----
Senator Ernst. Correct.
Mr. Brown.--Department.
Senator Ernst. Correct. Thank you.
General Shanahan. Senator, while the number may vary
depending on who wrote it, I think, in fiscal year 2018, the
number was 511 projects that had AI as their primary focus
across the Department. The question is, Are all those 511
projects towards a common end, in support of the National
Defense Strategy? This question of synchronization is essential
to where we're going in the JAIC, and it comes down to
governance and oversight. In section 238 of the NDA [Naitonal
Defense Authorization Act] actually directs governance and
oversight, for this very reason.
We have a lot of work to do in this area. I would like to
start by just getting our arms around all of the projects that
will come out in fiscal year 2020, to understand the amount of
funding, what the projects are for, not to threaten somebody's
budget. That is not the intent of the JAIC. But, we owe it to
the Department and to the Hill and to the public to be able to
account for all of those projects and the money that's being
spent.
I take that very seriously. We're still in the building
phase right now for the JAIC, but we are in early discussions
about what governance looks like for the JAIC and, How do we
bring all of us together to understand what are the projects
going towards? A $200,000 research project at University of
Michigan may be exactly what we need for a long-term insight
into a particular part of autonomous vehicles. The question is,
Do we know about it at a central level so that the Secretary
and the Deputy Secretary of Defense are comfortable about what
the Department is doing in artificial intelligence?
We take this very seriously. As Dr. Highnam said, we're
also in discussions, just between DARPA and us, on, Where is
that transition from DARPA, ready to field, over to the JAIC?
We are in early conversations of that. We don't have programs
identified yet.
Senator Ernst. Very good. Well, I appreciate that.
Going back to what Dr. Highnam said is, of course, Who
cares and who benefits? I think, bottom line, that is a great
way to put it. If you're not sharing information and going
through that synchronization, who cares and who benefits? We
don't really know. I appreciate that very much.
Thank you. I will step out. Ranking Member will take over
the meeting.
Thank you.
Dr. Highnam. If I may, one go-back on that. In all of our
research programs, we also seek transition of the technologies
that come out. We don't just do the research. This is Defense.
We're pushing it. We seek transition agreements with the end
users, wherever they may be, in the services or in the IC
[intelligence community].
Senator Ernst. Thank you.
Senator Peters [presiding]. Thank you, Madam Chairwoman.
I think I want to--I'd like to pick up a little bit, Mr.
Brown, on--you were discussing the commercial sector and how
we're reaching out to the commercial sector to be bringing in a
lot of this technology. Certainly, that's what we're seeing--
probably some of the most exciting advances are happening
outside of the DOD, in the commercial space, and, because of
all the applications from the financial industry, to banking to
insurance to automobiles--I mean, all of that is going to be
transformed in significant ways from artificial intelligence.
But, it's important for the DOD to be able to bring that in
and use it effectively. There are a number of factors that
usually, I think, stand in the way of that happening, from our
very cumbersome procurement process, to say the least, that we
have, that scares companies away from being involved with the
Federal Government, to a slow and often late budget process
that we have here. There are enhanced security reviews. I mean,
there's a long list of challenges. That's what I'd like you to
elaborate on, as to, What challenges do you envision, as we try
to adapt some of these commercial applications into military
use? Then, General Shanahan, if you'd follow up on Mr. Brown's
comments. Dr. Highnam, too.
Mr. Brown. So, Senator Peters, you're exactly right. I am
benefiting from the wisdom of folks who came before me in
setting up the Defense Innovation Unit, because we largely
address, by how we were formed, some of those constraints that
you talked about.
First, procurement process. We have set up a special
solicitation process. It's open--anyone can respond--where we
do not start with a list of detailed requirements assuming we
know how industry should solve a problem we might have in DOD.
But, we start with something very simple--sentence or
paragraph, saying, ``This is the problem we're trying to solve.
What can you offer us that will help address that?'' That gets
us away from, again, very detailed requirements to seeing, What
does the commercial sector offer?
Then we try and move at commercial speed and commercial
terms, meaning we don't have onerous requirements for IP
[intellectual property], and we don't take companies through
something that is unfamiliar to them. We'd like them to view
DOD and government as just another vertical as they look at
other commercial segments they want to pursue. Again,
commercial terms and speed are important for us there. Because
our mission is, How do we expand the national security
innovation base? How do we get more vendors working with us?
Then, as it relates to the budget process, that's something
that we are looking at now. How do we ensure that there's a
transition if we successfully prototype a use case? How can we
move quickly to get that fielded? We have to use a variety of
techniques. Some of them you've helped establish, like the
Rapid Innovation Fund. Fortunately, in the AI sector, that's
made much more easy with the partnership with JAIC, because now
we've got infrastructure, folks who can help make this
available to the rest of the services. We've talked about the
contract that we're working right now, production contract
being one we'll be able to draw from.
As it relates to security, we try and move away from
classified use cases and translate those to a commercial
problem. We try and work almost exclusively in an unclassified
realm. We're conscious of those constraints, and we have ways
to make it easier for commercial vendors to work with us.
I think another benefit is being able to work and access
the talent, the ideas that come from the folks in the
commercial sector, because we may not be able to track all the
talent--it's likely we will not be able to--in the AI fields
within the Department of Defense.
Senator Peters. All right. Thank you.
General?
General Shanahan. Senator, the legislation is clear:
commerciality, first and foremost. For the 2 years that I
worked in Project Maven, we took that approach. Now, there are
always going to be some unique problem sets within the
Department that require some in-house developments and in-house
solutions, but we went with commerciality every time. I would
say I was fortunate, fortunate in the form of a Marine Corps
colonel who was an operator, an intelligence professional, but
also a level-3 certified acquisition pro, and he was able to
work within the confine of the DFAR [Defense Federal
Acquisition Regulations Supplement]. People are surprised that
we use the DFAR to that effect. There are additional
authorities we've been granted. I haven't had to use them yet
in the JAIC because we're so new in the standup of the process.
But, there are ways to work the system, thanks to--as Michael
Brown just said, the existing solutions are already out there
in commercial industry. As I get further into standing up the
JAIC, what I'm looking for, as many arrows as possible in the
quiver of acquisition and contracting, able to pull for a
different situation on any given day, whether it's an other
transaction authority, commercial service, or just using
straight-up DFAR. But, it's not easy to do it, but there are
ways to work within the system, and we do put commerciality at
the beginning of every project.
Senator Peters. You say that it's not easy, but there are
ways. You believe that you have the authorities that you need,
at least at this moment? Or is there more that this committee--
--
General Shanahan. I do believe, at this moment, we have the
authorities we need.
Senator Peters. Okay.
General Shanahan. I reserve the right, 1 year into this, to
come back and make a different case.
[Laughter.]
Senator Peters. Yeah. Duly noted, General.
Dr. Highnam, did you have anything to add?
Dr. Highnam. Yes. From the research aspect, looking at our
investments in fiscal year 2018, about 50 percent of our AI
research investments were industry, about 14 percent were small
business. We have a very large coverage of picking up and
driving the development of the best ideas. About one-third went
to universities. Those are the sources.
For us, as we look ahead to technologies coming onboard,
maturing them, and, to the examples they gave earlier, reducing
the brittleness and just catering towards the engineering front
needed for large-scale military deployments, we're addressing
rigor, making sure they work, robustness. Second-wave
technology is applied aggressively to defense applications.
Then creating and proving out the third wave of technologies--
of AI technologies--again, creating them not from whole cloth,
but from working with the companies and working with the
schools to do that.
Senator Peters. Great.
Thank you.
Senator Shaheen.
Senator Shaheen. Thank you all very much. I'm sorry I
missed your testimony earlier.
In 2017, China laid out plans to become the leader in AI by
2030. What's our strategy to make sure, (a) that doesn't
happen, and (b) that we are the leader, as opposed to China?
Mr. Brown. I'll take a crack at that. This is obviously
much broader than a Department of Defense strategy. I think we
know well how to win a tech race, because we did it quite
effectively the last time we were involved in one with the
Soviets in the Cold War and afterwards. It starts with, What
are doing to invest in ourselves? All the breakthroughs that
Silicon Valley is benefiting from, even today in our economy,
as we look at some of these software IPOs [initial public
offerings]--Uber, Lyft, AirBNB--have come from federally funded
research. I credit DARPA and the other parts of the Federal
Government that create those breakthroughs--Internet, GPS
[global positioning system], miniaturized electronics, et
cetera. I think it starts with what we do in federally funded
research, education. What we did to focus on engineering and
science after Sputnik, need to do that again. Then the national
purpose. What we have now with the executive order, how do we
build on that to create a common purpose about this being
important? My concern would be that--How many Americans know
about the national order on artificial intelligence, and how
many young people are we reaching to inspire that this needs to
be their mission? Because this technology race, especially on
AI, is going to be multigenerational. It's not going to be
lasting one administration, or two. We've got to get the
national purpose behind this to support, then, what we can do
to leverage that in the Defense Department.
Senator Shaheen. Well, that raises the next question, which
is, Are we doing that?
Mr. Brown. Well, I think you could always say we could be
doing more.
Senator Shaheen. Okay. What more should----
Mr. Brown. There's no----
Senator Shaheen.--we be doing?
Mr. Brown. There's no time to waste in this race with
China. They have----
Senator Shaheen. So, what----
Mr. Brown.--some advantages, in terms of, today, probably
more patents that they've--there's more startup activity, in
terms of dollars invested. But, the U.S. still has a lot of
critical advantages, in terms of our education system, what
we're doing to actually pioneer things, in terms of hardware
technology to advance AI, the tensor processing units, et
cetera, the activities like DARPA is working on, with very
long-term research in mind. I feel like the U.S. still has a
lead there. We've just got to take more advantage of that. What
makes us special in this race?
Senator Shaheen. How concerned are we that, not only is
China making this commitment, but that they're stealing our
intellectual property, which includes AI, and that we have not
figured out how to adequately respond to that, I would say? You
all may not agree with that, but that's certainly my view. I
don't know who would like to answer that? General Shanahan?
General Shanahan. Yes, Senator. To just carry on to what
Mr. Brown was talking about earlier, it--this is not just a DOD
question--whole of DOD--it's not just whole of government, it's
whole of society, it's multigeneration to be able to build. If
I look at bringing in talent--AI talent into the JAIC, I can
ask the services to, ``Give me your best AI talent.'' There's
just not enough to go around. It will take decades to build
this. This is, one, for the executive order on AI. It's a
start, but there has to be an implementation plan, which I know
is coming. But, also, the National Security Commission on AI
will lay out some of these very factors about, How do we do
this as a society, everything from grade-school education to
military courses bringing in concepts of coding all the way
from the very beginning?
Now, to your other point, Senator, about intellectual
property theft, every one of us has a concern about that. It's
been taken much more seriously in the past 2 years than I would
have said 5 years ago, beginning to understand the scope of the
problem. Just using Huawei as an example, having a whole-of-
government approach to convince people not to use that
technology, because it has an entry point into places in China.
This is something we're working very hard at protecting our own
systems, protecting our data. Without getting into any details
in the project I worked on the Under Secretary of Defense for
Intelligence Project Maven, but also as we stand up the JAIC,
is protecting our data, doing everything we can to make sure
somebody doesn't understand what that data is, how we built our
algorithms. There is so much more than this. But, I believe the
sounding board of what China is doing, just within the past 2
years, is now making a difference. Much more to do, but we are
taking a different approach than we were in the past.
Senator Shaheen. Well, thank you. I appreciate that.
I would argue that, as we look at the education system,
that one of the things we should recognize is the importance of
immigration to that, and that, as we look back over the last 30
or 40 years, that one of the things that has been so important
to our system of higher education are those people from around
the world, the best and the brightest. When we have a system
that says, ``We don't want you to come here to college, and we
want you to go home as soon as you're done,'' that's not in our
interest. I would argue that that needs to be part of our
strategy, as well.
Thank you.
Senator Ernst [presiding]. Thank you, Senator Shaheen.
Senator Heinrich.
Senator Heinrich. I want to start by thanking our Chair and
Ranking Member for hosting this hearing. This is an incredibly
important topic, and one which we all need to be, educating our
peers about, because, as our guests today know, this is going
to be a bigger and bigger piece of what we focus on in the next
few years.
I'm really pleased to announce that this week we are
officially launching the Artificial Intelligence Caucus in the
U.S. Senate. Along with Senators Portman and Schatz, Gardner,
and our Chair and Ranking Member, we're looking forward to
trying to work together to strike that right balance in
developing the technology and the policy so that academia or
labs, private industry, and Federal entities like the ones we
have testifying here today, can harness this to the benefit of
the American people.
AI is, as you said, really going to impact every sector of
our economy, our society, not just the Department of Defense. I
want to start with Lieutenant General Shanahan and thank you
for your participation yesterday with the AI Commission. It's
my understanding that the services and other components in the
Pentagon right now have been directed to coordinate with the
JAIC, with the Joint Artificial Intelligence Center, regarding
any AI initiatives that cost more than 15 million annually. Is
that coordination happening?
General Shanahan. Senator, it's not fully in place yet. It
is----
Senator Heinrich. Okay.
General Shanahan.--my intent, through governance and
oversight within the JAIC, to put that structure in place, for
the very reason that you said. We have to know what they're
spending it on.
Senator Heinrich. It all starts with knowing what we're
doing----
General Shanahan. Yes.
Senator Heinrich.--and then building off of that. Do you
have the authorities that you need to be able to do this part
of your job effectively?
General Shanahan. I'd say section 238 will grant those
authorities. If I feel like we need any other authorities,
we'll go back through the Department. But, I believe I have
those authorities right now.
Senator Heinrich. As we look at this over the course of the
next couple budget years, what should we be measuring the JAIC
against, in terms of metrics, and by what timeline?
General Shanahan. When we talk about the JAIC, in
capability delivery, I'll divide it up into product delivery
and then the rest of the JAIC. Product delivery is, Are we
delivering on the national mission initiatives and component
mission initiatives? On the national mission initiatives, have
we put results in place that are making a difference, with a
return on investment? That won't be an instantaneous measure,
even with Project Maven, which has been going for almost 2 full
years right now. The return on investment takes a while to
measure in AI. As we talked about yesterday at the National
Security Commission, this is transformational. When you feel
the first sprint 1 algorithms, they are not game-changing,
they're designed for the operator to say, ``They're not good
enough. Here's what I need to do and get to sprint 2, and we'll
get to transformation.'' But, we need to show that we are
delivering capabilities.
For the component initiatives, I need to give an incentive
for the services and components to come to the JAIC. How do I
do that? One, funding. Two, joint common foundation, or a JAIC
common foundation. ``I have data for you to use. I have tools.
I have frameworks. I have some cloud and edge services. I have
a--I'm a place that--one-stop shopping,'' which is a term that
doesn't always work as well as it sounds on paper. But, I need
to give people an incentive to come in to the JAIC, to help
them accelerate their own AI initiatives.
Yes, sir.
Senator Heinrich. Yesterday at the AI Commission meeting, I
thought it was really helpful, what you said about the cultural
nature of this, and the multigenerational aspect of this. Talk
about how we manage that. Because the people who manage it,
yourself included, we're not going to have the same intuitive
access to this world that the people getting out of coding
schools right now have today. How, as the Pentagon, do you
manage this cultural transition within such a large
organization?
General Shanahan. Well, I would put culture and talent
management at the top two of my priorities in trying to change
the Department in bringing artificial intelligence into it. As
I said yesterday, there is a combination of top-down pressure
and bottom-up innovation. For the most part, I believe the
bottom-up innovation exists. We have to give it an outlet, a
vehicle to give people room to go out and try things new and
different, allow them to fail, and just show that they have a
different way of doing business, that we can listen to them.
There are now new programs in some of the departments, like
Kessel Run with the Air Force. There is more and more of a
culture change beginning to happen, but it's not part of the
institution yet. What we have to do is institutionalize it. We
have to give the top cover, in forms of resources, authorities,
and policies, as well as going out and giving capabilities to
people in the field.
One of the things I say is that, absent somebody getting to
play with AI, it's science fiction. They need----
Senator Heinrich. Right,
General Shanahan.--to see it, to smell it, to touch it and
really see what it can and cannot do. And part of that is
experimentation. It's almost like a war period between World
War I and World War II, where we can go out and actually
experiment with these capabilities. But, to do that, we have to
develop the capabilities. It's a little bit of a vicious cycle.
We have to get capabilities in the hand of operators and
analysts, try them out, wargame with them, try new operating
concepts, and then figuring out what works and doesn't work.
That cycle is a little slow in getting going right now. When
you ask about timeframe, I would say a year, in some respects,
in terms of delivering capabilities; 2 years to begin to say,
``Are we changing the Department?'' As you heard from Colonel
Cukor yesterday, we're 2 years into this, and I would say not
everybody accepts the change----
Senator Heinrich. Yeah.
General Shanahan.--that's coming.
Senator Heinrich. Great. Thank you, General. Appreciate it.
Senator Ernst. Thank you, Senator Heinrich.
I think we'll go ahead and do a second round of questions.
I will reserve my questions until the end so we can make sure
that the rest of our Senators have an opportunity.
Senator Peters, please go ahead.
Senator Peters. Thank you, Madam Chair.
General Shanahan, there is a concern out there by many
about the possibility that AI-enabled systems and autonomous
systems will cross some ethical lines, especially in
operational settings. I know that the DOD AI Strategy includes
efforts to think about AI ethics and safety issues are you're
developing the systems. As you just said, you've got to get
them out in the field, you've got to work them, but it probably
makes sense to be thinking about this on the front end, as
well, as we go forward.
For the panel, here, if you could highlight for us what
your biggest ethics-related concerns are for the possible use
of AI systems by the military, and how you're working to
address them.
General Shanahan. Yes, Senator. Every technology introduced
in the Department comes with a question of the lawful, safe,
and ethical use of that technology. AI is not different in that
respect. It has some differences, in terms of what you would
call ``explainable AI.'' Is an AI making decisions based on
data now instead of algorithms--rules-based algorithms that
have been programmed into it? We are thinking about this from
the very beginning.
Based on my work in Project Maven, I can tell you the
algorithms fielded are light years away from SKYNET and full
autonomous weapon systems. But, we know we have to start
thinking about the policy implications of that. If you were to
ask where the highest temperature is outside the Department,
it's on the question of lethal autonomous weapons. Autonomy in
weapon systems is governed by DOD policy today, and we are
partnered, and the JAIC is partnered, with the Defense
Innovation Board, who has a year-long project underway about AI
principles for defense, doing open hearings, being able to hear
from anybody that wants to come in and talk about their
concerns about the ethical, safe, and lawful use of artificial
intelligence in DOD. I will tell you, it's something we take
extremely seriously. We will go at this, as we have done with
other technologies, through a very rigorous and disciplined
test and evaluation, validation, and verification process. We
have not fielded an algorithm in Project Maven without having
gone through that rigorous and disciplined process. As early as
we are, and as brittle as those algorithms are, we put them
through that process. If we start talking about full autonomous
systems, that level of rigor and discipline will only continue
to increase.
But, in terms of what we are most concerned about is its
performance of algorithms. As Dr. Highnam said earlier, some of
the algorithms have failure modes that we have to take into
account. That DOD directive that I referred to, has several
sections on what we have to go through in the Department to be
ready to test and field technology that involves autonomy.
But, autonomous weapon systems with artificial and general
intelligence is what people seem to think is the worst case. I
think of artificial narrow intelligence. Anything we field will
be fielded in accordance with the Law of War, international
humanitarian law, rules of engagement, and commanders'
judgment. I mean, these are things that we take into account
for every technology, even more so because people don't know
all about the implications of artificial intelligence in a
weapon system.
Senator Peters. Yeah. As you go through that process, you
know, certainly, that's encouraging, that that thought process
is occurring within your organization, but I think we have to
also realize that some of our adversaries may not be
constrained by some of the same kinds of processes that we go
through, and could present unacceptable risk to us, from a
national security perspective, as well as the men and women who
go in harm's way facing autonomous systems that operate under a
completely--set of rules than what we would think is
appropriate here in the United States.
I think that leads to my last point, and that's thinking of
some of these higher-level concerns and policy concerns of what
we should be thinking about globally, in terms of these
technologies. I know, when we were talking about AI, we often
turn to technical experts and engineers. Often technical
experts and engineers make comments about ethics. But, I have
found that's a somewhat narrow approach, and we've made those
kinds of mistakes in the past. My sense is--are you and Dr.
Highnam--are you also working with philosophers and ethicists
and folks who think deeply about some of the moral questions
associated with these technologies? Or should we be doing more?
Dr. Highnam. One, there's always more to do. We had an AI
colloquium last week, a DARPA colloquium. About 700 people
there. One of the most interesting panels that we had was on
ethics, led by Richard Danzig, who used to be the Secretary of
the Navy. Fascinating discussion. It's very much a part of the
technical discussions that are going on. We are looking at
that.
Now, within--wearing a slightly more technical hat, there
are some issues, at the moment, that we are very concerned
about, with a technical solution. One is the implicit bias. The
field of data science and machine learning or machine training
have significant overlap. There's a tendency among people who
are human, as they build these systems, to pick datasets, to
cull datasets, to unintentionally put leads or, again, bias
into how they're doing things, which means a system could
preferentially recognize Joe or Jane, based on--just because of
the way it was trained. That's a piece of the puzzle. We have
research programs going directly against that. A large one
underway now is understanding group biases. But, again, this is
common to data science and machine training.
The second area is about the deployment of technology when
we don't fully understand its failure modes, back to the point
I made earlier. One of our programs, short autonomy, has a very
interesting premise, an important one. We have an autonomous
vehicle, a flight vehicle, a ground vehicle. You can make it a
lot more flexible in how it deals with unexpected conditions by
adding some second-wave AI technologies to the puzzle. A
condition shows up, it adapts and makes changes. But, it's--
again, if you don't really understand the failure modes, if you
don't have that assurance and sense, almost, of a cyber
assurance that this is going to behave itself and operate
within safe limits, then you put something on the street or in
the air that's--you really have to take a little--you have to
think hard about before you do that. A lot of our research,
again, is going into making technologies robust in that sense,
as well. We have multiple programs--research programs
addressing different aspects of this problem. It's a very
important problem.
Senator Peters. Great. Thank you.
Senator Ernst. Senator Shaheen.
Senator Shaheen. Thank you.
Dr. Highnam, did I understand you correctly when you said:
As we look at where AI is currently being developed in this
country, about 50 percent of it is in large businesses; 14
percent, small businesses; and a third from universities? Was
that----
Dr. Highnam. It's 50 percent in business, of which 14--so,
14 percent, overall----
Senator Shaheen. Ah. Okay.
Dr. Highnam.--then 36 percent, larger businesses; 34
percent universities; and the rest in service labs, energy
labs, and so on.
Senator Shaheen. The rest is from the public sector----
Dr. Highnam. Yes.
Senator Shaheen.--then.
Dr. Highnam. Yeah.
Senator Shaheen. If you think about past circumstances in
our history, whether it was the Manhattan Project or putting a
man on the Moon, can we analyze the sectors that provided that
technological innovation, and figure out whether this is the
right breakdown, in terms of where AI is coming from?
Dr. Highnam. The answer is certainly yes. It's not
something I've done. But, to comment on that, if I may, the--
this is the research phase. We're finding--these investments
are not on systems that are in any sense deployable.
Senator Shaheen. Sure.
Dr. Highnam. But, these are people--I'm finding them in
industry, with really good ideas, who propose to our research
programs. Then, as much of this technology evolves, they
publish some, they don't publish some, depending on----
Senator Shaheen. Right.
Dr. Highnam.--classification, and so on. But, a lot of the
time, we want industry to make these technologies, as they are
proven to work, to be commercial, to be incorporated into----
Senator Shaheen. Right.
Dr. Highnam.--other products that then the Department can
buy back. I think that the days are gone when we can think
about corralling hundreds of thousands or very large numbers of
experts in such a hot technology area. We--this is normal--
DARPA's normal business mode, but I'm certainly going to take
your question back for a look.
Senator Shaheen. Well, I was just thinking about, How do we
encourage more experimentation, more research? Thinking about
small businesses. Small businesses create 16 times more patents
than large businesses.
Dr. Highnam. Yep.
Senator Shaheen. You know, two out of every three jobs are
created from small business. Are there ways we can incentivize
small business to do more of that research and innovation that
we're looking at to provide the AI that we need? I would argue
that one program that is there that helps do that is the SBIR
program----
Dr. Highnam. Yeah.
Senator Shaheen.--Small Business Innovation Research
program.
Dr. Highnam. Two answers. One--I'm sorry, it was part of my
preamble--one of the things we've done in our AI campaign is to
set up something called AI Exploration. In that, we post a
topic of interest. Anybody is given 30 days to respond. It's
typically schools and small businesses who do that. Then we
award within 60 days after that. Ninety days from posting to
award, up to a million dollars per award, up to about 18 months
in duration. We've invested about 45 million so far, since
September, in this activity. Because, you're right, this is a
lot of the innovation, and this is us exploring in a space and
giving them the grounds to do that. We've also recently
revamped our small business approach to align it directly with
our research programs. We're also encouraging moving directly
to phase two. We also have an innovation accelerator, as well,
to advise small businesses on how to take things commercial
after they discover it, not just in AI, but across the board.
Senator Shaheen. Are there other policy changes that we
should be thinking about to promote--should we be encouraging
more set-aside for SBIR programs? Are there other ways in which
we can promote AI that we're not currently doing?
For any of you?
Dr. Highnam. We're seeing an awful lot of smart small
businesses come forward, teaming with schools, teaming with big
companies sometimes. Certainly those in the larger DARPA
ecosystem understand how to work us.
Mr. Brown. I think this just emphasizes what you've heard
from all of us, the need to work with commercial innovators in
?
Senator Shaheen. Right.
Mr. Brown.--AI. I saw an interesting statistic that came
from Congressional Research Service recently, that, in the
1960s, a third of the global R&D was U.S. defense-related, and
now that number is 3.7 percent. It just speaks to the need to
look outside. I think what you've heard from all of us--DIU,
that's our mission, so of course you'd expect me to say that,
but we heard it from Dr. Highnam, General Shanahan, that we
want to work with these successful innovators outside, and
bring that technology in, because, unlike the Manhattan Project
working on one specific goal, AI is a horizontal technology
that is infused in everything, or will be. That really speaks
to the beauty of the U.S. capitalistic free-market system so
that we can benefit from all that innovation happening across.
I think our challenge is, Where do we pull that in, from a
talent perspective, technology perspective, and proven use
cases? How do those apply to the Defense Department?
General Shanahan. Senator, it's about messaging that--as
been said, AI, unlike any other technology in the past, is
been--the equation has been completely turned around as
commercial and not government. The message of--the United
States Government, not just the Department of Defense, has an
interest in promoting AI from the smallest company up to the
biggest company. With Project Maven, we had no favorites.
Everybody was a player, smallest startup all to the biggest
companies in the United States. But, getting the message that
we want the business, and if they have their intellectual
property to work with the government on, we want to take that.
What I don't want to see is some of the best companies in
the United States, some of the best intellectual talent we have
out there being funded with VC [venture capital] money from
places like China. But, if they have to go somewhere, and we're
not giving them an opportunity, that's what's going to happen.
We have a role, I think. It's a very serious role, is to
communicate that we're serious about artificial intelligence,
we need the capabilities you bring to the table, and the three
of us represented here from AI now to AI next, and Mr. Brown in
between, sort of going out and doing the pilots and finding the
right companies out there, that is a message we need to
communicate. I think part of that, through the executive order
on AI, but also the National Security Commission on AI that
will come out with, I expect, some very weighty recommendations
about a societal change in how we're looking at artificial
intelligence.
Senator Shaheen. Well, thank you. I appreciate all of those
responses.
If we're going to continue to be competitive in the rest of
the world, then this needs to be part of our strategy. If
you're correct, Mr. Brown, that in our system that unleashes
all of this innovation in the private sector, then we should be
able to win that competition. But, I think that there are
policies that we need to put in place to encourage that, and we
ought to think about which ones make sense to get where we want
to go.
Thank you all very much.
Thank you, Madam Chair.
Senator Ernst. Thank you. Absolutely.
I will wrap up with just a couple questions. I do want to
thank you for the discussion. We've covered a lot of territory,
a lot of very interesting territory. I appreciate the
discussion on ethics, as well, with lethal autonomous weapons.
I think that's something that we need to fully vet and explore
even more.
But, what I'd like to do is turn back to the more mundane,
everyday uses of AI, if we can. General Shanahan, you had--I
think, had mentioned some of those uses. Of course, we have
companies, like Amazon and UPS and Walmart, and they do use AI
for those back-office types of tasks that you had mentioned
earlier. Can you walk us through some of those tasks and where
we might be able to utilize AI? Not big, sexy topics, but
certainly if we can streamline the way we do business within
the DOD, I think this would be helpful.
General Shanahan. Yes, Senator. When you talk about smart
automation, or, in the vernacular of the industry, ``robotic
process automation,'' it's not headline-grabbing, in terms of
big AI projects, but it may be where some of the most
efficiencies can be found. That's the case if you read the
dailies in industry, whether it's in medicine or in finance.
This is where early gains are being realized in AI. Some of the
other projects we take on in the Department are probably years
in the making of return on investment. These other areas, I
think, will be much shorter-term return on investment.
What we're trying to do in the JAIC--when I looked at this
just a couple of months ago, we weren't even concentrating on
this smart automation. I'm now trying to figure out how I stand
up a small office just focused on that. I don't see us leading
that, but it's leading others to find out how to incorporate
these technologies into their back-office functions.
I've already met with the Chief Management Officer of the
Department, as well as the Chief Data Officer of the
Department, to have these early discussions. I'm convinced
there will be lots of opportunities in back-office functions,
finance being, I'd say, the first one to take on, to help
augment people. I think people get very concerned, right off
the bat, about being replaced. There's not enough people to go
around, for the most part, so this is about augmenting people
and being able to do much more work than they were able to do
with the tools, which, in some cases, are far too old, manual,
laborious. These are about how to--if you see the
demonstrations of a bot versus a human doing the same sort of
manual task, there's no question who gets to the finish line
first.
Senator Ernst. Absolutely.
General Shanahan. We're early in this process right now.
But, that's one I'm very interested in taking.
Senator Ernst. No, I think that's really important. One of
the big discussions that we've had, just in the last year, was
the DOD audit, and how do we arrive at a clean audit through
such a large--what I describe as a large, you know, animal. Is
it practical to look at an application like that? Would it be
helpful to guide us towards a clean audit?
General Shanahan. I'd say the answer to that is yes.
Scoping it will be the challenge, is finding out how big this
is to go after the audit. But, I know the Chief Data Officer,
Michael Conlin, is looking at applications like this. There are
big decisions made in the Department with data done in a very
manually intensive way. If those decisions can be made faster
and better, that, of course, is something that the leadership
of the Department is interested in.
The answer to your question is yes. It's a question of
understanding the scope and the scale of doing it.
Senator Ernst. Very good.
Of course, the inventory purposes, acquisition, program
spending, you name it, I think that AI can help in those areas.
It has been mentioned, maintenance, as well--predictive
maintenance on equipment and aircraft, so forth, would be very
helpful, as well.
General Shanahan. Senator, if I may just add on to the
point. Whether it's smart automation or predictive
maintenance--as we're finding very early, the problems
themselves are not massive, but the lessons learned are what
we're really catching on to.
Just one use case of a helicopter, seemingly simple. But,
everything we're learning about data management, which would be
no surprise to anybody in industry who's dealing with
artificial intelligence and machine learning. But, those are
what we're trying to collate and bring up to a higher level for
the Department about understanding what different standards,
policies, authorities need to be in place to make this happen
against all the different aircraft in the Department of
Defense.
To your other point about--I call it a flywheel effect.
Once a few people begin to understand what smart automation
does, it will catch on. But, nobody believes it yet, because
they haven't the benefit of actually seeing it work.
Senator Ernst. Absolutely.
General Shanahan. But, that's what we have to do, is--we
have to show--it's the show-me piece. We have to have people
believe it's real, and not just science fiction.
Senator Ernst. Absolutely.
With that, I will go ahead and wrap up this hearing today.
I do want to thank the members of our panel for being here. The
flywheel effect, it starts here, as well, in Congress, and
making sure that we are educating others on artificial
intelligence and the applications for our DOD.
Thank you, again, to the witnesses for being here and for
educating us on what your jobs entail and how we can better use
artificial intelligence. Thank you, gentlemen, very much.
This concludes the hearing of Emerging Threats and
Capabilities.
[Whereupon, at 3:53 p.m., the Subcommittee adjourned.]
[Questions for the record with answers supplied follow:]
Questions Submitted by Senator Joni K. Ernst
ai to improve processes
1. Senator Ernst. Dr. Highnam and Mr. Brown, how can DOD
[Department of Defense] use AI to improve its business and back office
processes to save money for readiness missions?
Dr. Highnam. There are many commercial companies offering AI-based
solutions for business and back office processes. However, the
implementation, customization and deployment effort for such a project
is complex and subject to schedule and performance risks. Back-office
operations have been studied in-depth for decades, and many
sophisticated computer algorithms are routinely used in currently
deployed enterprise solutions. Nonetheless, an AI technology system
requires a software layer that captures user actions and business
processes in a computer-readable form. To be successful, AI methods
must access disjoint databases and combine financial reports with
computer-interpretable information about human decision processes. A
graduated approach to adopting AI techniques would first automate
repeatable processes while also capturing human responses to
nonstandard situations. Computer-readable information about human
responses would then provide the data required to test if and when AI
methods might further improve business processes. Offerings claiming to
provide value through the application of AI technologies should be
vetted thoroughly and their life-cycle-costs analyzed and compared to
best practice in the commercial world and to current solutions.
Superior solutions would both enable end users to automate repeatable
processes and learn from user behavior to recommend courses of action
for non-standard situations.
Mr. Brown. Based on routinely collected data, artificial
intelligence (AI), data analytics and machine learning technologies are
well-suited to make predictions to improve back office processes. For
example, commercially available software for robotic process automation
is the use of software with AI and machine learning capabilities to
handle high-volume, repeatable tasks that previously required humans to
perform. These tasks can include queries, calculations, and maintenance
of records and transactions.
2. Senator Ernst. Dr. Highnam and Mr. Brown, for example, can AI-
enabled systems help DOD achieve a clean audit, better track inventory
and acquisition program spending, or assist in the personnel assignment
process?
Dr. Highnam. The goal of an audit is to ensure compliance with a
complex set of regulations. To achieve this goal requires extensive
common-sense and organization-specific knowledge. For example, a
financial auditor must be deeply familiar with the Generally Accepted
Accounting Principles (GAAP) and understand they should be applied to
the organization in question. For the DOD, a thorough knowledge of the
Defense Federal Acquisition Regulation Supplement (DFARs) would also be
essential. The application of such knowledge is well beyond current AI
capabilities. AI tools can assist auditors by enabling more rapid
analysis of data relevant to the audit. For example, machine-learning
systems can be trained to detect fraudulent transactions. However, the
use of such tools requires years of training and experience. The DOD
may be able to achieve cost savings in audits through judicious
incorporation of automated processes and AI technologies. Nonetheless,
the DOD faces unique challenges related to scale, expiration of funds,
and operational diversity. It is unlikely that common commercial
products will fully address DOD use cases without extensions or
modifications. Graduated solutions should first advance the level of
automation while providing a software foundation for future AI
deployments. Finally, decades of research have produced well-
established procedures and mathematical concepts for managing
inventory, such as logistic regression for demand forecasting and
Economic Order Quantity estimation. Machine learning techniques may
potentially improve demand forecasts, but DOD should undertake a
careful analysis of any AI-based solution to ensure that its lifecycle
costs and complexity do not exceed current solutions, without
substantial increase in accuracy and timeliness. Similar caveats apply
to other areas of business process automation, such as personnel
assignment.
Mr. Brown. Although not a focus area of DIU at the moment, there
are a number of commercially available data analytics platforms that
make this information more visible.
3. Senator Ernst. Dr. Highnam and Mr. Brown, what specifically have
your organizations done to support AI for back-office functions at DOD?
Dr. Highnam. DARPA creates foundational AI components that
generalize to a variety of application areas. DARPA recently engaged in
a knowledge exchange with DFAS and DISA to explore whether DARPA AI
components might support back-office functions within the DOD. Several
near term investments would need to be made to lay the foundation for
future pilot testing and deployment of AI technologies from DARPA and
other organizations. Specifically, the DOD would need to invest in a
software layer that enables users to automate routine business
processes and stores the actions taken by humans to address non-
standard situations. Ideally, the software layer would connect disjoint
databases and software tools in a manner that builds on users' current
workflows, captures an ontology of user actions, and stores relevant
data in a computer-readable form. The software layer would provide a
foundation for pilot tests, extensions, and rapid deployment of AI
components from DARPA and other organizations. There will be a follow-
up meeting between DARPA, DFAS, and DISA.
Mr. Brown. DIU has been working with a commercial solution on AF
strategic planning choices to predict necessary POM inputs for given
decisions. The use cases center around using integrated data (current
and historical: budget; financial execution; manpower; personnel; and
logistics/equipment) to identify 3rd, 4th and 5th order effects of a
given planning choice. The intent is to improve fidelity in the
strategic planning process, and understand quantitatively the likely
outcomes on readiness, manning, equipment needs, training impacts, and
other weapon systems. There is both a quality and speed component to
decision making. Using a tool increases the throughput of decisions
through the budgeting process, and provides more fidelity so senior
leaders and analysts can ask the questions that will help them improve
decisions.
dod technology acquisition
4. Senator Ernst. Dr. Highnam, Mr. Brown, and Lt. General Shanahan,
how does the DOD acquisition process delay the development and
implementation of AI and other emerging technologies and what efforts
have you taken to address this problem?
Dr. Highnam. DOD acquisition is often process driven, rather than
mission driven like DARPA. The state of the art of emerging technology
like AI evolves much faster than DOD acquisition timelines. For this
reason, DARPA uses streamlined or rapid acquisition authorities to
connect with the best people and technology at the speed of innovation.
For example, DARPA has used ``Other Transaction'' authorities for
decades because they allow flexibility and often faster results than
traditional government contracting. Recently, DARPA launched the
Artificial Intelligence Exploration (AIE) program to execute forty-
eight (48) separate Other Transaction awards totaling $45 million, for
rapid feasibility studies of AI concepts lasting eighteen (18) months.
To keep pace with industry changes, DARPA's AIE awards had an average
timeline of less than ninety (90) days from announcement to award.
Mr. Brown. DIU's mission is to look beyond traditional DOD vendors
and focus on the commercial technology community. To attract these
commercial solutions, DIU developed the Commercial Solutions Opening
(CSO) process in 2016, a three-phase, competitive merit-based business
model leveraging prototype Other Transaction authority. This process
has enabled DIU to focus on balancing speed, flexibility, and
collaboration to award prototype projects to leading-edge, dual-use
technology companies that might otherwise not do business with DOD. A
potential future problem would be the length of time to get new vendors
cleared to solve some of our problems.
Lt. Gen. Shanahan. The Department of Defense's acquisition process
is still oriented primarily towards materiel weapon systems rather than
software development. Modern software development requires a different
process, with end-users providing feedback at every stage and
developers continually refining the model after deployment. In pursuit
of our mission needs, the JAIC is taking advantage of current
regulations and policies wherever possible to implement proven
iterative development strategies for software and emerging technologies
such as AI. Based on my previous experience with Project Maven, there
is often sufficient flexibility within the Defense Federal Acquisition
Regulation Supplement (DFARS), but to navigate the DFARS successfully
requires government personnel with a commensurate level of experience
and deep understanding of DOD acquisition and contracting regulations.
Further simplification of existing regulations will be helpful. The
JAIC is seeking to hire people with the requisite agile acquisition
expertise while also reviewing a wide range of acquisition and
contracting options designed to allow maximum flexibility and agility
when pursuing commercial AI solutions. The JAIC will work closely with
USD (Research and Engineering) and USD (Acquisition and Sustainment) on
improving AI-related acquisition and contracting policies and
authorities. Congress has been especially helpful in crafting
legislation designed to expedite fielding commercial AI technologies.
5. Senator Ernst. Dr. Highnam, Mr. Brown, and Lt. General Shanahan,
have you faced any roadblocks in the Department to speeding up
acquisition of these technologies?
Dr. Highnam. DARPA has not faced any roadblocks. DARPA consistently
uses a wide variety of existing acquisition flexibilities granted by
Congress, such as the ``Other Transaction authority'', which we have
used for decades. The Artificial Intelligence Exploration (AIE) program
was an example of DARPA using ``Other Transaction authorities'' to make
forty-eight (48) awards with an average timeline of less than ninety
(90) days from announcement to award. DARPA greatly appreciates the
acquisition autonomy and flexibility that Congress and the Department
grants us due to our unique R&D mission of pursuing breakthrough
technology and avoiding strategic surprise. DARPA's success so far has
resulted from this autonomy and flexibility, and our future success
depends on our ability to use a variety of special acquisition
authorities to pursue innovative R&D.
Mr. Brown. DIU has had a beneficial partnership with Army
Contracting Command--New Jersey (ACC-NJ) for contract execution, but
reliance on contracting organizations with other primary missions has
been unable to meet DIU's capacity needs and higher DOD demand. In
November 2018, the Office of the Under Secretary of Defense for
Acquisition and Sustainment (OUSD (A&S)) granted DIU authority to
execute ``Other Transaction agreements.'' In April 2019, DIU utilized
this new authority to release its first solicitation, the initial step
in awarding a prototype contract.
Lt. Gen. Shanahan. Not at this point, although we acknowledge the
risks inherent in not taking full advantage of a wide range of
acquisition and contracting practices that support agile development,
fielding, and sustainment. During the JAIC's initial standup in fiscal
year 2019, to avoid incurring delays in each of our three primary
projects, the JAIC has primarily taken advantage of existing contract
vehicles. As we prepare to scale operations in fiscal year 2020, we are
using lessons learned from other DOD-wide AI projects such as Maven to
create and combine a wide variety of new and existing contract
vehicles, respectively, that will provide the JAIC and others in DOD
with access to state-of-the-art AI technology and services, at the
necessary speed and with the required agility. We will be prepared to
request new or modified policies, authorities, and legislation as
necessary.
__________
Questions Submitted by Senator Mazie K. Hirono
ethical challenges
6. Senator Hirono. General Shanahan, the ethical use of AI is an
issue that must be adequately addressed before AI can be deployed in a
widespread and meaningful way. Who within DOD is responsible for issues
regarding the ethical use of AI within DOD?
Lt. Gen. Shanahan. Leadership in values and ethics is at the core
of everything we do in the JAIC. The responsibility for the ethical use
of AI will be widely shared across the Department, spanning system
developers, trainers, leaders, end users, and more. The breadth of this
shared responsibility is a recognition that many different communities
within DOD have a critical role to play in getting this right. The JAIC
will play a central role in helping craft policies, guidelines, and
authorities related to the ethical, safe, and lawful use of AI
technologies. This includes working with allies, partners, and
international organizations on questions of international norms for AI,
while precluding unnecessary restrictions on our development and
fielding of AI capabilities. The Department is fully committed to the
ethical use of AI from concept through deployment and sustainment. This
includes ensuring sufficient rigor and discipline throughout test and
evaluation as well as validation and verification processes and
procedures.
7. Senator Hirono. General Shanahan, are there separate efforts to
address ethical issues related to the uses of AI and the issues related
to AI systems potentially generating errors or unwanted outcomes?
Lt. Gen. Shanahan. The Defense Innovation Board (DIB) is developing
principles to guide current and future AI research, applications, and
activities as part of the AI Principles for Defense project. Upon
completion of the project, the DIB will make a recommendation to the
Secretary of Defense on proposed AI principles for defense, at which
point the Department will conduct an internal review process on
adopting such principles. Additionally, a major component of DARPA's
`AI Next' campaign, titled Explainable AI, focuses on enabling AI
systems to explain their actions. Progress in ``explainable AI'' will
help users understand, appropriately trust, and use AI systems in an
ethical and safe manner. Prior to fielding any AI capabilities, the
Department will ensure sufficient rigor and discipline in test and
evaluation processes and procedures with particular attention to
independent evaluations and assessments, errors, biases, and unwanted
outcomes. This will include establishing dynamic feedback loops that
take into account feedback from operational users on algorithm or model
performance on deployed systems.
predictive maintenance
8. Senator Hirono. General Shanahan, one of the areas of focus for
JAIC [Joint Artificial Intelligence Center] that seems particularly
promising is predictive maintenance. The ability to improve readiness
by limiting downtime with predictive maintenance could have a
tremendous positive effect on maintenance efforts across DOD. What are
the major hurdles to widespread deployment of predictive maintenance AI
within DOD?
Lt. Gen. Shanahan. I agree that machine learning (ML) looks
particularly promising in terms of augmenting, accelerating, and
automating current maintenance practices. The JAIC funded predictive
maintenance as one of its initial National Mission Initiatives because
of the near-term opportunity to increase aircraft readiness, aligning
with the Secretary of Defense's direction to improve aircraft readiness
across the Department. However, one of the earliest identified
challenges, which matches lessons learned from Project Maven and other
previous DOD AI/ML projects, centers on data: its quality,
accessibility, and relationship to ground truth. Existing maintenance
data is often not at the level necessary for AI systems development.
Critical data elements are often missing, incorrect, or at too low
resolution. Additionally, data inconsistencies exist across the
Services for related systems. One of the JAIC's roles will be to help
develop and implement DOD-wide policies, authorities, and governance
designed to improve data readiness across the entire Department. This
will include supporting efforts to digitize end-to-end processes, from
the back office to the tactical edge.
9. Senator Hirono. General Shanahan, where does DOD need to invest
to widely deploy predictive maintenance?
Lt. Gen. Shanahan. Predictive maintenance should be employed widely
across the Department. AI-enabling investments in data generation,
curation, data access, and maintenance are necessary precursors to
enable predictive maintenance AI to be successful at scale. Data
curation is one of the most important investments, along with needed
requirements to improve data quality at the point of use. In the
interim, the JAIC is relying on AI and domain experts to conduct
standardized post-processing of inaccurate and missing data. While our
initial H-60 helicopter use case is relatively small, the lessons we
are identifying in this project--particularly as they relate to data
management--will be applicable to every future project in DOD.
Investments in ontology, component standards, and use of a common
infrastructure are addressing challenges with data access. Shared
ontology ensures a common data lexicon across the Joint Force--an
important step toward scaling new AI-capabilities. The JAIC is
supporting a number of ongoing standards-related efforts including
those led by the National Institute for Standards and Technology. If
such data standards are established and uniformly adopted, then it will
help DOD ensure future equipment procurements include government
ownership of data needed to efficiently apply AI analysis.
10. Senator Hirono. General Shanahan, will predictive maintenance
AI be deployable at sea, where ships have traditionally been limited in
computing power and communications bandwidth?
Lt. Gen. Shanahan. The JAIC's Predictive Maintenance National
Mission Initiative uses historical H-60 helicopter data. However, the
Navy maintenance community has already raised the challenge of limited
communication bandwidth as partially explaining missing data that is
necessary for AI analysis. The Navy is working through the connectivity
requirement between the Integrated Mechanical Diagnostics System (IMDS)
and TRIAD shipboard servers to ensure maximum use of AI technology at
the point of use. This is an area where the JAIC will work closely with
the Navy and Marine Corps to identify and address specific challenges
related to deploying AI-enabled predictive maintenance solutions at sea
or in any degraded, disconnected, intermittent, or bandwidth-limited
environment.
__________
Questions Submitted by Senator Martin Heinrich
commercial industry engagement
11. Senator Heinrich. Mr. Brown, I am aware that you have DIU
[Defense Innovation Unit] offices set up in Silicon Valley, Boston, and
Austin. I'd like to make sure that AI startup companies, say in
Albuquerque, Santa Fe, Des Moines, or Detroit have the same
opportunities to present their AI solutions for your problem sets. How
should they go about doing so?
Mr. Brown. DIU is interested in getting the best commercial
solutions from around the country and growing the national security
innovation base. To date, we have received submissions from over 800
companies in 42 states and the District of Columbia. We post an Area of
Interest (AOI) solicitation on our website, www.diu.mil, to which
companies across the country can submit proposals.
12. Senator Heinrich. Mr. Brown, is New Mexico within your upcoming
outreach efforts? If not, I would like to extend a personal invitation
to visit New Mexico and meet with several cutting-edge AI companies.
Mr. Brown. Thank you for the invitation. We have recently met with
your congressional lead about our processes and opportunities to work
with DIU. We have a number of companies that have responded to our
AOIs. We will definitely look to visit New Mexico in upcoming travels
nearby.
the cloud and ai
13. Senator Heinrich. Mr. Brown and Lt. Gen. Shanahan, as you know,
the cloud contract and process itself, JEDI, has received significant
attention. In general, how important is the establishment of a cloud to
supporting the deployment of future AI capabilities?
Mr. Brown. The establishment of a cloud solution to support the
deployment of future AI capabilities is very important. Most vendors in
the AI/ML space are moving toward cloud services to provide their most
advanced capabilities, so it is a high priority to establish and
migrate to cloud infrastructure to stay current with the vendor base.
Large training data sets and substantial/scalable computational
resources are two core capabilities that enable effective artificial
intelligence applications. Cloud solutions enable those datasets to be
shared across geographically separated commands, pooling together
larger datasets. Cloud computing allows a far more efficient allocation
of DOD capital, as the precise amount of computational resources are
spun up to solve the specific problem and no idle capacity is being
paid for.
Lt. Gen. Shanahan. Achieving the most effective results possible
from AI-enabled technologies will depend on a future of enterprise
cloud. The cloud provides massive and elastic compute power, lower
latency, unlimited storage, and scalability. It will also enable local
and global dynamic algorithm and model updates at the required speeds
that will be essential to future multi-domain operations. The JAIC's
objective is to use AI to solve large and complex problems to ensure
that the Services and Components have real-time access to ever-
improving libraries of data sets, reusable tools, frameworks, and
standards. This cannot be achieved without the foundational framework
established by a hybrid solution of general purpose and fit-for-purpose
enterprise clouds, in accordance with DOD's enterprise cloud strategy
14. Senator Heinrich. Mr. Brown and Lt. Gen. Shanahan, is it
possible to deploy AI capabilities without cloud or clouds in place?
Mr. Brown. Yes. Many of our prototypes are either ``on premises''
(non-cloud) deployments, or hybrids that have both on premise and cloud
elements. It is possible to deploy trained algorithms at the edge with
a ``train in cloud, deploy at edge'' concept of operations.
Furthermore, it is also possible to deploy additional processing power
at the edge with some of the more advanced hardware available from
vendors. However, the core benefits of AI/ML solutions come into play
when we pool together many data sources in a common operating
environment.
Lt. Gen. Shanahan. It is possible to develop and field AI-enabled
capabilities without a cloud. However, such `one-off' solutions will
limit the overall effectiveness of deployed capabilities--everything
from training through inference at the tactical edge and dynamic
updates to deployed algorithms/models. It is possible to deploy limited
AI capabilities in traditional data centers and on-premise
infrastructures; data centers are generally cost prohibitive and suffer
from a number of other limitations lowering the ability to achieve
maximum effectiveness from AI technologies. AI deployment at scale
requires the elasticity of compute and storage that is afforded by
cloud computing, while also accounting for compute hosted on platforms
and sensors at the tactical edge. The full potential of an AI-enabled
future depends on enterprise cloud that is optimized for AI, with
sufficient backup measures in place to account for operations in
degraded, disconnected, intermittent, or bandwidth-limited
environments.
ethics and safety
15. Senator Heinrich. Dr. Highnam and Lt Gen Shanahan, how is the
Department of Defense addressing ethics and safety concerns surrounding
military applications of AI?
Dr. Highnam. The safe use of AI technologies by the military and
others is an important concern of AI researchers and practitioners.
DARPA is addressing this concern in multiple research programs,
including its Assured Autonomy program. The Assured Autonomy program is
developing rigorous design and analysis technologies for continual
assurance of learning-enabled autonomous systems to guarantee safety
properties in uncertain environments. These include new techniques for
modeling and system design, formal verification, simulation-based
testing, machine learning, and safety-assured learning. The
technologies being developed in the Assured Autonomy program will
enable the DOD to more rapidly and efficiently deploy learning-enabled
autonomous systems that can be trusted to operate safely in uncertain
environments. DARPA's recent AI Colloquium featured an ethics panel
with in-depth discussion of ethical issues relevant to AI in military
contexts: https://youtu.be/jSxCWLJt0wY.
Lt. Gen. Shanahan. The Department continues it full commitment to
the ethical, safe, and lawful use of AI. Leadership in values and
ethics is at the core of everything we do in the JAIC. One of the key
pillars of DOD's AI Strategy is leading in military ethics and AI
safety. JAIC will support ethical implementation of AI by consulting
with leaders across academia, industry, and the international
community; investing in resilient, robust, reliable, secure, and
explainable AI systems; developing and improving policies that that
consider technical strengths and limitations of AI; and pioneering
approaches for AI test, evaluation, verification, and validation.
Concurrently, the JAIC will implement AI as required to maintain
battlefield overmatch in alignment with the national defense strategy.
There is no inherent contradiction between the ethical and safe
integration of artificial intelligence across the range of military
operations, and meeting the Department's enduring mission to deter war
and protect the security of our nation. Artificial intelligence is a
critical capability to ensure we field a lethal, resilient, and rapidly
adapting Joint Force. The responsibility for the ethical use of AI will
be widely shared across the Department, spanning system developers,
trainers, leaders, end users, and more. The breadth of this shared
responsibility is a recognition that many different communities within
DOD have a critical role to play in getting this right, although the
JAIC will play a central role in helping craft policies, guidelines,
and authorities related to the ethical, safe, and lawful use of AI
technologies. The Department is fully committed to the ethical use of
AI from concept through deployment and sustainment. This includes
ensuring sufficient rigor and discipline in test and evaluation as well
as validation and verification. Prior to fielding any AI capabilities,
the Department will ensure sufficient rigor and discipline in test and
evaluation processes and procedures. This will include dynamic feedback
loops that take into account feedback from operational users on
deployed algorithms or models.
component mission initiatives
16. Senator Heinrich. Lt Gen Shanahan, you stated that you expect
to see a rapid growth of Component Mission Initiatives, or CMI's in
fiscal year 2020. Can you talk more about why and what those new CMI's
will look like?
Lt. Gen. Shanahan. Since the establishment of the JAIC, we have
been interacting closely with the military Services and other DOD
Components to explore how the JAIC can support their requests for AI-
enabled solutions to a wide variety of potential challenges. This
includes working with them on potential future CMIs. At present, these
range from operational planning through predictive medicine,
identifying individuals at risk of harming themselves, force
protection, and information operations. We are further developing our
internal JAIC governance framework to shape, prioritize, and recommend
CMIs to ensure there are no duplicative efforts and to ascertain proper
levels of support, from personnel resourcing to funding. Additionally,
the establishment of the JAIC Common Foundation (JCF) will enable us to
better support CMIs in the coming fiscal year.
17. Senator Heinrich. Lt Gen Shanahan, does the budget request
reflect the rapid growth of CMI's you anticipate?
Lt. Gen. Shanahan. Yes. We anticipate allocating about $30 million
to support six to ten CMIs in fiscal year 2020. Dedicated funding for
CMIs will incentivize the services and components to rely on the JAIC
while supporting the maturity of the JAIC Common Foundation (JCF), a
crucial enabling capability for CMIs and National Mission Initiatives
(NMI). As the JCF matures it will become a repository of tools,
frameworks, processes, and data to accelerate new CMIs throughout the
DOD. We expect a growing demand for CMI support through fiscal year
2020 and beyond; the current fiscal year 2020 budget request is
currently projected to accommodate all expected CMI support through the
end of fiscal year 2020.
computer science and recruitment
18. Senator Heinrich. Lt Gen Shanahan, you stated that in terms of
AI, adding funding and people will not by themselves spark the
necessary level of institutional change we need, and that we need a
deeper foundation of people who understand the technology and computer
sciences. Do you believe the Department currently prioritizes
recruitment for computer sciences?
Lt. Gen. Shanahan. There is more work to do in this area. Similar
to the lessons learned across DOD over the past decade with the rapid
growth of cyberspace, there is awareness the Department needs to
accelerate efforts to grow, recruit, retain, and promote people with
the kinds of skills necessary to thrive in an AI-enabled future. This
is a multi-generational challenge, though there are a number of
different near- and medium-term efforts that can begin to improve
current shortfalls. These include targeted recruiting; recruiting
incentives; introducing AI and coding principles in grade schools, high
schools, and military entry-level education and training programs;
retention bonuses; intermediate and advanced training throughout a
career; exchanges with industry; and relying on commercial companies to
accelerate the breadth and depth of AI experience across the DOD. The
Services' Reserve Officer Training Corps and Service Academies remain
committed to producing future officers with strong academic foundations
in computer science. Moreover, the Services collaborate with industry
leaders to advance the essential skills needed by these officers in
many technical fields. For the enlisted force, the Services primarily
select individuals with related qualifications and train them
internally for AI-oriented positions. Specialized tests are
administered by the Services to identify individuals with an aptitude
for assignment into computer career fields. Our civilian workforce also
helps to meet the needs of the Department not fulfilled by military
personnel. Concomitant with efforts across DOD to close significant
gaps in AI knowledge and expertise, we need a whole-of-society approach
to further develop a cadre of people with the requisite skills in AI
and AI-associated fields. Absent such an approach, the DOD will
continue to compete with industry, academia, and other government
agencies for an extremely limited pool of AI expertise. It is vital to
grow the entire base of expertise across society, rather than focus
only on the very small pool of AI talent that already exists within
DOD.
19. Senator Heinrich. Lt Gen Shanahan, what are your thoughts on
establishing ``computer sciences'' as a core competency within the
military? (i.e., career tracks, mission specialties)
Lt. Gen. Shanahan. Under Title 10 of the United States Codes the
responsibility for training falls within the Services. Each Service
reviews and updates their core competencies to best reflect the
national defense strategy, changes in the operational environment, and
available technology. Given the immediate need for substantially more
expertise in AI and AI-related skills--from computer scientists to data
architects, data engineers, system engineers and more--I am the
strongest possible proponent for an approach in which the Services
consider how best to represent this need in their core competencies.
The JAIC will work closely with the Services, OSD, and DOD Components
to propose and advance necessary authorities, policies, and legislation
designed to grow a more qualified AI force.
20. Senator Heinrich. Lt Gen Shanahan, given how important STEM
[Science, Technology, Engineering, and Math] is to your efforts, should
there be someone in the Department entrusted with recruiting
individuals specifically for that type of skill set?
Lt. Gen. Shanahan. Each Service is best positioned to determine the
optimum approach to meet their recruiting mission and training
requirements. As the recruiting environment has become more
challenging, I am confident, if deemed appropriate, each Service will
consider other initiatives, including specialty recruiting, that will
enable them to achieve their recruiting missions. I view one of the
JAIC's key responsibilities as helping the Services and other
Components assess their AI workforce strengths, limitations, and
shortfalls, and helping them craft strategies centered on AI talent
recruiting, retention, and promotion.
breakout of dod ai funding
21. Senator Heinrich. Lt Gen Shanahan, the fiscal year 2020 budget
request includes a $927 million investment in Artificial Intelligence,
with money directed toward the Joint Artificial Intelligence Center and
Advanced Image Recognition (Project Maven). Can you please provide the
breakdown on how the $927 million will be allocated?
Lt. Gen. Shanahan. The $927 million is allocated among three major
AI programs in DOD: Joint Artificial Intelligence Center ($268
million), Project Maven ($250 million), and Defense Advanced Research
Projects Agency (DARPA) ($409 million). (The overall budget for all DOD
AI programs is classified and can be provided separately upon request.)
The Joint Artificial Intelligence Center was established in June 2018
to accelerate the delivery of AI-enabled capabilities. Project Maven's
efforts are focused on ISR full motion video and expanding to other
areas that include high altitude still imagery and captured enemy
materials. Through its ``AI Next'' campaign, DARPA is developing
advanced AI theory and applications making it possible for machines to
adapt contextually to changing situations.
22. Senator Heinrich. Lt Gen Shanahan, one impediment to accurately
evaluating funding levels for AI is the lack of a stand-alone AI
Program Element (PE) in DOD funding tables. As a result, AI R&D
[Research and development] appropriations are spread throughout
generally titled PEs and incorporated into funding for larger systems
with AI components. As the lead coordinating entity for AI, do you have
visibility on the Department of Defense's total investment in AI,
including Services and components, and can you provide that breakdown
to the Committee by PE and amounts?
Lt. Gen. Shanahan. I agree that we need better visibility on the
entire range of spending on AI and AI-related projects across DOD. The
Department of Defense's total AI funding was classified Secret by the
Acting Secretary of Defense and can be provided separately upon
request. Current procedures require the military Services and DOD
Components to coordinate with the JAIC on all AI-related projects over
$15 million. For fiscal year 2020, we are relying on data calls by OSD-
CAPE and USD-Comptroller to gather all AI-related funding information.
DOD CIO will capture AI investments in the IT Budget starting in fiscal
year 2021. Through updates in policy and institution of DOD AI
governance, Components will be mandated to enter their AI projects in
the DOD portal to be reflected in the Select & Native Programming Data
Input System for Information Technology SNaP-IT exhibit. These are
necessary first steps towards gaining a more comprehensive
understanding of all AI-related spending across the Department. As part
of our evolving JAIC governance framework, we will propose new
processes and procedures to ensure greater oversight of DOD-wide AI
spending; this will include spending by the DOD components of the
intelligence community.
__________
Questions Submitted by Senator Gary C. Peters
google workforce objections to maven
23. Senator Peters. Lt. Gen. Shanahan, the decision by Google to
pull out the Maven program raises many concerns. What is your
assessment of why Google pulled out of the Maven project and what steps
are you taking and what steps are these companies taking to ensure that
we can work with the best minds in Silicon Valley on these important
efforts?
Lt. Gen. Shanahan. Google's contribution to Project Maven was as a
subcontractor to a prime contractor. They were one of many
subcontractors associated with the project and they had no direct
contractual relationship with Project Maven or the JAIC. They completed
their statement of work as subcontracted and publicly withdrew from
future consideration of a potential subcontract renewal or extension.
The decision was based on entirely internal factors and Google only
provided to the government limited information as to the private
business considerations of that decision. It would be inappropriate for
the Department to speculate as to the underlying rationale for that
decision as private industry has the right to engage in contracts as
they see fit. Nonetheless, the Google/Maven issue underscores that the
DOD needs to improve its efforts to acquire, configure, and deploy
commercial artificial intelligence technology, which requires adapting
to the culture and business practices of a new, more diverse group of
industry partners. The vast majority of current AI companies do not
rely on DOD as a substantial part of their AI business model. Their
employees may have different motivations and ethical concerns about
engaging in national security work. Continued engagement and two-way
transparency with these companies are vital to maintaining our
competitive military edge and ensuring the Department has access to the
nation's top technology and talent. We need to continue to message that
the United States military is the most important global institution for
the preservation and expansion of widely-shared values, including
privacy, liberty, and equality. Communicating the key messages that the
benefits of AI-enabled technologies include enhancing protection of
U.S. and allied forces, reducing the potential for civilian casualties
and collateral damage, and saving lives. One of our key missions in the
JAIC is to build and strengthen relationships across U.S. AI technology
hubs. The Department wants to work with those companies who will
support the full range of DOD missions, while acknowledging that every
company has an inherent right to rely on its own internal review
mechanisms to decide on whether and how to work with DOD on AI-related
projects.
24. Senator Peters. Mr. Brown, I know DIU is trying to forge
relationships with the private sector as well--do you see a way for us
to work with private sector companies who at the corporate or employee
level has expressed concern about working with the Department of
Defense on our toughest military challenges?
Mr. Brown. Transparent, open and frequent communications with
industry, academia and the public will be critical to our success. DIU
was designed to rebuild some of these bridges, develop new
relationships and facilitate conversations that would have commercial
solutions providers engage with the Department of Defense.
value of long term research
25. Senator Peters. Dr. Highnam and Mr. Brown, there is a lot of
understandable and deserved interest in the great advances that the
commercial sector is making in AI. I know that the automotive industry
is making great strides in developing new AI-enabled systems to support
driverless cars and intelligent systems in traditional passenger cars.
But I think that the Federal Government and this committee can play an
important role in creating the environment and funding the research for
that commercial innovation to thrive. Can you describe the role that
federally-funded and university research plays in supporting commercial
sector and Silicon Valley innovation, and what more should we be doing
at the federal level to continue that support?
Dr. Highnam. In terms of fostering commercial innovations in AI-
enabled systems, DARPA has historically played the role of making
pivotal early technology investments in high risk, high payoff
technologies. For example, DARPA's early investments in computer
vision, machine learning, and autonomous control in the Grand Challenge
and Urban Challenge programs spurred the development of technologies
that led to the self-driving car technologies being pursued in the
commercial sector today. (DARPA's earlier investments in the same
technical areas in the early 1980s created the basis for the
Challenges.) Given the growing role and importance of AI technologies
in our daily life, fostering an even broader, more robust AI research
ecosystem is an area where the Federal Government, including DARPA, can
continue to play an important role by: 1) funding early stage research
that inspires commercial application development; 2) funding academia
to ensure a talent pipeline for commercial AI organizations; and 3)
funding programs that address new or previously unrecognized research
issues that often arise when technology meets real world deployments.
Mr. Brown. Galvanizing a generational investment in science and
technology is one of the most important proactive steps we can take in
a technology race to ensure we reinforce the best aspects of U.S.
innovation and entrepreneurial behavior. This requires a foundation of
leading science and technology development and a return to the
technology prominence the U.S. enjoyed in the 1960s--developed as a
response to the previous tech race that we won. The nation, including
the commercial technology ecosystems around the country, are still
benefiting from the technology breakthroughs that came from this
investment in basic science and research in the 1960s and 1970s
(internet, GPS, semiconductors, etc.). The field of AI and Deep
Learning in particular has advanced through public challenges such as
ImageNet, which was hosted by Stanford University and got teams to
compete on building algorithms to identify various images. DIU is
building on this legacy with xView, a series of competitions that
challenge participants to build algorithms applicable to human
assistance and disaster response. The upcoming iteration of xView will
challenge participants to build algorithms to do automated building
damage assessment, which is critical in post-disaster environments.
These challenges not only deliver cutting edge capabilities to the
department, but also spur research interest in areas critical to DOD.
ai workforce
26. Senator Peters. Lt. Gen. Shanahan, for us to move forward in AI
we are going to need a workforce, both in uniform and on the civilian
side, that is capable of being a smart user, buyer, regulator, and
researcher of AI systems. What steps are you taking to build that
workforce and do you need any special hiring, pay authorities or other
authorities in order to compete for AI talent with the private sector?
Lt. Gen. Shanahan. The JAIC is challenged with the hyper-
competitive market for top talent, specifically our difficulty in
offering more competitive compensation packages. One of the key
attributes of the Cyber Excepted Service (CES) is the Departments
ability to offer more competitive market based compensation packages.
While CES resolves many of the key issues, CES does not resolve all the
concerns. To that end, the JAIC would benefit from Congressional
assistance in the following two areas: Relief for the current federal
pay cap ($166,500; Executive Level IV). Authorities to offer retention
incentives internally to DOD employees. Specifically, we would benefit
from the ability to offer incentives for DOD employees moving
internally to the Department. Current authorities only allow for
retention incentives to be given if the employee is leaving Government.
need for good data for ai systems:
27. Senator Peters. Lt. Gen. Shanahan, AI and machine learning
systems require data sets to learn from, and DOD often struggles in
collecting, clean and useful data sets related to their challenges. AI
and machine learning systems can only learn and be more useful if they
are given raw material, namely data to process and work on to begin to
better answer questions and perform required processes. But the
government in general and DOD specifically has a reputation for not
having good data on either its operational or back office systems and
processes, and especially for not sharing that data with others. How is
the JAIC working to create policies on the generation and sharing of
data for AI systems to learn from and get better?
Lt. Gen. Shanahan. Data is the fuel of AI. The Department
recognizes that bad, dirty, or unshared data is a major impediment to
implementing AI at scale. We need to move from a software-based to a
data-first paradigm, refocusing on data as the product and establishing
data-oriented architectures. With particular attention to designing AI-
ready data-driven workflows. Over the past year, the Office of the
Chief Management Officer and OUSD (Comptroller) have been collaborating
to extract, standardize, and curate data from the Department's
operational and back office systems and processes. This data is now
resident in the Defense Repository of Common Enterprise Data (DRCED)
soon to be known as ADVANA. DRCED includes a detailed cost baseline of
the Department's Reform Lines of Business, comprising 75 percent of the
Department's unclassified budget. In addition to this catchment of
data, DRCED provides users with a selection of tools for data
wrangling, data analysis, and data visualization, and machine learning.
As Project Maven demonstrated, and as we learned immediately with our
initial two National Mission Initiatives (NMI), data management is one
of the most resource-intensive and time-consuming aspects of the AI
delivery pipeline. Success in any DOD AI project will rest on a
foundation of good data management. We acknowledge the level of effort
involved in improving DOD-wide data management; one of our key
responsibilities within the JAIC will be to work closely with CMO and
all other stakeholders to address and solve data management problems,
and to serve as a central hub for helping the Services and Components
do the same. This will include establishing standards, setting new
policies, and providing the necessary authorities to begin to shift the
Department in the right direction for future data management. For some
systems, the processes are not yet ready for AI since they are not
data-driven. We will work with our mission partners to design workflows
that are data driven--where users visualize data to perform their tasks
and record information during tasks. These systems can perform many
task with many users across the operation units--aggregating
information for each task level. We are adopting a user-based design
principal for all of our systems--where end-users are engaged in
designing, testing, and providing feedback so that the best AI-enabled
delivery is possible. Once the user workflows are data driven, then AI
can be added to augment the workflow. The JAIC will use tools,
libraries, and framework resources provided by the JAIC Common
Foundation (JCF) to build AI/ML to enhance the data driven workflows
through our NMIs and CMIs. In addition, each NMI/CMI contributes to the
data, models, libraries, and services in the JCF to be shared to the
community at large
research community data
28. Senator Peters. Dr. Highnam and Mr. Brown, I imagine the
research community has challenges with accessing, maintaining, and
sharing useful datasets, especially with respect to classified data.
How do your research teams get the data they need to develop new AI
systems and concepts and do we need to make any policy adjustments to
make that more streamlined?
Dr. Highnam. DARPA AI research programs often need data that may be
unclassified or classified. In some cases, it may be possible for a new
program to re-use existing data from other on-going or completed
programs, and in some cases the new program may need to create new
data. A special case arises when we need to generate unclassified
synthetic data as a surrogate for classified real-world data in order
to permit participation by researchers at universities. Separately, we
anticipate that the recently established DOD Joint Artificial
Intelligence Center (JAIC), with part of its mission to establish a
common set of AI standards, tools, shared data, reusable technology,
etc. will be a step towards improving accessibility of datasets to the
AI research community. DARPA has a long history of working data issues
and does not seek specific policy adjustments at this time.
Mr. Brown. For machine learning efforts to be successful, large
volumes of data are necessary to train algorithms. Large datasets do
exist, but there are challenges leveraging them for a variety of
reasons--some of which are cultural, some of which are regulatory.
While we have the obligation to safeguard unauthorized disclosure of
data, we also have an urgent need to bring machine learning
capabilities into the Defense Department. We would welcome more
comprehensive data-sharing agreements across DOD organizations to
streamline data access challenges that often slow the development of
critical technology. In addition, clearances pose a barrier to getting
some of the best companies and people working on DOD's most important
problems. Efforts to streamline and accelerate the clearance process
would be beneficial.