[Senate Hearing 114-562]
[From the U.S. Government Publishing Office]
S. Hrg. 114-562
THE DAWN OF ARTIFICIAL INTELLIGENCE
=======================================================================
HEARING
BEFORE THE
SUBCOMMITTEE ON SPACE, SCIENCE,
AND COMPETITIVENESS
OF THE
COMMITTEE ON COMMERCE,
SCIENCE, AND TRANSPORTATION
UNITED STATES SENATE
ONE HUNDRED FOURTEENTH CONGRESS
SECOND SESSION
__________
NOVEMBER 30, 2016
__________
Printed for the use of the Committee on Commerce, Science, and
Transportation
[GRAPHIC NOT AVAILABLE IN TIFF FORMAT]
U.S. GOVERNMENT PUBLISHING OFFICE
24-175 PDF WASHINGTON : 2017
_________________________________________________________________________________________
For sale by the Superintendent of Documents, U.S. Government Publishing Office,
http://bookstore.gpo.gov. For more information, contact the GPO Customer Contact Center,
U.S. Government Publishing Office. Phone 202-512-1800, or 866-512-1800 (toll-free).
E-mail, [email protected].
SENATE COMMITTEE ON COMMERCE, SCIENCE, AND TRANSPORTATION
ONE HUNDRED FOURTEENTH CONGRESS
SECOND SESSION
JOHN THUNE, South Dakota, Chairman
ROGER F. WICKER, Mississippi BILL NELSON, Florida, Ranking
ROY BLUNT, Missouri MARIA CANTWELL, Washington
MARCO RUBIO, Florida CLAIRE McCASKILL, Missouri
KELLY AYOTTE, New Hampshire AMY KLOBUCHAR, Minnesota
TED CRUZ, Texas RICHARD BLUMENTHAL, Connecticut
DEB FISCHER, Nebraska BRIAN SCHATZ, Hawaii
JERRY MORAN, Kansas EDWARD MARKEY, Massachusetts
DAN SULLIVAN, Alaska CORY BOOKER, New Jersey
RON JOHNSON, Wisconsin TOM UDALL, New Mexico
DEAN HELLER, Nevada JOE MANCHIN III, West Virginia
CORY GARDNER, Colorado GARY PETERS, Michigan
STEVE DAINES, Montana
Nick Rossi, Staff Director
Adrian Arnakis, Deputy Staff Director
Jason Van Beek, General Counsel
Kim Lipsky, Democratic Staff Director
Chris Day, Democratic Deputy Staff Director
Clint Odom, Democratic General Counsel and Policy Director
----------
SUBCOMMITTEE ON SPACE, SCIENCE, AND COMPETITIVENESS
TED CRUZ, Texas, Chairman GARY PETERS, Michigan, Ranking
MARCO RUBIO, Florida EDWARD MARKEY, Massachusetts
JERRY MORAN, Kansas CORY BOOKER, New Jersey
DAN SULLIVAN, Alaska TOM UDALL, New Mexico
CORY GARDNER, Colorado BRIAN SCHATZ, Hawaii
STEVE DAINES, Montana
C O N T E N T S
----------
Page
Hearing held on November 30, 2016................................ 1
Statement of Senator Cruz........................................ 1
Prepared statement of Dr. Andy Futreal, Chair, Department of
Genomic Medicine, The University of Texas MD Anderson
Cancer Center.............................................. 5
Statement of Senator Peters...................................... 2
Statement of Senator Nelson...................................... 4
Statement of Senator Schatz...................................... 42
Statement of Senator Thune....................................... 44
Statement of Senator Daines...................................... 47
Witnesses
Eric Horvitz, Technical Fellow and Director, Microsoft Research--
Redmond Lab, Microsoft Corporation; Interim Co-Chair,
Partnership on Artificial Intelligence......................... 7
Prepared statement........................................... 8
Dr. Andrew W. Moore, Dean, School of Computer Science, Carnegie
Mellon University.............................................. 17
Prepared statement........................................... 19
Greg Brockman, Co-Founder and Chief Technology Officer, OpenAI... 26
Prepared statement........................................... 28
Dr. Steve A. Chien, Technical Group Supervisor, Artificial
Intelligence Group, Jet Propulsion Laboratory, National
Aeronautics and Space Administration........................... 31
Prepared statement........................................... 33
Appendix
Letter dated November 30, 2016 to Hon. Ted Cruz and Hon. Gary
Peters from Marc Rotenberg, EPIC President and James Graves,
EPIC Law and Technology Fellow................................. 51
THE DAWN OF ARTIFICIAL INTELLIGENCE
----------
WEDNESDAY, NOVEMBER 30, 2016
U.S. Senate,
Subcommittee on Space, Science, and Competitiveness,
Committee on Commerce, Science, and Transportation,
Washington, DC.
The Subcommittee met, pursuant to notice, at 2:35 p.m., in
room SR-253, Russell Senate Office Building, Hon. Ted Cruz,
Chairman of the Subcommittee, presiding.
Present: Senators Cruz [presiding], Thune, Daines, Peters,
Nelson, and Schatz.
OPENING STATEMENT OF HON. TED CRUZ,
U.S. SENATOR FROM TEXAS
Senator Cruz. This hearing will come to order.
Good afternoon. Welcome to each of the witnesses. Thank you
for joining us. Thank you everyone, for attending this hearing.
Throughout history, mankind has refused to accept the
complacency of the status quo and has instead looked to harness
creativity and imagination to reshape the world through
innovation and disruption. The Industrial Revolution, Henry
Ford's moving assembly line, the invention of flight and
commercial aviation, and, more recently, the creation of the
Internet have all acted as disruptive forces that have not only
changed the way we live, but have been engines for commerce
that have offered consumers enormous freedom.
Today, we're on the verge of a new technological
revolution, thanks to the rapid advances in processing power,
the rise of big data, cloud computing, mobility due to wireless
capability, and advanced algorithms. Many believe that there
may not be a single technology that will shape our world more
in the next 50 years than artificial intelligence. In fact,
some have observed that, as powerful and transformative as the
Internet has been, it may be best remembered as the predicate
for artificial intelligence and machine learning.
Artificial intelligence is at an inflection point. While
the concept of artificial intelligence has been around for at
least 60 years, more recent breakthroughs, such as IBM's chess-
playing Deep Blue victory over world champion Gary Kasparov,
advancements in speech recognition, the emergence of self-
driving cars, and IBM's computer Watson's victory in the TV
game show Jeopardy have brought artificial intelligence from
mere concept to reality.
Whether we recognize it or not, artificial intelligence is
already seeping into our daily lives. In the healthcare sector,
artificial intelligence is increasingly being used to predict
diseases at an earlier stage, thereby allowing the use of
preventative treatment, which can help lead to better patient
outcomes, faster healing, and lower costs. In transportation,
artificial intelligence is not only being used in smarter
traffic management applications to reduce traffic, but is also
set to disrupt the automotive industry through the emergence of
self-driving vehicles. Consumers can harness the power of
artificial intelligence through online search engines and
virtual personal assistants via smart devices, such as
Microsoft's Cortana, Apple's Siri, Amazon's Alexa, and Google
Home. Artificial intelligence also has the potential to
contribute to economic growth in both the near and long term. A
2016 Accenture report predicted that artificial intelligence
could double annual economic growth rates by 2035 and boost
labor productivity by up to 40 percent.
Furthermore, market research firm Forrester recently
predicted that there will be a greater-than-300-percent
increase in investment in artificial intelligence in 2017
compared to 2016. While the emergence of artificial
intelligence has the opportunity to improve our lives, it will
also have vast implications for our country and the American
people that Congress will need to consider, moving forward.
Workplaces will encounter new opportunities, thanks to
productivity enhancements. As artificial intelligence becomes
more pervasive, Congress will need to consider its privacy
implications. There is also a growing interest in this
technology from foreign governments who are looking to harness
this technology to give their countries a competitive advantage
on the world stage.
Today, the United States is the preeminent leader in
developing artificial intelligence. But, that could soon
change. According to The Wall Street Journal, ``The biggest
buzz in China's Internet industry isn't about besting global
tech giants by better adapting existing business models for the
Chinese market; rather, it's about competing head-to-head with
the U.S. and other tech powerhouses in the hottest area of
technological innovation: artificial intelligence.'' Ceding
leadership in developing artificial intelligence to China,
Russia, and other foreign governments will not only place the
United States at a technological disadvantage, but it could
have grave implications for national security.
We are living in the dawn of artificial intelligence. And
it is incumbent that Congress and this subcommittee begin to
learn about the vast implications of this emerging technology
to ensure that the United States remains a global leader
throughout the 21st century. This is the first congressional
hearing on artificial intelligence. And I am confident it will
not be the last, as this growing technology raises
opportunities and potential threats at the same time.
I look forward to hearing from our distinguished panel of
experts today.
And, at this point, I'll yield to our subcommittee's
Ranking Member, Senator Peters, to give an opening statement.
STATEMENT OF HON. GARY PETERS,
U.S. SENATOR FROM MICHIGAN
Senator Peters. Well, thank you, Chairman Cruz, for calling
this very important meeting.
And I'd like to thank the witnesses for taking the time, as
well. And look forward to hearing your expertise on this
exciting area.
You know, certainly, our Nation's history has always been
defined, in my mind, by our ongoing search for the next
innovation, the next big thing that's going to advance our
economy and our society. Our mastery of manufacturing and
automation helped the United States establish its industrial
and military might.
Looking to the future, we must continue to harness this
American drive to discover, to create the next big advancement,
and to keep our Nation on the cutting edge. Over the last few
decades, technology has changed the way we make and how we make
it. Today, we boast an ever-evolving innovation ecosystem
that's rooted in robotics, machine learning, and, the subject
of today's hearing, artificial intelligence.
AI products like our smartphones, intelligent personal
assistant, or our banks' fraud detection technology are already
improving the day-to-day lives of Americans. New advances in
computer processing and cloud computing are driving rapid
expansion in AI development in industries as diverse as
healthcare, transportation, education, and security. And
they've all gone through enormous change, as all of you know.
We as a society need to help foster this broader ecosystem
so we can capitalize on its benefits, including cleaner energy,
new economic growth, improved safety and health, and greater
accessibility for the elderly and disabled populations.
Being from Michigan, I've had the opportunity to experience
firsthand the development of self-driving cars that are made
possible by a combination of advanced automatic braking and
lane-changing systems, cameras, sensors, high-performance
computing, deep learning systems, 3D high-definition maps, and
artificial intelligence.
Just last year, over 35,000 people died in motor vehicle
crashes, but research suggests that about 94 percent of those
accidents were the result of human error. With safe deployment,
AI-fueled automated vehicles could significantly decrease this
number, saving countless thousands of lives.
In addition, the World Economic Forum estimates that the
digital transformation of the automotive industry alone--just
the automotive industry alone--will generate $67 billion in
value for the sector and yield $3.1 trillion in societal
benefits before 2025. That's billions injected into our economy
to boost our competitiveness, and billions saved due to major
reductions in auto accident injuries and deaths, environmental
degradation caused by traffic congestion and air pollution.
More broadly, U.S. technology companies spent eight and a half
billion dollars on AI in 2015, more than four times the amount
spent in 2010. And experts predict incredible growth in the
coming years in the healthcare, marketing, and finance sectors.
A critical component of this subcommittee's work is
promoting and preserving American competitiveness. And, while
the U.S. leads the world in investment and discoveries in AI,
as we all know and as Chairman mentioned, China is quickly
catching up. In 2014, Chinese scientists overtook U.S.
scientists in terms of new papers published and citations of
published papers in the area of deep learning, a cutting edge
of AI research.
Like computer processing, nanotechnology, and
biotechnology, AI has the power to dramatically change and grow
the economy, but we must invest in research and in STEM
education to maintain our competitive advantage. Analysts by
the Council of Economic Advisors shows that doubling or
tripling all research investment, something that I fully
support, will allow us to continue to grow. However, targeting
increases in areas of high economic productivity, like AI, may
offer benefits with much smaller budgetary impact. In fact, a
recent report from the White House recommends that current
government spending on AI research should be doubled or even
quadrupled to achieve optimal economic growth. I certainly look
forward to hearing your comments on that.
We've already seen enormous gains from Federal investment
in this area through NASA's work to develop AI applications for
use in robotic spacecraft missions. These applications include
planning, spacecraft autonomy, image processing, and rover
autonomy. NASA also utilizes AI and Earth-observing satellites
to optimize observations of natural disasters, like volcanic
eruptions. And I look forward to hearing more about this
pioneering work by NASA and the ways in which other industries
have benefited from date.
And finally, while we must strive to optimize the full
economic potential of AI, we must also address its potential
impacts on the workforce. While new jobs will be created
because of AI, we also have to think critically about the steps
we can take today and in coming years to make sure that
American workers are not left behind.
The Subcommittee plays a unique role in studying emerging
technologies and examining ways to promote and harness
scientific advancement for the greater good. I appreciate the
Chairman's interest in this important issue. I look forward to
working with him and continuing to advance development at AI
for the greater good. And that's why I look with great
anticipation to the testimony from each of you.
Thank you.
Senator Cruz. Thank you, Senator Peters.
We'd now recognize the Ranking Member of the full
committee, Senator Nelson, for an opening statement.
STATEMENT OF HON. BILL NELSON,
U.S. SENATOR FROM FLORIDA
Senator Nelson. Thank you, Mr. Chairman.
Indeed, AI has helped the space program quite a bit. I'm
delighted that a representative from JPL is here. JPL are the
wiz kids, the rock stars. And, in fact, they are rocket
scientists.
There's another part about AI, and that is the replacement
of jobs. We've got to prepare for that. For example, Elon Musk
recently predicted, in an interview with CNBC, that robots
could eventually take many jobs away from folks and that they
would have to depend on the government in order to have a
living. Elon used the example of truck drivers, who could be
displaced in the future by autonomous vehicles and those kind
of advancements that would allow trucks to drive themselves.
And yet, if a whole occupation is suddenly displaced, what do
we do? We just came through an election where the loss of jobs
was a big topic. Maybe truck drivers don't want to be trained
to go on a computer. So, what are we going to do for the
future? This is just another challenge that we face as
technology advances. And that's why we're here today.
So, thanks. I'm looking forward to it.
Senator Cruz. Thank you, Senator Nelson.
I'd now like to introduce our witnesses.
The first witness is Dr. Eric Horvitz, who is the interim
Co-Chair of the Partnership on Artificial Intelligence and
serves as the Managing Director of the Microsoft Research
Redmond Lab. Dr. Horvitz's research contributions span
theoretical and practical challenges with computing systems
that learn from data and that can perceive, reason, and decide.
His efforts have helped to bring multiple systems and services
into the world, including innovations in transportation,
healthcare, aerospace, e-commerce, online services, and
operating systems.
Dr. Andrew Moore is the Dean of the School of Computer
Science at Carnegie Mellon University. Dr. Moore's background
is in statistical machine learning, artificial intelligence,
robotics, and statistical computation of large volumes of data,
including decision and control algorithms.
Mr. Greg Brockman is the Cofounder and Chief Technology
Officer of OpenAI, a nonprofit artificial intelligence research
company. Prior to OpenAI, Mr. Brockman was the CTO of Stripe, a
financial technology company that builds tools enabling Web
commerce.
And Dr. Steve Chien is the Technical Group Supervisor of
the Artificial Intelligence Group and the Senior Research
Scientist in the Mission Planning and Execution Section at
NASA's Jet Propulsion Laboratory.
I would also note that Dr. Andrew Futreal was set to
testify at our hearing today, but, unfortunately, due to
weather issues, he was unable to travel to D.C. today. Dr.
Futreal is the Chair of the Department of Genomic Medicine at
The University of Texas MD Anderson Cancer Center, in my
hometown of Houston. We thank him for his previously submitted
written testimony. And if there are no objections, I would like
to submit Dr. Futreal's testimony into the hearing record.
[The prepared statement of Dr. Futreal follows:]
Prepared Statement of Dr. Andy Futreal, Chair, Department of Genomic
Medicine, The University of Texas MD Anderson Cancer Center
Subcommittee Chairman Cruz, Ranking Member Peters, and members of
this committee, thank you all very much for the opportunity to testify
before you today. My name is Andy Futreal and I am Chair of the
Department of Genomic Medicine at The University of Texas MD Anderson
Cancer Center.
We are now entered into a completely unprecedented time in the
history of medicine. We have the ability to investigate the fundamental
molecular underpinnings of disease, to leverage technology and
computational capabilities with the real prospect of fundamentally
altering the natural history of disease. We can now determine each
individual's genetic blueprint with relative speed and accuracy at a
cost of less than a millionth of the price tag of the first human
genome sequenced just a little more than 13 years ago. We are moving
into an era of tackling the sequencing of very large groups of
individuals and defining the role of common variation, that which is
shared by more than 1-5 percent of the population, in health, risk and
disease. The challenge of reducing this watershed of data into
practical implementation to improve human health and provide better
care for patients is upon us. The opportunities to improve and tailor
healthcare delivery--the right drug for the right patient at the right
time with the right follow-up--are being driven by exploiting
computational approaches and so-called ``big data''. AI and machine
learning approaches have the potential to help drive insights and
deliver improved standards of care. Taking oncology as the proving
ground where a very great deal of these efforts are currently focused,
there are several challenges, opportunities and issues that present
themselves.
The clinically meaningful implementation of machine-assisted
learning and AI is, of course, crucially dependent on data--lots of it.
Herein lies perhaps the biggest challenge. Substantial and varied
clinical data is generated on every patient cared for every day. These
data are generally held in non-interoperable systems whose principle
purpose is to facilitate tracking of activities/services/tests for
billing purposes. The richest clinical data is effectively locked in
various dictated and transcribed notes detailing patients' clinical
course, responses, problems and outcomes from the various treatments/
interventions undertaken. We need to further develop capabilities to
both get these data from their source systems and standardize their
ongoing collection as practically as possible.
As well, a proportion of those under our care take part in research
studies, generating research data in both the clinical and more
translational/basic science realms. These data, including increasing
amounts of detailed large-scale genomic sequencing information, are not
generally available for integration with clinical data on a per-patient
or aggregate basis in a way that would facilitate implementation of
advanced analytics. The ability to purposefully integrate clinical and
research data for analytics, without the need for predetermining and
rigidly standardizing all data inputs up front is what is needed.
There are substantial opportunities for AI, again anchoring in
oncology by way of example. Perhaps the most concise way of framing
where we need to be headed, in my view, is the concept of real-time
``patients like mine'' analytics. Leveraging clinical, molecular,
exposure and lifestyle data of patients that have been treated before
to understand and predict what the best choices are for the current
patient. But even more so, not just choice of therapeutic but how to
improve and intercede as needed in management such that positive
outcome chances are maximized. We need to make predictive analytics the
norm, learning from every patient to improve the outcome of the next.
Importantly, we need to be thinking now about training our best and
brightest in the next generation of physicians and medical
professionals to drive this progress, as it will take a new wave of
computationally savvy individuals to build, train and grow these
systems. Further, we need to think carefully about how we promote data
sharing, particularly in the clinical arena. Open access is a laudable
goal, but one that must be tempered with the relevant privacy and
security practices. Facilitated collaboration on specific topics with
honest broker mechanisms to demonstrate rapid progress and real value
in data sharing early will, I think, be key.
At MD Anderson, we have been exploring the possible utilities of AI
and related technologies in collaboration with IBM. We are utilizing
the Watson platform for cognitive computing to train an expert system
for patient-centric treatment recommendation and management. Currently,
we are evaluating performance in the context of lung cancer. Future
work reflects the challenges and opportunities that the entire field
faces--namely that of what to deploy in the near-term where
dissemination of expert knowledge in the context of rule-based
approaches could have significant impact on potentially improving
standard of care and where to take efforts in the longer term with
learning, AI type approaches.
The ability to have data-driven, AI empowered point-of-care
analytics holds the promise of improving the standard of care in
medically underserved areas, of guaranteeing that every patient--
regardless of zip code--can be assured of up-to-date and appropriate
care taking into account their own particular data and circumstance. A
massive undertaking to be sure, but one that is, I believe, within our
collective grasp.
I thank you again for the opportunity to testify before this
committee and I would be happy to answer any questions you may have.
Senator Cruz. With that, we will move to the testimony,
although I will note for each of you that your experience and
wisdom in this topic will be welcomed and sorely needed, and
intelligence, artificial or otherwise, is not something we deal
with often, because this is the United States Congress.
[Laughter.]
Senator Cruz. And with that, Dr. Horvitz, you may give your
testimony.
STATEMENT OF ERIC HORVITZ, TECHNICAL FELLOW AND
DIRECTOR, MICROSOFT RESEARCH--REDMOND LAB,
MICROSOFT CORPORATION; INTERIM CO-CHAIR, PARTNERSHIP ON
ARTIFICIAL INTELLIGENCE
Dr. Horvitz. Thank you, Chairman Cruz, Ranking Member
Peters, and members of the Subcommittee. And good afternoon.
And thank you for hosting this discussion on AI. It's fabulous.
To start, AI is not one thing. AI is a constellation of
disciplines, including computer vision, machine learning,
language understanding, reasoning and planning, and robotics,
but they're all aimed at a shared aspiration, the scientific
understanding of thought and intelligent behavior, and in
developing computing systems based on these understandings.
Many advances over the 60-year history of AI have now--are
now actually part of our daily lives. Just consider the AI
route-planning algorithms we use daily in our GPS systems. And,
while we've seen many advances over time, it's clear that we're
now at an inflection point. We're seeing an acceleration of AI
competencies in many areas. And the inflection is driven by a
confluence of several factors. And these include the
unprecedented quantities of data that have come available with
the widespread digitization of our lives, increases in
computing power over time, and the recent jumps in the prowess
of our algorithms, the methods we use, particularly machine-
learning methods that learn to predict and to diagnose from
data.
The advances are putting unprecedented technologies in the
hands of people, including real-time speech-to-speech
translation among languages now available from Microsoft's
Skype services, and computer vision for assisting drivers. AI
technology that is already available today could save thousands
of lives and many billions of dollars if properly translated
into practice. And these key--and the key opportunities before
us include healthcare, transportation, education, as well as
agriculture, manufacturing, and increasing accessibility for
those with special needs. Other directions that are critical
include using AI advances to enhance the resilience and
capacity of critical infrastructure, like our electrical power
grid and road network.
So, let's take as an example healthcare, to start. It's--AI
is a veritable sleeping giant for healthcare. AI technologies
will be extremely valuable for handling acute as well as
chronic illnesses and for making our hospitals safer. As an
example, we've built systems that can predict patient outcomes
and that make patient-specific recommendations to allocate
scarce resources. And these prototypes have been applied to
address such challenges as chronic diseases, hospital
readmissions, hospital-associated infections, and catching
preventable errors in hospitals responsible for over a quarter
of a million deaths per year in the United States.
And, while on the topic of saving lives, advances in
pattern-recognition systems enable us to develop effective
automated braking and control systems that will keep us safer.
We could take a big cut out of those 30,000 deaths that we've
become accustomed to tolerating every year in our country. And
we don't often think about the 300,000 incapacitating injuries
on our roads every year.
Numerous other AI innovations can help with safe driving.
This week, we just published--our team just published results
that show how we can leverage data from Minneapolis to build a
routing system that identifies the safest routes to take,
considering a multitude of factors at any moment, even how the
sun is shining and the glare that it produces.
Moving forward, key research directions in AI include
focusing on human-AI collaboration, ensuring the robustness,
safety, and security of AI systems, and identifying and
addressing the ethical, legal, economic, and broader societal
influences of AI.
As an example of important research, we need to endow
systems with the ability to explain their reasoning to people.
People need to trust systems to use them effectively. And such
trust requires transparency. There's also important work to be
done with developing robust and resilient AI systems,
especially when these systems are used for high stakes, safety-
critical applications amidst the complexities of the open
world.
We must also be aware that AI systems can present new kinds
of attack surfaces that can be disrupted by cyberattacks, so
it's important to address rising cyber vulnerabilities as we
develop and field these more capable AI systems.
Another area of importance is the influence of AI on our
Nation's workforce and economy. And, while estimates vary, the
economic influence of AI will likely be in multiples of
trillions of dollars. However, along with the expected benefits
come concerns with how AI will affect jobs and income
disparities. And these are important issues. We must closely
reflect, monitor, and plan to ensure a smooth transition to a
world where AI plays a more important role.
One thing for sure is that we urgently need to prioritize
the training and retraining of the U.S. workforce so that our
workforce skills are aligned with needs. We also need to double
down on investments in STEM education and training. And, moving
forward, continued strong and public and private sector support
of research and studies on the scientific and socioeconomic
challenges of AI will be critical to ensuring that people in
society get the most out of AI advances.
So, in summary, we expect AI advances to raise our quality
of life, to empower citizens in new ways. But, with--as with
any technical advance, we need to invest efforts to study and
address potential challenges, concerns, inequities that may
come along with the benefits that we expect from AI.
Thanks very much.
[The prepared statement of Dr. Horvitz follows:]
Prepared Statement of Eric Horvitz, Technical Fellow and Director,
Microsoft Research--Redmond Lab, Microsoft Corporation
``Reflections on the Status and Future of Artificial Intelligence''
Chairman Cruz, Ranking Member Peters, and Members of the
Subcommittee, my name is Eric Horvitz, and I am a Technical Fellow and
Director of Microsoft's Research Lab in Redmond, Washington. While I am
also serving as Co-Chair of a new organization, the Partnership on
Artificial Intelligence, I am speaking today in my role at Microsoft.
We appreciate being asked to testify about AI and are committed to
working collaboratively with you and other policymakers so that the
potential of AI to benefit our country, and to people and society more
broadly can be fully realized.
With my testimony, I will first offer a historical perspective of
AI, a definition of AI and discuss the inflection point the discipline
is currently facing. Second, I will highlight key opportunities using
examples in the healthcare and transportation industries. Third, I will
identify the important research direction many are taking with AI.
Next, I will attempt to identify some of the challenges related to AI
and offer my thoughts on how best to address them. Finally, I will
offer several recommendations.
What is Artificial Intelligence?
Artificial intelligence (AI) refers to a set of computer science
disciplines aimed at the scientific understanding of the mechanisms
underlying thought and intelligent behavior and the embodiment of these
principles in machines that can deliver value to people and society.
A simple definition of AI, drawn from a 1955 proposal that kicked
off the modern field of AI, is pursuing how ``to solve the kinds of
problems now reserved for humans.'' \1\ The authors of the founding
proposal on AI also mentioned, ``We think that a significant advance
can be made in one or more of these problems if a carefully selected
group of scientists work on it together for a summer.'' While progress
has not proceeded as swiftly as the optimistic founders of the field
may have expected, there have been ongoing advances over the decades
from the sub-disciplines of AI, including machine vision, machine
learning, natural language understanding, reasoning and planning, and
robotics.
---------------------------------------------------------------------------
\1\ McCarthy, J., Minsky, M.L., Rochester, N., Shannon, C.E. A
Proposal for the Dartmouth Summer Project on Artificial Intelligence,
Dartmouth University, May 1955.
---------------------------------------------------------------------------
Highly visible AI achievements, such as DeepBlue's win over the
world chess champion, have captured the imagination of the public. Such
high-profile achievements have relayed a sense that the field is
characterized by large jumps in capabilities. In reality, research and
development (R&D) in the AI sub-disciplines have produced an ongoing
stream of innovations. Numerous advances have become part of daily
life, such as the widespread use of AI route-planning algorithms in
navigation systems.\2\ Many applications of AI execute ``under the
hood'', including methods that perform machine learning and planning to
enhance the functioning of computer operating systems or to better
retrieve and rank search results. In some cases, AI systems have
introduced breakthrough efficiencies without public recognition or
fanfare. For example, in the mid- to late-1990s leading-edge machine
vision methods for handwriting recognition were pressed into service by
the U.S. Postal Service to recognize and route handwritten addresses on
letters automatically.\3\ High-speed variants of the first machines now
sort through more than 25 billion letters per year, with estimated
accrued savings of hundreds of millions of dollars.
---------------------------------------------------------------------------
\2\ Hart, P.E., Nilsson, N.J., Raphael, B. A Formal Basis for the
Heuristic Determination of Minimum Cost Paths. IEEE Transactions on
Systems Science and Cybernetics Vol. SSC-4, No. 2, July 1968.
\3\ Kim, G and Govindaraju, V., Handwritten Word Recognition for
Real-Time Applications, Proceedings of the Third International
Conference on Document Analysis and Recognition, August 1995.
---------------------------------------------------------------------------
AI at an Inflection Point
Over the last decade, there has been a promising inflection in the
rate of development and fielding of AI applications. The acceleration
has been driven by a confluence of several factors. A key influence
behind the inflection is the availability of unprecedented streams of
data, coupled with drops in the cost of storing and retrieving that
data. Large quantities of structured and unstructured databases about
human activities and content have become available via the digitization
and the shift to the web of activities around commerce, science,
communications, governance, education, and art and entertainment.
Other contributing factors include dramatic increases in available
computing power, and jumps in the prowess of methods for performing
machine learning and reasoning. There has been great activity in the
machine learning area over the last thirty years with the development
of a tapestry of algorithms for transforming data into components that
can recognize patterns, perform diagnoses, and make predictions about
future outcomes. The past thirty years of AI research also saw the rise
and maturation of methods for representing and reasoning under
uncertainty. Such methods jointly represent and manipulate both logical
and probabilistic information. These methods draw from and extend
methods that had been initially studied and refined in the fields of
statistics, operations research, and decision science. Such methods for
learning and reasoning under uncertainty have been critical for
building and fielding AI systems that can grapple effectively with the
inescapable incompleteness when immersed in real-world situations.
Over the last decade, there has been a renaissance in the use of a
family of methods for machine learning known as neural networks.\4\ A
class of these algorithms referred to as deep neural networks are now
being harnessed to significantly raise the quality and accuracy of such
services as automatic speech recognition, face and object recognition
from images and video, and natural language understanding. The methods
are also being used to develop new computational capabilities for end
users, such as real-time speech-to-speech translation among languages
(e.g., now available in Microsoft's Skype) and computer vision for
assisting drivers with the piloting of cars (now fielded in the Tesla's
models S and X).
---------------------------------------------------------------------------
\4\ Neural network algorithms are descendants of statistical
learning procedures developed in the 1950s, referred to as perceptrons.
With neural networks, representations of patterns seen in training data
are stored in a set of layers of large numbers of interconnected
variables, often referred to as ``neurons''. The methods are inspired
loosely (and in a very high-level manner) by general findings about the
layering of neurons in vertebrate brains. Seven years ago, a class of
neural networks, referred to as deep neural networks, developed decades
earlier, were shown to provide surprising accuracies for pattern
recognition tasks when trained with sufficient quantities of data.
---------------------------------------------------------------------------
Key Opportunities
AI applications explored to date frame opportunities ahead for
leveraging current and forthcoming AI technologies. Pressing AI methods
that are currently available into service could introduce new
efficiencies into workflows and processes, help people with
understanding and leveraging the explosion of data in scientific
discovery and engineering, as well as assist people with solving a
constellation of challenging real-world problems.\5\
---------------------------------------------------------------------------
\5\ Multiple AI applications in support of people and society are
presented here: E. Horvitz, AI in Support of People and Society, White
House OSTP CCC AAAI meeting on AI and Social Good, Washington, D.C.,
June 2016. (access video presentation)
---------------------------------------------------------------------------
Numerous commercial and societal opportunities can be addressed by
using available data to build predictive models and then using the
predictive models to help guide decisions. Such data to predictions to
decisions pipelines can deliver great value and help build insights for
a broad array of problems.\6\ Key opportunities include AI applications
in healthcare and biomedicine, accessibility, transportation,
education, manufacturing, agriculture, and for increasing the
effectiveness and robustness of critical infrastructure such as our
electrical power grid.
---------------------------------------------------------------------------
\6\ E. Horvitz. From Data to Predictions and Decisions: Enabling
Evidence-Based Healthcare, Data Analytic Series, Computing Community
Consortium, Computing Research Association (CRA), September 2010.
---------------------------------------------------------------------------
Healthcare and transportation serve as two compelling examples
where AI methods can have significant influence in the short-and
longer-term.
Healthcare. AI can be viewed as a sleeping giant for healthcare.
New efficiencies and quality of care can be obtained by leveraging a
coupling of predictive models, decision analysis, and optimization
efforts to support decisions and programs in healthcare. Applications
span the handling of acute illnesses, longer-term disease management,
and the promotion of health and preventative care. AI methods show
promise for multiple roles in healthcare, including inferring and
alerting about hidden risks of potential adverse outcomes, selectively
guiding attention, care, and interventional programs where it is most
needed, and reducing errors in hospitals.
On-site machine learning and decision support hinging on inference
with predictive models can be used to identify and address potentially
costly outcomes. Let's consider the challenge of reducing readmission
rates. A 2009 study of Medicare-reimbursed patients who were
hospitalized in 2004 found that approximately 20 percent of these
patients were re-hospitalized within 30 days of their discharge from
hospitals and that 35 percent of the patients were re-hospitalized
within 90 days.\7\ Beyond the implications of such readmissions for
health, such re-hospitalizations were estimated to cost the Nation
$17.4 billion in 2004. Studies have demonstrated that predictive
models, learned from large-scale hospital datasets, can be used to
identify patients who are at high risk of being re-hospitalized within
a short time after they are discharged--and that such methods could be
used to guide the allocation of special programs aimed at reducing
readmission.\8\
---------------------------------------------------------------------------
\7\ Coleman, E. Jencks, S., Williams, M. Rehospitalizations among
Patient in the Medicare Fee-for-Service Program, The New England
Journal of Medicine, 380:1418-1428, April 2009.
\8\ Bayati, M., Braverman, M., Gillam, M. Mack, K.M., Ruiz, G.,
Smith, M.S., Horvitz, E. Data-Driven Decisions for Reducing
Readmissions for Heart Failure: General Methodology and Case Study.
PLOS One Medicine. October 2014.
---------------------------------------------------------------------------
AI methods can also play a major role in reducing costs and
enhancing the quality of care for the difficult and ongoing challenge
of managing chronic disorders. For example, congestive heart failure
(CHF) is prevalent and expensive. The illness affects nearly 10 percent
of people over 65 years. Medical costs and hospitalizations for CHF are
estimated to be $35 billion per year in the U.S. CHF patients may hover
at the edge of physiological stability and numerous factors can cause
patients to spiral down requiring immediate hospitalization. AI methods
trained with data can be useful to predict in advance potential
challenges ahead and to allocate resources to patient education,
sensing, and to proactive interventions that keep patients out of the
hospital.
Machine learning, reasoning, and planning offer great promise for
addressing the difficult challenge of keeping hospitals safe and
efficient. One example is addressing the challenge with hospital-
associated infections.\9\ It is estimated that such infections affect
10 percent of people who are hospitalized and that they are a
substantial contributor to death in the U.S. Hospital-associated
infections have been linked to significant increases in hospitalization
time and additional costs of tens of thousands of dollars per patient,
and to nearly $7 billion of additional costs annually in the U.S. The
CDC has been estimated that 90 percent of deaths due to hospital-
associated infections can be prevented. A key direction is the
application of predictive models and decision analyses to estimate
patients' risk of illness and to guide surveillance and other
preventative actions.
---------------------------------------------------------------------------
\9\ Wiens, J., Guttag, J., and Horvitz, E., Patient Risk
Stratification with Time-Varying Parameters: A Multitask Learning
Approach. Journal of Machine Learning Research (JMLR), April 2016.
---------------------------------------------------------------------------
AI methods promise to complement the skills of physicians and
create new forms of cognitive ``safety nets'' to ensure the effective
care of hospitalized patients.\10\ An Institute of Medicine (IOM) study
in 2000 called attention to the problem of preventable errors in
hospitals.\11\ The study found that nearly 100,000 patients die in
hospitals because of preventable human errors. The IOM estimate has
been revised upward by several more recent studies. Studies in October
2013 and in May 2016 estimated that preventable errors in hospitals are
the third leading cause of death in the U.S., only trailing behind
heart disease and cancer. The two studies estimated deaths based in
preventable error as exceeding 400,000 and 250,000 patients per year,
respectively.\12\,\13\ AI systems for catching errors via
reminding and recognizing anomalies in best clinical practices could
put a significant dent in the loss of nearly 1,000 citizens per day,
and could save tens of thousands of patients per year.
---------------------------------------------------------------------------
\10\ Hauskrecht, M., Batal, I., Valko, M., Visweswaran, S., Cooper,
G.F., Clermont, G., Outlier Detection for Patient Monitoring and
Alerting, Journal of Biomedical Informatics, Volume 46, Issue 1,
February 2013, Pages 47-55.
\11\ To Err Is Human: Building a Safer Health System, Institute of
Medicine: Shaping the Future, November 1999.
\12\ James, John T. A New, Evidence-based Estimate of Patient Harms
Associated with Hospital Care, Journal of Patient Safety, September
2013.
\13\ Daniel, M., Makary, M. Medical Error--The Third Leading Cause
of Death in the U.S., BMJ, 353, 2016.
---------------------------------------------------------------------------
The broad opportunities with the complementarity of AI systems and
physicians could be employed in myriad ways in healthcare. For example,
recent work in robotic surgery has explored how a robotic surgeon's
assistant can work hand-in-hand to collaborate on complex surgical
tasks. Other work has demonstrated how coupling machine vision for
reviewing histological slides with human pathologists can significantly
increase the accuracy of detecting cancer metastases.
Transportation. AI methods have been used widely in online services
and applications for helping people with predictions about traffic
flows with doing traffic-sensitive routing. Moving forward, AI methods
can be harnessed in multiple ways to make driving safer and to expand
the effective capacity of our existing roadway infrastructure.
Automated cars enabled by advances in perception and robotics promise
to enhance both flows on roads and to enhance safety. Longer-range
possibilities include the fielding of large-scale automated public
microtransit solutions on a citywide basis. Such solutions could
transform mobility within cities and could influence the overall
structure and layout of cities over the longer-term.
Smart, automated driver alerting and assistance systems for
collision avoidance show promise for saving hundreds of thousands of
lives worldwide. Motor vehicle accidents are believed to be responsible
for 1.2 million deaths and 20-50 million non-fatal injuries per year
each year. NHTSA's Fatality Analysis Reporting System (FARS) shows that
deaths in the U.S. due to motor vehicle injuries have been hovering at
rates over 30,000 fatalities per year. In addition to deaths, it is
important to include a consideration of the severe injuries linked to
transportation. It is estimated that 300,000 to 400,000 people suffer
incapacitating injuries every year in motor vehicles; in addition to
the nearly 100 deaths per day, nearly one thousand Americans are being
incapacitated by motor vehicle injuries every day.
Core errors based in the distraction of drivers and problems with
control lead to road departures and read-end collisions. These expected
problems with human drivers could be addressed with machine perception,
smart alerting, and autonomous and semi-autonomous controls and
compensation. AI methods that deliver inferences with low false-
positive and false-negative rates for guiding braking and control could
be pressed into service to save many thousands of lives and to avoid
hundreds of thousands of life-changing injuries. Studies have found
that a great proportion of motor vehicle accidents are caused by
distraction and that nearly 20 percent of automobile accidents are
believed to be failures to stop. Researchers have estimated that the
use of smart warning, assisted braking, and autonomous braking systems
could reduce serious injuries associated with rear-end collisions by
nearly 50 percent.\14\
---------------------------------------------------------------------------
\14\ Kusano, K.D. and Gabler, H.C., Safety Benefits of Forward
Collision Warning, Brake Assist, and Autonomous Braking Systems in
Rear-End Collisions, IEEE Transactions on Intelligent Transportation
Systems, pages 1546-1555. Volume: 13(4), December 2012.
---------------------------------------------------------------------------
Myriad of opportunities. Healthcare and transportation are only two
of the many sectors where AI technologies offer exciting advances. For
example, machine learning, planning, and decision making can be
harnessed to understand, strengthen, monitor, and extend such critical
infrastructure such as our electrical power grid, roads, and bridges.
In this realm, AI advances could help to address challenges and
directions specified in the Energy Independence and Security Act of
2007 on the efficiency, resilience, and security of the U.S. power
grid. In particular, there is opportunity to harness predictive models
for predicting the load and availability of electrical power over time.
Such predictions can lead to more effective plans for power
distribution. Probabilistic troubleshooting methodologies can jointly
harness knowledge of physical models and streams of data to develop
models that could serve in proactive and real-time diagnoses of
bottlenecks and failures, with a goal of performing interventions that
minimize disruptions.
In another critical sector, AI methods can play an important role
in the vitality and effectiveness of education and in continuing-
education programs that we offer to citizens. As an example, data-
centric analyses have been employed to develop predictive models for
student engagement, comprehension, and frustration. Such models can be
used in planners that create and update personalized education
strategies.\15\,\16\ Such plans could address conceptual
bottlenecks and work to motivate and enhance learning. Automated
systems could help teachers triage and troubleshoot rising challenges
with motivation/engagement and help design ideal mixes of online and
human-touch pedagogy.
---------------------------------------------------------------------------
\15\ K. Koedinger, S. D'Mello., E. McLauglin, Z. Pardos, C. Rose.
Data Mining and Education, Wiley Interdisciplinary Reviews: Cognitive
Science, 6(4): 333-353, July 2015.
\16\ Rollinson, J. and Brunskill, E., From Predictive Models to
Instructional Policies, International Educational Data Mining Society,
International Conference on Educational Data Mining (EDM) Madrid,
Spain, June 26-29, 2015.
---------------------------------------------------------------------------
Key Research Directions
R&D on AI continues to be exciting and fruitful with many
directions and possibilities. Several important research directions
include the following:
Supporting Human-AI collaboration. There is great promise for
developing AI systems that complement and extend human abilities \17\.
Such work includes developing AI systems that are human-aware and that
can understand and augment human cognition. Research in this realm
includes the development of systems that can recognize and understand
the problems that people seek to solve, understanding human plans and
intentions, and to recognize and address the cognitive blind spots and
biases of people.\18\ The latter opportunity can leverage rich results
uncovered in over a century of work in cognitive psychology.
---------------------------------------------------------------------------
\17\ Licklider, J. C. R., ``Man-Computer Symbiosis'', IRE
Transactions on Human Factors in Electronics, vol. HFE-1, 4-11, March
1960.
\18\ Presentation: Horvitz, E., Connections, Sustained Achievement
Award Lecture, ACM International Conference on Multimodal Interaction
(ICMI), Seattle, WA, November 2015.
---------------------------------------------------------------------------
Research on human-AI collaboration also includes efforts on the
coordination of a mix of initiatives by people and AI systems in
solving problems. In such mixed-initiative systems, machines and people
take turns at making contributions to solving a
problem.\19\,\20\ Advances in this realm can lead to methods
that support humans and machines working together in a seamless, fluid
manner.
---------------------------------------------------------------------------
\19\ E. Horvitz. Principles of Mixed-Initiative User Interfaces.
Proceedings of CHI '99, ACM SIGCHI Conference on Human Factors in
Computing Systems, Pittsburgh, PA, May 1999.
\20\ E. Horvitz. Reflections on Challenges and Promises of Mixed-
Initiative Interaction, AAAI Magazine 28, Special Issue on Mixed-
Initiative Assistants (2007).
---------------------------------------------------------------------------
Recent results have demonstrated that AI systems can learn about
and extend peoples' abilities.\21\ Research includes studies and
methods that endow systems with an understanding about such important
subtleties as the cost of an AI system interrupting people in different
contexts with potentially valuable information or other contribution
\22\ and on predicting information that people will forget something
that they need to remember in the context at hand.\23\
---------------------------------------------------------------------------
\21\ E. Kamar, S. Hacker, E. Horvitz. Combining Human and Machine
Intelligence in Large-scale Crowdsourcing, AAMAS 2012, Valencia, Spain,
June 2012.
\22\ E. Horvitz and J. Apacible. Learning and Reasoning about
Interruption. Proceedings of the Fifth ACM International Conference on
Multimodal Interfaces, November 2003, Vancouver, BC, Canada.
\23\ E. Kamar and E. Horvitz, Jogger: Investigation of Principles
of Context-Sensitive Reminding, Proceedings of International Conference
on Autonomous Agents and Multiagent Systems (AAMAS 2011), Tapei, May
2011.
---------------------------------------------------------------------------
Causal discovery. Much of machine learning has focused on learning
associations rather than causality. Causal knowledge is a critical
aspect of scientific discovery and engineering. A longstanding
challenge in the AI sub-discipline of machine learning has been
identifying causality in an automated manner. There has been progress
in this realm over the last twenty years. However, there is much to be
done on developing tools to help scientists find rich causal models
from large-scale sets of data.\24\
---------------------------------------------------------------------------
\24\ See efforts at the NIH BD2K Center for Causal Discovery:
http://www.ccd.pitt.edu/about/
---------------------------------------------------------------------------
Unsupervised learning. Most machine learning is referred to as
supervised learning. With supervised learning, data is directly or
indirectly tagged by people who provide a learning system with specific
labels, such as the goals or intentions of people, or health outcomes.
There is deep interest and opportunity ahead with developing
unsupervised learning methods that can learn without human-authored
labels. We are all familiar with the apparent power that toddlers have
with learning about the world without obvious detailed tagging or
labeling. There is hope that we may one day better understand these
kinds of abilities with the goal of harnessing them in our computing
systems to learn more efficiently and with less reliance on people.
Learning physical actions in the open world. Research efforts have
been underway on the challenges of enabling systems to do active
exploration in simulated and real worlds that are aimed at endowing the
systems with the ability to make predictions and to perform physical
actions successfully. Such work typically involves the creation of
training methodologies that enable a system to explore on its own, to
perform multiple trials at tasks, and to learn from these experiences.
Some of this work leverages methods in AI called reinforcement
learning, where learning occurs via sets of experiences about the best
actions or sequences of actions to take in different settings. Efforts
to date include automatically training systems to recognize objects and
to learn the best ways to grasp objects.\25\
---------------------------------------------------------------------------
\25\ J. Oberlin, S. Tellex. Learning to Pick Up Objects Through
Active Exploration, IEEE, August 2015.
---------------------------------------------------------------------------
Integrative intelligence. Many R&D efforts have focused on
developing specific competencies in intelligence, such as systems
capable of recognizing objects in images, understanding natural
language, recognizing speech, and providing decision support in
specific healthcare areas to assist pathologists with challenges in
histopathology. There is a great opportunity to weave together multiple
competencies such as vision and natural language to create new
capabilities. For example, natural language and vision have been
brought together in systems that can perform automated image
captioning.\26\,\27\ Other examples of integrative
intelligence involve bringing together speech recognition, natural
language understanding, vision, and sets of predictive models to
support such challenges as constructing a supportive automated
administrative assistant.\28\ There is much opportunity ahead in
efforts in integrative intelligence that seek to weave together
multiple AI competencies into greater wholes that can perform rich
tasks.
---------------------------------------------------------------------------
\26\ H. Fang, S. Gupta, F. Iandola, R. Srivastava, L. Deng, P.
Dollar, J. Gao, X. He, M. Mitchell, J. Platt, C. Zitnick, G. Zweig,
From Captions to Visual Concepts and Back, CVPR 2015.
\27\ Vinyals, O., Toshev, A., Bengio S. Dumitru, E., Show and Tell:
A Neural Image Caption Generator, CVPR 2015.
\28\ D. Bohus and E. Horvitz. Dialog in the Open World: Platform
and Applications, ICMI-MLMI 2009: International Conference on
Multimodal Interaction, Cambridge, MA. November, 2009.
---------------------------------------------------------------------------
Advances in platform and systems. Specific needs for advances with
data-center scale systems and innovative hardware have come to the fore
to support the training and execution of large-scale neural network
models. New research at the intersection of learning and reasoning
algorithms, computing hardware, and systems software will likely be
beneficial in supporting AI innovations. Such research is being fielded
in platforms that are becoming available from large companies in the
technology sector.
Development tools and ``democratization of AI''. New types of
development tools and platforms can greatly assist with development,
debugging, and fielding of AI applications. R&D is ongoing at large IT
companies on providing developers with cloud-based programmatic
interfaces (e.g., Microsoft's Cognitive Services) and client-based
components for performing valuable inference tasks (e.g., detect
emotion in images). Also, learning toolkits are being developed that
enable researchers and engineers to do machine learning investigations
and to field classifiers (e.g., Microsoft's CNTK and Google's
TensorFlow). Other development environments are being developed for
creating integrative AI solutions that can be used by engineers to
assemble systems that rely on the integration of multiple competencies
(natural language understanding, speech recognition, vision, reasoning
about intentions of people, etc.) that must work together in a tightly
coordinated manner in real-time applications.
Challenges
Economics and jobs. Over the last several years, the AI
competencies with seeing, hearing, and understanding language have
grown significantly. These growing abilities will lead to the fielding
of more sophisticated applications that can address tasks that people
have traditionally performed. Thus, AI systems will likely have
significant influences on jobs and the economy. Few dispute the
assertion that AI advances will increase production efficiencies and
create new wealth. McKinsey & Company has estimated that advanced
digital capabilities could add 2.2 trillion U.S. dollars to the U.S.
GDP by 2025. There are rising questions about how the fruits of AI
productivity will distributed and on the influence of AI on jobs.
Increases in the competencies of AI systems in both the cognitive and
physical realms will have influences on the distribution, availability,
attraction, and salaries associated with different jobs. We need to
focus attention on reflection, planning, and monitoring to address the
potential disruptive influences of AI on jobs in the U.S.--and to work
to understand the broad implications of new forms of automation
provided by AI for domestic and international economics. Important
directions for study include seeking an understanding of the needs and
value of education and the geographic distribution of rising and
falling job opportunities.
There is an urgent need for training and re-training of the U.S.
workforce so as to be ready for expected shifts in workforce needs and
in the shifts in distributions of jobs that are fulfilling and
rewarding to workers. In an economy increasingly driven by advances in
digital technology, increasing numbers of jobs are requiring a degree
in one of the STEM (science, technology, engineering, and math) fields.
There is growing demand for people with training in computer science,
with estimates suggesting that by 2024, the number of computer and
information analyst jobs will increase by almost 20 percent. For
companies to thrive in the digital, cloud-driven economy, the skills of
employees must keep pace with advances in technology. It has been
estimated as many as 2 million jobs could go unfilled in the U.S.
manufacturing sector during the next decade because of a shortage of
people with the right technical skills.\29\ Investing in education can
help to prepare and adapt our workforce to what we expect will be a
continuing shift in the distribution of jobs, and for the changing
demands on human labor.
---------------------------------------------------------------------------
\29\ The Manufacturing Institute and Deloitte, ``The skills gap in
U.S. manufacturing: 2015 and beyond.'', 2015.
---------------------------------------------------------------------------
Beyond ensuring that people are trained to take on fulfilling,
well-paid positions, providing STEM education and training to larger
number of citizens will be critical for U.S. competitiveness. We are
already facing deficits in our workforce: The Bureau of Labor
Statistics estimates that there are currently over 5 million unfilled
positions in the U.S. Many of those jobs are those created due to new
technologies. This suggests that there are tremendous opportunities for
people with the right skills to help U.S. companies to create products
and services that can, in turn, drive additional job creation and
create further economic growth.
Shortages of people who have training in sets of skills that are
becoming increasingly relevant and important could pose serious
competitive issues for companies and such shortages threaten the long-
term economic health of the U.S. Without addressing the gap in skills,
we'll likely see a widening of the income gap between those who have
the skills to succeed in the 21st century and those who do not. Failing
to address this gap will leave many people facing an uncertain future--
particularly women, young people, and those in rural and underserved
communities. Working to close this divide will be an important step to
addressing income inequality and is one of the most important actions
we can take.
Safety and robustness in the open world. Efforts to employ AI
systems in high-stakes, safety critical applications will become more
common with the rising competency of AI technologies.\30\ Such
applications include automated and semi-automated cars and trucks,
surgical assistants, automation of commercial air transport, and
military operations and weapon systems, including uses in defensive and
offensive systems. Work is underway on ensuring that systems in safety
critical areas perform robustly and in accordance with human
preferences. Efforts on safety and robustness will require careful,
methodical studies that address the multiple ways that learning and
reasoning systems may perform costly, unintended actions.\31\ Costly
outcomes can result from erroneous behaviors stemming from attacks on
one or more components of AI systems by malevolent actors. Other
concerns involve problems associated with actions that are not
considered by the system. Fears have also been expressed that smart
systems might be able to make modifications and to shift their own
operating parameters and machinery. These classes of concern frame
directions for R&D.
---------------------------------------------------------------------------
\30\ T.G. Dietterich and E.J. Horvitz, Rise of Concerns about AI:
Reflections and Directions. Communications of the ACM, Vol. 58 No. 10,
pages 38-40, October 2015.
\31\ Amodei, D., Olah, C., Steinhardt, J., Christiano, P.,
Schulman, J., Mane, D., Concrete Problems in AI Safety, arXiv, 25 July
2016.
---------------------------------------------------------------------------
Efforts and directions on safety and robustness include the use of
techniques in computer science referred to as verification that prove
constraints on behaviors, based on offline analyses or on real-time
monitoring. Other methods leverage and extend results developed in the
realm of adaptive control, on robust monitoring and control of complex
systems. Control-theoretic methods can be extended with models of
sensor error and with machine learning about the environment to provide
guarantees of safe operation, given that assumptions and learnings
about the world hold.\32\ Such methods can provide assurance of safe
operation at a specified tolerated probability of failure. There are
also opportunities for enhancing the robustness of AI systems by
leveraging principles of failsafe design developed in other areas of
engineering.\33\ Research is also underway on methods for building
systems that are robust to incompleteness in their models, and that can
respond appropriately to unknown unknowns faced in the open world.\34\
Beyond research, best practices may be needed on effective testing,
structuring of trials, and reporting when fielding new technologies in
the open world.
---------------------------------------------------------------------------
\32\ D. Sadigh, A. Kapoor, Safe Control Under Uncertainty with
Probabiliistic Signal Temporal Logic, Robotics: Science and Systems,
RSS 2016.
\33\ Overview presentation on safety and control of AI can be found
here: E. Horvitz, Reflections on Safety and Artificial Intelligence,
White House OSTP-CMU Meeting on Safety and Control of AI, June 2016.
(view video presentation).
\34\ One challenge that must be considered when fielding
applications for safety critical tasks is with transfer of applications
from the closed world of test scenarios to the open world of fielded
technologies. Systems developed in the laboratory or in test facilities
can be surprised by unexpected situations in the open world--a world
that contains unmodeled situations, including sets of unknown unknowns
stemming from incompleteness in a system's perceptions and
understandings. Addressing incompleteness and unknown unknowns is an
interesting AI research challenge.
---------------------------------------------------------------------------
A related, important area for R&D on safety critical AI
applications centers on the unique challenges that can arise in systems
that are jointly controlled by people and machines. Opportunities
include developing systems that explicitly consider human attention and
intentions, that provide people with explanations of machine inferences
and actions, and that work to ensure that people comprehend the state
of problem solving--especially as control is passed between machines
and human decision making. There is an opportunity to develop and share
best practices on how systems signal and communicate with humans in
settings of shared responsibility.
Ethics of autonomous decisions. Systems that make autonomous
decisions in the world may have to make trades and deliberate about the
costs and benefits of rich, multidimensional outcomes--under
uncertainty. For example, it is feasible that an automated driving
system may have to reason about actions that differentially influence
the likelihood that passengers versus pedestrians are injured. As
systems become more competent and are granted greater autonomy in
different areas, it is important that the values that guide their
decisions are aligned with the values of people and with greater
society. Research is underway on the representation, learning,
transparency, and specification of values and tradeoffs in autonomous
and semi-autonomous systems.
Fairness, bias, transparency. There is a growing community of
researchers with interest in identifying and addressing potential
problems with fairness and bias in AI systems.\35\ Datasets and the
classifications or predictions made by systems constructed from the
data can be biased. Implicit biases in data and in systems can arise
because of unmodeled or poorly understood limitations or constraints on
the process of collection of data, the shifting of the validity of data
as it ages, and using systems for inferences and decisions for
populations or situations that differ greatly from the populations and
situations that provided the training data As an example, predictive
models have been used to assist with decision making in the realm of
criminal justice. Models trained on datasets have been used to assist
judges with decisions about bail and about the release of people
charged with crimes in advance of their court dates. Such decisions can
enhance the lives of people and reduce costs. However, great caution
must be used with ensuring that datasets do not encode and amplify
potential systematic biases in the way the data is defined and
collected. Research on fairness, biases, and accountability and the
performance of machine-learned models for different constituencies is
critically important. The importance of this area will only grow in
importance as AI methods are used with increasing frequency to advise
decision makers about the best actions in high-stakes settings. Such
work may lead to best practices on the collection, usage, and the
sharing of datasets for testing, inspection, and experimentation.
Transparency and openness may be especially important in applications
in governance.
---------------------------------------------------------------------------
\35\ See Fairness, Accountability, and Transparency in Machine
Learning (FATML) conference site: http://www.fatml.org/
---------------------------------------------------------------------------
Manipulation. It is feasible that methods employing machine
learning, planning, and optimization could be used to create systems
that work to influence peoples' beliefs and behavior. Further, such
systems could be designed to operate in manner that is undetectable by
those being influenced. More work needs to be done to study, detect,
and monitor such activity.
Privacy. With the rise of the centrality of data-centric analyses
and predictive models come concerns about privacy. We need to consider
the potential invasion in the privacy of individuals based on
inferences that can be made from seemingly innocuous data. Other
efforts on privacy include methods that allow data to be used for
machine learning and reasoning yet maintains the privacy of
individuals. Approaches include methods for anonymizing data via
injecting noise \36\, sharing only certain kinds of summarizing
statistics, providing people with controls that enable them to trade
off the sharing of data for enhanced personalization of services \37\,
and using different forms of encryption. There is much work to be done
on providing controls and awareness to people about the data being
shared and how it is being used to enhance services for themselves and
for larger communities.
---------------------------------------------------------------------------
\36\ Differential privacy ref: Dwork, C.: Differential Privacy. In:
Proceedings of the 33rd International Colloquium on Automata, Languages
and Programming (ICALP) (2), pp. 1-12 (2006).
\37\ A. Krause and E. Horvitz. A Utility-theoretic Approach to
Privacy in Online Services, Journal of Artificial Intelligence
Research, 39 (2010) 633-662.
---------------------------------------------------------------------------
Cybersecurity. New kinds of automation can present new kinds of
``attack surfaces'' that provide opportunities for manipulation and
disruption by cyberattacks by state and non-state actors. As mentioned
above, it is critical to do extensive analyses of the new attack
surfaces and the associated vulnerabilities that come with new
applications of AI. New classes of attack are also feasible, including
``machine learning attacks,'' involving the injection of erroneous or
biased training data into datasets. Important directions include
hardware and software-based security and encryption, new forms of
health monitoring, and reliance on principles of failsafe design.
Recommendations
We recommend the following to catalyze innovation among our basic
and applied AI communities across government, academia, industry, and
non-profit sectors:
Public-sector research investments are vital for catalyzing
innovation on AI principles, applications, and tools. Such
funding can leverage opportunities for collaborating and
coordinating with industry and other sectors to help facilitate
innovation.
Research investments are needed at the intersection of AI,
law, policy, psychology, and ethics to better understand and
monitor the social and societal consequences of AI.
Governments should create frameworks that enable citizens
and researchers to have easy access to government curated
datasets where appropriate, taking into consideration privacy
and security concerns.
With the goal of developing guidelines and best practices,
governments, industry, and civil society should work together
to weigh the range of ethical questions and issues that AI
applications raise in different sectors. As experience with AI
broadens, it may make sense to establish more formal industry
standards that reflect consensus about ethical issues but that
do not impede innovation and progress with AI and its
application in support of people and society.
In an era of increasing data collection and use, privacy
protection is more important than ever. To foster advances in
AI that benefit society, policy frameworks must protect privacy
without limiting innovation. For example, governments should
encourage the exploration and development of techniques that
enable analysis of large datasets without revealing individual
identities.
We need to invest in training that prepares people for high-
demand STEM jobs. Governments should also invest in high-
quality worker retraining programs for basic skills and for
certifications and ongoing education for those already in the
workforce. A first step is to identify the skills that are most
in demand. Governments can develop and deliver high-quality
workforce retraining programs or provide incentives and
financial resources for private and nonprofit organizations to
do so.
Thank you for the opportunity to testify. I look forward to
answering your questions.
Senator Cruz. Thank you, Dr. Horvitz.
Dr. Moore.
STATEMENT OF DR. ANDREW W. MOORE, DEAN, SCHOOL OF COMPUTER
SCIENCE, CARNEGIE MELLON UNIVERSITY
Dr. Moore. Thank you, Chairman Cruz, Ranking Member Peters,
and members of the Subcommittee, for convening this really
important meeting.
What I want to do with this testimony is to offer three
takeaways. First, what AI is and what it isn't right now. The
``artificial'' in artificial intelligence is there for a
reason. Second, I will explain why things are changing so
quickly right now. And third, I want to talk about the thing
that keeps me awake at night regarding U.S. competitiveness in
AI.
So, what is AI? When one of my students decides to build an
AI, they always end up doing two things. One, they make sure
the computer can perceive and understand the world through
computer vision and through big data. Second, they do massive
search, which means, in all these examples we've talked about,
such as a car finding a route, the computer searches to find
the best of billions, and sometimes quintillions, of
possibilities, and does that very quickly and efficiently. And
that's happening all the time. When you ask Siri a question,
what's happening up there in the cloud is, it is essentially
trying out about 10 billion possible answers to your question
and searching over them to score each one as to how likely it
is to make you satisfied. If an autonomous car is about to hit
a deer, and it's about .2 of a second to impact, it can spend
the first 20th of a second gaming out millions of possibilities
of what it can do, what the deer is going to do, to maximize
the chance that people are going to survive. So, that's what's
going on in AI. It is perception through big data and search.
What has really changed in the last couple of years is,
through the efforts of the United States in big data, we now
have computers which can perceive enough of this to be useful.
And this means that there's now a big land grab going on for
researchers and entrepreneurs and students for finding all the
places where we can use this. For example, this picture here is
of Sebastien LePage, who is V.P. of Operations at Kinova, a
robotics arm company, working with Carnegie Mellon faculty on
the question of, if you are unable to use anything below your
neck, but you want to indicate by nodding at something that
you'd like to pick it up, perhaps nod again to say you want to
bring it to your mouth, searching over the billions of possible
things the arm can do safely to find the one which is most
likely to make you, as the user, happy. This is an
unambiguously good use of technology applicable to tens of
thousands of our veterans, for example.
But, what excites me, and the reason I'm in this business
right now, is that there are thousands of stories like this
happening, and my students and faculty and folks all across the
country who have the skills are exploring the ways of doing
this.
For example, one of our undergraduate students, working by
herself using open AI tools, managed to quickly come up with a
system for checking open Internet records to detect sex
traffickers. And her algorithm has saved, you know, a couple of
hundred young women from sex trafficking. That's one person who
had the skills able to do this.
Another example, this wasn't possible 12 months ago. Now we
have drones zooming quickly through forests, able to dodge
trees as they're going, because they can now afford to plan so
fast to avoid trees in sub-second time. Many other examples, so
many that I could keep us going for half an hour. They all just
make me so excited.
But, the thing that keeps me awake at night on all of this
is the talent war. I really, really beseech that, together, we
can get a million of the current middle-school students in the
country to become AI experts over the next 5 to 10 years. At
the moment, we have a tiny fraction--I would say less than 1
percent--of the people available to work in this area who could
work in this area. If you duplicated these four panel members a
hundred times each, we still would have too much to do when it
comes to taking advantage of all these opportunities.
I estimate that, when an Internet company hires one of our
students, they're making 5 to 10 million dollars per student
just by having that person on their payroll. And so, the
bidding wars for these talents are huge. And our students,
instead of necessarily moving to work on AI for the Veterans
Administration or AI for helping protect our warfighter, the
majority of them are simply going off to Internet companies,
which is fine, but I want them in all the other sectors of our
economy, as well.
Similarly, if you look at every one of the big advances in
artificial intelligence that are now active, they came from
university professors--the majority of them from U.S.
university professors. We are beginning to lose many of our
university professors to industry, and that is damaging our
seed corn. Our university professors who are AI experts are
looking for sustained, stable funding, not necessarily lots of
funding, so that they can realize their dreams of doing things
like this.
So, I'm very excited. I'm very grateful for this
subcommittee for shining a light on this important issue. I
think the future is bright, but it really is an AI race at the
moment.
[The prepared statement of Dr. Moore follows:]
Prepared Statement of Dr. Andrew W. Moore, Dean, School of Computer
Science, Carnegie Mellon University
Thank you Chairman Cruz, Ranking Member Peters, and Members of the
Subcommittee for convening this important hearing on Artificial
Intelligence (AI). I am honored to be here and to be joined by
colleagues who are advancing the science, technology, business models,
critical applications, and policy considerations of AI in the service
of the United States and humanity.
My name is Andrew Moore. I am the Dean of the School of Computer
Science at Carnegie Mellon University and former Vice President at
Google responsible for Machine Learning technology. I appreciate your
leadership in focusing on the future of science, innovation, and
American competitiveness and on the role that AI can play. The policies
and strategies we adopt over the next several years will determine if
the United States wins the race to lead this technological revolution
as well as the resulting benefits for our citizens.
Introduction: Perspectives on the Future of AI from a Journey in
Computer Science
Building upon fifty years of research, strategic Federal
investments, dramatic advances in machine learning, and the explosion
in available digital data, we no longer describe AI as a technology
from the future: it is around us in our phones, our vehicles and in
defense of our borders. AI tools are already making doctors better at
diagnosing diseases and ensuring patients obtain the latest effective
treatments.
AI-empowered personalized learning will enable teachers to better
reach and engage every student. Powerful new AI cyber tools will
provide a new and more definitive defense against a world increasingly
populated by hackers intent on criminal or state-sponsored attacks on
American institutions, businesses and citizens. Adaptive, learning
robotic systems will enable small manufacturers to cost-effectively
change product lines more rapidly--even realizing mass production
economies from ``quantity one'' to compete with foreign firms utilizing
cheap labor. The ability to combine autonomous vehicles with public
transit will unlock urban congestion, transform land use, enhance
safety, and enable cities to focus on the most critical human elements
of mobility. And, the potential applications of AI as powerful tools in
national defense and homeland security will make us safer, even in the
face of growing threats. In each of these areas, powerful opportunities
exist to eradicate the barriers of distance, economic isolation, and
limited economic opportunities, as well as making us a smarter, more
productive, healthier, safer nation.
Some economists assert that increased deployment of AI could
represent a powerful economic stimulus for the nation--perhaps adding
as much as 2 points to annual GDP growth by 2035.\1\ There are also
economists who warn that the advance of AI applications could
exacerbate income inequality and threaten a wide number of middle
income jobs.\2\
I am not an economist by training. I bring to this hearing
perspectives shaped by my journey over three decades as a computer
scientist and a technology business leader. As a university researcher
I had the opportunity to develop machine learning capabilities that
enable emergency room physicians to better predict the illnesses and
patient levels they are likely to confront as weather and virus
outbreak patterns evolve. This experience provided a window on how
powerful AI applications can be to improve the delivery of vital
services to those in need.
At Google I helped develop advanced machine learning platforms to
more effectively connect consumers to information by making search
engine algorithms smarter and more powerful. That experience also
taught me how AI tools can democratize access to information and
unleash the energy of entrepreneurs to capitalize on the power of these
platforms to bring products to consumers in a way that would have never
been possible before.
For example, enabling consumers to see the 200,000 new dresses that
are produced each day in the world helps to unleash the creativity and
entrepreneurship of dress makers and fashion designers in an
unprecedented way, whether they are large companies or a small startup,
in a major city or a rural community.
But, as this Committee knows well, we face far broader and more
daunting and important challenges as a nation than matching consumers
with dresses.
Now, as Dean of the #1 Computer Science School in the U.S., I have
the wonderful opportunity to engage with a new generation of students--
and their faculty mentors--who are drawn to computer science because
they want to focus their careers on applying AI to tackle our biggest
societal challenges. They arrive at this with the clear eyed
recognition that, as has been true with all new innovation, they must
also address the potential negative impacts these technologies may
bring. These experiences make me very optimistic that we can harness
the power of AI to grow our economy and improve our quality of life
while also acting definitively to mitigate any potential disruptions
this new technology, like any new technology, can bring. New technology
will always come. We must contribute to its use for good.
My journey as a computer scientist leaves me certain that AI can
create fundamentally new economic opportunities and be a powerful
resource for addressing our most pressing challenges in areas of
security, health care, better food production, and a new era of growth
in manufacturing. At the same time, it can fundamentally transform the
nature of work, as well as create new challenges in areas such as
privacy. The key is a focused national strategy to nurture and attract
the best talent, including applying new AI learning tools to aid
workers in need of retraining; to enhance discovery and
commercialization; and to create a business and regulatory environment
that rewards innovation.
Carnegie Mellon and the AI Revolution
My perspective has been heavily shaped by the culture of discovery
at the School of Computer Science at Carnegie Mellon. The development
of Artificial Intelligence was launched 60 years ago at a seminal
gathering at Dartmouth University in the summer of 1956. Two of the
four scientists who led that session, Allen Newell and Herbert Simon,
were CMU faculty and had already created the first AI program. Since
that time, with strong support from Federal research agencies, our
faculty have pursued disruptive innovations that have help fuel the
development of AI. These innovations include multithreaded computing,
speech and natural language understanding, computer vision, software
engineering methodology, self-driving robotic platforms, distributed
file systems and more.
Today well over 100 faculty and 1,000 students at Carnegie Mellon
are engaged in AI-related research and education. In addition to
advancing breakthroughs fundamental to the building blocks of AI
systems, Carnegie Mellon faculty and student researchers have applied
advances in AI to the early detection of disease outbreaks, combating
sex trafficking rings, detection of emerging terror threats in social
media, and to the development of cognitive tutoring tools that are now
deployed in middle schools, high schools, and colleges in every state
in the Nation. CMU alumni and faculty (typically on leave) hold leading
positions in each of the major companies driving AI development,
including at Microsoft, IBM, Google, Amazon, and Apple. CMU spin-off
companies have been a catalyst to advancing AI innovations.
Fundamental Building Blocks of AI Systems
AI is defined as ``the scientific understanding of the mechanisms
underlying thought and intelligent behavior and their embodiment in
machines.'' \3\ As we strategize on the next AI steps at Carnegie
Mellon University, it helps us to break AI research into two broad
categories: Autonomous AIs and Cognitive Assistant AIs. An Autonomous
System has to make low level decisions by itself, for example a car
that only has half a second to react to a collision simply cannot wait
for a human. Or a constellation of satellites that has lost
communications with the ground needs to figure out what they should be
observing and transmitting to the ground while trading off the need to
protect their advanced sensors against an energy attack. Cognitive
Assistants, on the other hand, work hand in hand with a human: our
smart phones telling us how to get to our kid's dental appointments are
a simple example. Much more advanced examples include CMU faculty Sidd
Srinivasa's work on intelligent robot arms controlled by humans in
wheelchairs with high spinal cord injuries.
AI involves transforming raw data--often massive amounts of raw
data--into usable, actionable information. This cycle is known as
``data to knowledge to action.'' The graphic below captures the
``stack'' of elements that constitute AI. It is intended to show all of
the areas that are important for ongoing AI research and development,
to continue to expand our science and technology.
The foundation is the device and hardware layer that includes
powerful computer processing and storage capabilities. The data science
kernel layer includes architectures for processing massive amounts of
data--essential to managing the explosion of digital data available
through the Internet and the growing global network of sensors. The
Machine Learning (ML) layer includes algorithms that automate the
detection of patterns and gather insights from large data sets far
faster than humans could, even in many lifetimes. The modeling layer
includes statistical methods and tools for prediction--the ability to
move from the recognition of patterns in data to the ability to
understand how complex real-world systems and structures behave. We
mean ``systems'' in a general sense: from biological entities, to
behaviors, to farms, to cities, to societies, to the cosmos. One
example system is triage of inspection of cargo by U.S. Customs.
Another is detecting and managing the response to potential false
alarms by emergency responders. The decision support layer includes
management information systems software that assembles facts, diagnoses
status and evaluates potential actions. As an example, decision support
applications are vital to enable autonomous vehicles to rapidly react
to changing traffic patterns. They are also in use in flexible
manufacturing systems in American factories. Decision support
capabilities also include tools to detect human emotion and intent and
create profiles from the physics of speech. Each of these layers builds
on the layers below it.
These building block layers power the two major application areas
of AI--autonomous systems and capabilities to augment human
performance. One application developed by a team of Carnegie Mellon
University School of Computer Science researchers, led by Rita Singh,
illustrates how the components of the AI ``stack'' can be applied to
dramatically enhance intelligence analysis and crime-solving
capabilities of organizations that deal with voice-based crimes.
The world is increasingly communicating through voice: an estimated
700 centuries worth of speech is transmitted over cellphones alone each
day. While more people are talking than ever before, even more people
are listening. There are 4 billion views of YouTube videos daily. These
and other Internet-accessible videos have voice embedded in them. The
tremendous outreach of voice today allows for a dangerous world where
more and more crimes can be committed and propagated through voice
alone. These crimes include those that affect people's personal
security, such as harassment, threats, extortion through fraudulent
phone calls etc., all the way to societal crimes that affect national
security, like hoax calls, criminal propaganda, communication in
organized crime, terrorist indoctrination etc.
The CMU team is developing technologies that utilize the power of
machine learning and AI to profile people through their voices. They
are able to describe the physical appearance of a person, background
and demographic facts about the person, and also the person's
surroundings entirely from their voice. In recent work with the U.S.
Coast Guard Investigative Services, the team analyzed scores of Mayday
calls from hoax callers transmitted over national distress channels,
and has provided physical descriptions of the perpetrators, and of
their location and the equipment used that were sufficiently accurate
to enable significant success in the investigative process.
It is noteworthy that the U.S. law enforcement and security
agencies as well as first responders are faced with hoax calls on a
daily basis, and these collectively cost the Nation billions of dollars
in misdirected and misused resources each year. Hoax calls are just one
example. The ability to track and describe humans through their voice
is useful in several disciplines of national intelligence, where voice
is part of the intelligence information gathered.
Our work builds on the fact that humans can make judgments about
people from their voices, like their gender, emotional state, their
state of health, and many others. The CMU team utilizes powerful AI
techniques to achieve super-human capabilities that enable machines to
make faster, more accurate, more abundant and deeper assessments of
people from their voices. This is made possible by advances in AI,
computing, machine learning and other related areas, and over two
decades of developments in automatic speech and audio processing
capabilities at CMU. The team hopes to be able to build physically
accurate holograms of humans from their voices in the future.
This work, and that of many others, demonstrates the power of AI to
dramatically help with judgments that humans make and in doing so
augment human capabilities. This case is also illustrative of what we
at Carnegie Mellon believe will be a dominant pattern of AI deployment:
work in close synergy with humans. The nature of work tasks will
evolve, potentially dramatically in certain cases, and will demand new
and different skills. AI systems that augment and complement human
capabilities will help us as individuals and as a nation through this
transition and beyond.
Similar examples of AI already touch our daily lives. Smartphone
applications that personalize services are based upon AI algorithms.
Other AI applications are helping forecast crop yields, analyzing
medical samples, and helping deploy police and fire resources.
Autonomous systems are at work on city streets, on American farms, and
patrolling the sea and air for our national defense.
Intelligent AI systems will also include mobile robots and
intelligent processing and decisionmaking among the sensory and
actuation capabilities of the ``Internet of things.'' AI systems may
always have limitations and will therefore be in a symbiotic/
coexistence relationship with humans, and with other AI systems.
Designing and building these systems and relationships is a fruitful
area for advances.
Perhaps most critically, judgments that humans make in the area of
national intelligence are vital to our safety and security. Combined
with the wealth of data available today (including through
crowdsourcing), AI is the future power source of these decisions--
processing far more possibilities and scenarios than humans could
alone, and working closely with humans to keep us protected.
And, we are just at the start of this AI revolution.
The Inflection Point and Emerging AI applications and Capabilities
Two specific breakthroughs in the last five years have created the
inflection point that makes this hearing so timely and essential. The
first is the rapid advancement in digital datasets that are central to
AI applications. Current estimates of the world's digital data are
approaching 1.3 zettabytes or about 1.3 trillion gigabytes.\4\ Fueled
by both research and applications, as well as a strong commitment to
increasing access to government data, this explosion includes digital
biomedical data, mapping data, traffic data, astronomical data, data
from sensors monitoring machines and buildings, and data from social
media capturing consumer trends from restaurants to travel patterns.
Advanced AI applications are catalyzed by the availability of this
data.
The second major breakthrough is the development of deep learning
techniques in machine learning. Deep learning involves a statistical
methodology for solving problems in very large and very complex
datasets. The term ``deep'' is derived from the ability of these
learning methodologies to automatically generate new models and
abstractions of the data. Deep learning brings about the potential for
self-learning capabilities that are the central to dramatic advances in
AI applications. More critically, deep learning creates the potential
for advancing beyond narrow AI--applications focused on one specific
task--to general AI that creates a platform for undertaking a wide
range of complex tasks and responding in complex environments.
Thoughts on the Policy Implications of the Emerging AI Revolution
The potential transformative impact of these future applications of
AI to transform our economy, generate economic opportunity and address
critical challenges to our security and quality of life is clear.
However, the future--especially the future of U.S. leadership in this
area--is not assured. Drawing upon my experiences as a researcher in
machine learning, a technology business leader committed to developing
AI capabilities, and now as a computer science dean engaging with the
aspirations of faculty and students, here are selected thoughts on some
of the key elements of a strategy to ensure continued U.S. leadership.
Winning the Talent War
We need a comprehensive set of policies and incentives that
addresses the skills needed to win in the AI-driven economy of the 21st
Century. These policies must address the talent pipeline, from computer
scientists per se to the workers impacted by new applications.
The starting point is a recognition that we are already engaged in
an international war for talent. Based upon my experience in a leading
technology company, a computer science graduate with expert level AI
training adds between $5 million and $10 million to the bottom line of
a company.
These people are very rare for two reasons. First, they need to
have the natural abilities to deal with logic and math and software on
a massive scale. Second, they need to survive very intense training
that covers many disciplines at once, including algorithms, robotics,
security, ethics, advanced probability and human-centered design.
As a result of the rarity of these skills, young AI experts are
being heavily competed for around the globe. We see crazy bidding wars
taking place from Beijing to Boston to Pittsburgh to Paris. The United
States is not winning in the rate of production of these young experts,
and we have recommendations below on how to get back on track.
Secondly, AI is one area where international innovation is coming
primarily from universities. It is North American professors and their
graduate students who have introduced all of the following great
advances in AI in recent years: self driving, deep learning, advanced
human recognition, emotion detection, provable AI safety, spoken dialog
systems, autonomous helicopters, intelligent traffic control, and many
others. These have all been taken into the corporate and military
worlds through technology transition and through many professors and
students transitioning with their technology. The success of AI
professors has had great benefit for the economy and security, but it
is getting harder and harder to entice new AI geniuses to replenish the
ranks of North American professors. The concerns about their retention
are twofold: it is increasingly lucrative to abandon an academic
position and also increasingly hard to raise funding for university
research. These professors are very important because they are the ones
producing thousands of AI experts for the country every year. If the
U.S. loses many of these professors--and fails to continue the pipeline
from graduate school--the supply if U.S. AI experts will dry up.
We will need a balanced set of policies and incentives to ensure
that we can provide the talent companies need while securing our long
term capacity for research and innovation. This requires recognizing
the imperatives of retaining top faculty and supporting graduate
students. To support faculty retention we may wish to consider
strategies utilized by some of our international competitors who issue
competitive ``star grants'': multi-year awards to the top 100
researchers to enable them and inspire them to continue their academic
research careers. To maintain our base of graduate students who are
central to our research leadership, consideration should be given to
expanding fellowships focused explicitly on AI-related fields and
expanding the number of multi-year, broad-based research awards that
enable faculty to provide support for students throughout their
graduate studies and within ambitious projects.
We also need to move aggressively to build the pipeline of computer
science talent. The Every Student Succeeds Act, the ESEA
reauthorization passed by this Congress, makes an important start by
emphasizing the importance of computer science in STEM education. It is
also increasingly vital to foster stronger collaborations across the
education spectrum: for example, between research universities and
community colleges and between higher education institutions and K-12
to enhance curricula, teacher education, and student engagement.
As has been vital in all periods of discovery and innovation, it is
essential that the United States retain its ability to attract the best
and brightest talent from around the world to study here, work here,
perform world-class research and development here, and start American
companies, all of which serve as engines for growth and national
prosperity.
For example, Carnegie Mellon is now engaged in a collaboration with
Microsoft's TEALS program and Pittsburgh Public Schools to enhance the
ability of teachers to introduce computational concepts throughout the
curriculum, by drawing on volunteer computer scientists who understand
the importance and urgency of computer science education. Similar
collaborations are taking place across the Nation. We will need to
explore how best to incentivize formal and informal learning
initiatives in all communities.
Winning the talent war will also require fundamentally new
approaches to workforce training. Many workforce programs tend to focus
on shifting individuals to new careers or training workers to operate a
specific type of equipment. Neither model is likely to be completely
applicable to empower workers to thrive as AI applications impact a
wide range of industries.
It will not be necessary for workers to have a computer science
degree to thrive in the AI economy. But the capacity and skills to work
with advanced machines and understand computational processes will be
essential. This will require a mix of technical skills and an
understanding of data analytics. This new workforce environment is
already taking shape. There are construction firms using advertisements
highlighting the opportunity to work alongside robots as a benefit in
their efforts to attract skilled workers. Advanced manufacturing is
another area that will build on the strength of robotics, while
requiring more and more tech-savvy workers.
We have two great resources in creating a skill development
environment for the AI era. First, more than in any other period of
technological development, we have the power of intentionality. We can
advance AI research and innovations with explicit consideration of the
human engagement and models of human/machine interaction in mind. It
will be vital for workers and workforce development professionals to
become integral to the AI research process to realize this opportunity.
Second, the AI revolution itself will give us unprecedented tools
for workers to develop new skills. AI is already creating the capacity
to personalize training for the individual worker, for example by
understanding and modeling each learner's path through a curriculum,
and blend technical and academic content that is targeted to the
specific job. Combined with innovations like wearable computing
devices, entirely new, more powerful approaches to on the job training
are being deployed.
Creating a National Framework for AI Research and Innovation
The amazing AI application that describes individuals solely
through their voice is built on over 20 years of federally funded
research. The next wave of breakthroughs in AI will take place in
academic labs, startups and major companies. We will need a national
research and innovation framework tailored to this ecosystem.
The starting point is Federal research focused on the critical
fundamental gaps impeding AI development. The recent reports prepared
by the White House National Science and Technology Council, with
extensive input from academic and industry researchers, is an excellent
starting point for identifying cross-cutting foundational research
areas.\5\ As noted in the NSTC reports, we will need to develop a
science of safety, dependability, and trust for AI systems. Traditional
verification methodologies and approaches are not fully applicable to
systems that learn and continually improve. This effort will require
both investments in advancing new methodologies and the creation of
test beds.
This focus on the science of safety and trust must also include
engagement on issues of privacy and the ethics of AI deployment.
Through a gift from K&L Gates, Carnegie Mellon University is launching
a new initiative focused on ethics, trust, and privacy. Federal support
that helps engage computer scientists, social scientists, legal and
policy experts, and industry leaders will also be key.
Another critical gap highlighted in the White House reports
involves the imperative for continued research focused on systems for
human-computer interaction. Research advances will ensure the effective
design of AI systems with user friendly interfaces that work seamlessly
alongside humans in a variety of settings. Future AI systems must be
able to adapt to different challenges, such as providing flexible
automation systems that switch from worker to machine operation and
systems designed to address situations where the operator is overloaded
by the complexity of his or her tasks.
Finally, it will also be critical to invest in the foundational
capabilities for scaling AI systems. The most critical need is to
collaborate across industry and government to improve access to the
knowledge that fuels the capabilities of AI systems. One promising
dialogue in this area is well underway. Representatives of agencies,
universities, and industry have worked on the development of a
collaborative AI infrastructure initiative initially called The Open
Knowledge Network (TOkeN). TOkeN would provide a vital core
infrastructure for AI development--interfaces to large data and
knowledge bases that can accelerate the ability of AI systems to create
products and services, broadly speaking, in health care, education,
climate and planetary sciences, energy, manufacturing, and a host of
other areas. TOkeN would be an open webscale, machine-readable
knowledge network aspiring to include every known concept from the
world of science, business, medicine, and human affairs--including both
raw data and semantic information. The creation of TOkeN would enable
the rapid expansion of AI applications for diagnosing disease,
designing new products or production processes, and serving our
citizens in many other ways.
The collaborators intend that TOkeN, if implemented, would
represent the kind of foundational infrastructure that was created to
launch the Internet era. In the early 1980s, proprietary, disconnected
islands of technology prevented the scaling of applications and
services--the Internet connected them. Today, islands of proprietary
and disconnected data and knowledge sets are impeding academic research
and industry innovation. With a relatively limited investment we can
create the foundation for scalable AI development and accelerate
innovation.
In addition to a focused research agenda we will need a research
framework that recognizes the nonlinear nature of AI innovation. Basic
and applied development is taking place in universities, startups, and
companies. We need to incentivize collaboration across this ecosystem.
The Computing Community Consortium (CCC) has advanced thoughts on how
new models of public/private, industry/academic partnerships can be
crafted to meet this challenge.\6\
One powerful tool to stimulate this collaboration is Federal
support for grand challenges that bring together companies, students,
faculty, and often state and local governments to apply innovations to
address particular critical societal objectives and opportunities. The
DARPA grand challenges have helped advance both the development of
autonomous vehicles and automated cyber defense capabilities. AI grand
challenges focused on issues such as education, manufacturing, or
opportunities to expand economic opportunity in rural areas would have
a catalytic impact on both fundamental research and commercial
applications.
Align Research and Development with Smart Regulatory and Procurement
Initiatives
The development and scaling of AI innovations will demand new
regulatory paradigms. Initial positive steps have been undertaken to
help advance the deployment of autonomous vehicles but we must summon
federal, state, and local, as well as industry and citizen
collaboration to craft smart regulations that advance AI and tap its
power to more efficiently realize public policy objectives for health
and safety. Without progress on regulatory issues AI development will
stagnate or, more likely, innovations born in the U.S. will take root
abroad, impeding national competitiveness. Combining regulatory
experiments and test beds with strategic procurement initiatives to
help advance AI products and services will be vital.
We need an ``All In'' Approach
Synergistic engagement among the Federal Government and our
``laboratories of democracy,'' the states, has been a powerful tool for
U.S. science since the efforts to revitalize the competitiveness of the
U.S. semiconductor industry in the 1980s. For example, Federal research
and commercialization investments in the life sciences have catalyzed
billions of dollars of state and local initiatives.\7\ These state and
local efforts help augment research infrastructure, train workers,
expand K-12 curricula, and incubate and nurture startups. Engagement of
the states in AI policy is particularly critical as we seek to advance
STEM education and workforce training initiatives, foster an innovative
regulatory environment, and continually cultivate a vibrant environment
for incubating AI startups.
Conclusion
Thank you once again for convening this hearing and for the
opportunity to join my distinguished colleagues to share thoughts on
the direction and implications of advances in Artificial Intelligence.
My experiences as a researcher, business leader, and dean lead me to
believe that applications of AI will begin to accelerate rapidly across
a host of industries. I believe these applications will expand economic
opportunity and contribute to addressing major societal challenges in
health care, food production, security and defense, energy, and the
environment and education. The ``democratizing'' power of AI
applications to bring new capabilities to individuals on the job, in
schools, and in our homes and communities is at the heart of this
potential.
My experiences have also made me greatly aware that we are in a
global race for talent and innovation. Focused attention on the impact
these applications may make on the nature of work in a host of
industries and the challenges they bring to our privacy is vital. This
will require drawing upon the very best American traditions of
collaboration across government, industry and academia.
It will also require research investments to advance innovation in
key gap areas that are core to advancing AI and sparking innovation,
entrepreneurship and new products and services. We will need an
innovative focus on regulatory environments that will be transformed by
AI. We must nurture our talent resources: from retaining top
researchers, to attracting the best and brightest from across the
globe, to creating a national pipeline to nurture students in every
community and creative new approaches to support existing workers. I
speak with confidence in stating that the university research,
education, and industry communities stand ready to engage in helping to
ensure that the AI revolution expands opportunities to all Americans.
End Notes and References
1. Why Artificial Intelligence is the Future of Growth, Mark Purdy
and Paul Daugherty, Accenture, 2016, P.19.
2. See for example the report on research conducted by Forrester,
``AI will eliminate 6 percent of jobs in five years, says report,''
Harriet Taylor, CNBC, September 12, 2016.
3. See American Association for the Advancement of Artificial
Intelligence, http://www.aaai.org/
4. ``World's Internet traffic to surpass one zettabyte in 2016,''
James Titcomb, The Telegraph, February 4, 2016.
5. The National Artificial Intelligence Research and Development
Strategic Plan, National Science and Technology Council, Networking and
Information Technology Research and Development Subcommittee, October
2016. See pages 16-22, and Preparing for the Future of Artificial
Intelligence, National Science and Technology Council, Committee on
Technology, October, 2016.
6. The Future of Computing Research: Industry-Academic
Collaborations Version 2, Computing Community Consortium, 2016.
7. For example, in 2001 Pennsylvania committed $2 billion in its
tobacco settlement funding allocation to support research by
universities and health research institutions, support venture
investments in the life sciences and fund regional cluster initiatives.
In 2008, Massachusetts committed $1 billion for a 10 year initiative
for capital investments in research infrastructure and start-ups.
Michigan invested $1 billion in 1999 over 20 years to support the
growth of life sciences corridors. For a summary of some of these
initiatives and other state efforts see ``Successful State Initiatives
that Encourage Bioscience Industry Growth,'' Peter Pellerito, George
Goodno, Biotechnology Industry Organization (BIO), 2012.
Senator Cruz. Thank you, Dr. Moore.
Mr. Brockman.
STATEMENT OF GREG BROCKMAN, CO-FOUNDER AND CHIEF TECHNOLOGY
OFFICER, OpenAI
Ranking Member Peters, distinguished members of the
Subcommittee, as well as their staff. This is a really
important session, and I'm honored to be giving this testimony
today.
I'm Greg Brockman, Co-Founder and Chief Technology Officer
of OpenAI. OpenAI is a nonprofit AI research company with a
billion dollars in funding. Our mission is to build safe,
advanced AI technology, and to ensure that its benefits are
distributed to everyone. We're chaired by technology executives
Sam Altman and Elon Musk.
The U.S. has led essentially all technological
breakthroughs of the past 100 years. And they've consistently
created new companies, new jobs, and increased American
competitiveness in the world. AI has the potential to be our
biggest advance yet.
Today, we have a lead, but we don't have a monopoly, when
it comes to AI. This year, Chinese teams won the top categories
in a Stanford annual image recognition context. South Korea
declared a billion-dollar AI fund. Canada actually produced a
lot of the technologies that have kicked off the current boom.
And they recently announced their own renewed investment into
AI.
So, right now I would like to share three key points for
how the U.S. can lead in AI:
The first of these is that we need to compete on
applications. But, when it comes to basic research, that should
be open and collaborative. Today, AI applications are
broadening. They're helping farmers decide which fields to
seed. They're helping doctors identify cancers. But, the
surprising thing is that industry is not just capitalizing on
the advances that have been made to date. Companies like
Facebook, Google, Microsoft, they're all performing the basic
scientific research, the kind of work that you would expect to
see just in academia. And they're trying to create the new AI
building blocks that can then be assembled into products.
And even more surprisingly, these industrial labs, they're
publishing everything that they discover. They are not holding
back any secrets. And the reason they do this is because
publication allows them to pool their resources to make faster
breakthroughs and to attract world-class scientists. Now, these
companies, they stay competitive by publishing the basic
research, but they don't talk about how they put this stuff
together to actually make products, to actually make the things
that are going to make dollars for the company. For example,
IBM Watson, Microsoft Cortana, there aren't many papers on how
those are built. And the thing that's happened is that this
openness has concentrated the world's AI research and
corresponding commercial value all around the United States.
This includes attracting many of the Canadian scientists who
really kicked off this AI boom. They're here now. In fact, one
of them is one of my cofounders at OpenAI. And, importantly,
this has allowed us to define the cultures, the values, and the
standards of the global AI community.
Now, this field is moving so quickly that basic research
advances tend to find their way into products in months, not
years. And so, the government can directly invest in American
innovation and economic value by funding basic AI research.
The second thing that we need to do is that we need public
measurement and contests. There's really a long history of
contests causing major advances in the field. For example, the
DARPA Grand Challenge really led directly to the self-driving
technology that's being commercialized today. But, really
important, as well, measures and contests help distinguish hype
from substance, and they offer better forecasting. And so, good
policy responses and a healthy public debate are really going
to depend on people having clear data about how the technology
is progressing. What can we do? What still remains science
fiction? How fast are things moving? So, we really support
OSTP's recommendation that the government keep a close watch on
AI advancement, and that it work with industry to measure it.
The third thing that we need is that we need industry,
government, and academia to start coordinating on safety,
security, and ethics. The Internet was really built with
security as an afterthought. And we're still paying the cost
for that today.
With AI, we should consider safety, security, and ethics as
early as possible--and that means today--and start baking these
into the technologies--into the fundamental building blocks
that are being created today.
Academic and industrial participants are already starting
to coordinate on responsible development of AI. For example, we
recently published a paper, together with Stanford, Berkeley,
and Google, laying out a roadmap for AI safety research. Now,
what would help is feedback from the government about what
issues are most concerning to it so that we can start
addressing those from as early a date as possible.
As the Chairman said in his opening statement, Accenture
recently reported that AI has the potential to double economic
growth rates by 2035, which would really make it into the
engine for our future economy. The best way to create a good
future is to invent it. And we have that opportunity with AI by
investing in open, basic research, by creating competitions and
measurement, and by coordinating on safety, security, and
ethics.
Thank you for your time, and I look forward to the Q&A.
[The prepared statement of Mr. Brockman follows:]
Prepared Statement of Greg Brockman, Co-Founder
and Chief Technology Officer, OpenAI
Thank you Chairman Cruz, Ranking Member Peters, distinguished
members of the Subcommittee. Today's hearing presents an important
first opportunity for the members of the Senate to understand and
analyze the potential impacts of artificial intelligence on our Nation
and the world, and to refine thinking on the best ways in which the
U.S. Government might approach AI. I'm honored to have been invited to
give this testimony today.
By way of introduction, I'm Greg Brockman, co-founder and Chief
Technology Officer of OpenAI. OpenAI is a non-profit AI research
company. Our mission is to build safe, advanced AI technology and
ensure that its benefits are distributed to everyone. OpenAI is chaired
by technology executives Sam Altman and Elon Musk.
The U.S. has led the way in almost all technological breakthroughs
of the last hundred years, and we've reaped enormous economic rewards
as a result. Currently, we have a lead, but hardly a monopoly, in AI.
For instance, this year Chinese teams won the top categories in a
Stanford University-led image recognition competition. South Korea has
declared a billion dollar AI fund. Canada produced some technologies
enabling the current boom, and recently announced an investment into
key areas of AI.
I'd like to share 3 key points for how we can best succeed in AI
and what the U.S. Government might do to advance this agenda. First, we
need to compete on applications, but cooperate on open, basic research.
Second, we need to create public measurement and contests. And third,
we need to increase coordination between industry and government on
safety, security, and ethics.
I. Competition and Cooperation
AI applications are rapidly broadening from what they were just a
few years ago: from helping farmers decide which fields to seed, to
warehouse robots, to medical diagnostics, certain AI-enabled
applications are penetrating and enabling businesses and improving
everyday life. These and other applications will create new companies
and new jobs that don't exist today--in much the same way that the
Internet did. But even discovering the full range of applications
requires significant scientific advances. So industry is not just
working on applications: companies like Facebook, Google, and Microsoft
are performing basic research as well, trying to create the essential
AI building blocks which can later be assembled into products.
Perhaps surprisingly, the industry labs are publishing everything
they discover. Publication allows them to pool their resources to
create faster breakthroughs, and to attract top scientists, most of
whom are motivated more by advancing society and improving the future,
than personal financial gain.
Companies stay competitive by publishing their basic research, but
not the details of their products. The inventor of a technique is
usually the first to deploy it, as it has the right in-house
infrastructure and expertise. For example, AI techniques developed by
Google's subsidiary DeepMind to solve Atari video games were applied to
increase the efficiency of Google's own data centers. DeepMind shared
their basic techniques by publishing the Atari research papers, but did
not share their applied work on data center efficiency.
Openness enables academia and industry to reinforce each other.
Andrew Moore of Carnegie Mellon University says it's not unusual that
between 10 and 20 percent of the staff he hires will take leaves of
absence to work in industry or found a startup. Pieter Abbeel, a
researcher at OpenAI, splits his time between OpenAI and the University
of California at Berkeley; likewise, Stanford Professor Fei-Fei Li is
spending time at both Stanford and Google; and many other companies and
organizations work with academics. This ensures that the private sector
is able to master the latest scientific techniques, and that
universities are able to understand the problems relevant for industry.
Openness has concentrated the world's AI research activity around
the U.S. (including attracting many of the Canadian scientists who
helped start the current AI boom), and allowed us to define its culture
and values. Foreign firms like China's Baidu have opened U.S.-based
research labs and have also started publishing. As AI becomes
increasingly useful, the pool of experts we're gathering will be
invaluable to ensuring that its economic activity also remains centered
on the U.S.
Recommendations--
We recommend the following, to ensure that our basic AI research
community remains the strongest in the world:
A. Maintain or increase basic research funding for AI: In 2015,
the government's unclassified investment in AI-related
technology was approximately $1.1 billion, according to The
National Artificial Intelligence Research and Development
Strategic Plan report from the National Science and Technology
Council.\1\ As highlighted by Jason Furman, Chairman of the
Council of Economic Advisers, there's evidence that the
socially optimal level of funding for basic research is two to
four times greater than actual spending.\2\ Given that it only
takes months for a basic AI advance to result in new companies
and products, usually by whoever made the advance, we support
increasing funding for basic research in this domain. If we
want these breakthroughs to be made in the U.S., we'll need to
conduct basic research across a number of subfields of AI, and
encourage the community to share their insights with each
other. We'll need to allow our academics to freely explore
ideas that go against consensus, or whose value has high
uncertainty. This is supported by history: companies like
Google and Microsoft rely on AI technologies that originated
with a small group of maverick academics.
---------------------------------------------------------------------------
\1\ National Science and Technology Council, Networking and
Information Technology Research and Development Subcommittee. 2016.
``The National Artificial Intelligence Research and Development
Strategic Plan'' report: https://www.whitehouse.gov/sites/default/
files/whitehouse_files/microsites/ostp/NSTC/
national_ai_rd_strategic_plan.pdf
\2\ Furman, Jason. 2016. ``Is This Time Different? The
Opportunities and Challenges of Artificial Intelligence'' report:
https://www.whitehouse.gov/sites/default/files/page/files/20160707
_cea_ai_furman.pdf
B. Increase the supply of AI academics: Industry has an
insatiable demand for people with AI training, which will only
increase for the foreseeable future. We need to grow the supply
of people trained in AI techniques; this will let us make more
research breakthroughs, give industry the people it needs to
commercialize the basic science, and train the next generation
of scientists. NSF could explore adjusting its policies to
allow more competitive salaries for those working on Federal
---------------------------------------------------------------------------
academic grants.
C. Enhance the professional diversity of the AI field: Today,
AI consists mostly of individuals with degrees in computer
science, mathematics, and neuroscience, with a significant
gender bias towards men. As AI increases its societal impact,
we need to increase the diversity of professional views within
the AI community. Government can explore making more
interdisciplinary research grants available to incentivize
experts in other fields, such as law or agriculture or
philosophy, to work with AI researchers. We also support the
White House's Computer Science for All initiative, and the
OSTP's recommendation that government should create a Federal
workforce with diverse perspectives on AI.
II. The Need For Public Measurement and Contests
Objective measures of progress help government and the public
distinguish real progress from hype. It's very easy to sensationalize
AI research, but we should remember that advanced AI has seemed just
around the corner for decades. Good policy responses and a healthy
public debate hinge on people having access to clear data about which
parts of the technology are progressing, and how quickly. Given that
some AI technologies, such as self-driving cars, have the potential to
impact society in a number of significant ways, we support OSTP's
recommendation that the government keep a close watch on the
advancement of specific AI technologies, and work with industry to
measure the progression of the technology.
Also, having a measurable goal for AI technologies helps
researchers select which problems to solve. In 2004, DARPA hosted a
self-driving car competition along a 150-mile course in the Mojave
Desert--the top competitor made it only seven miles. By 2007, DARPA
hosted an Urban Challenge to test self-driving cars on a complex, urban
environment, and six of the eleven teams completed the course. Today,
Uber, Google, Tesla, and others are working on commercializing self-
driving car technology.
Similarly, when Fei-Fei Li and her collaborators at Stanford
launched the image recognition ImageNet competition in 2010, it was
designed to be beyond the capabilities of existing systems. That
impossibility gave the world's research community an incentive to
develop techniques at the very edge of possibility. In 2012, academics
won first place using a neural network-based approach, which proved the
value of the technique and kickstarted the current AI boom. The winning
ImageNet team formed a startup and were subsequently hired by industry
to create new products. One member, Ilya Sutskever, is one of my co-
founders at OpenAI, and the other two members work at Google. This
shows how competitions can provoke research breakthroughs, and
translate into an economic advantage for industry.
We're moving from an era of narrow AI systems to general ones.
Narrow AI systems typically do one thing extremely well, like
categorize an image, transcribe a speech, or master a computer game.
General AI systems will contain suites of different capabilities; they
will be able to solve many tasks and improvise new solutions when they
run into trouble. They will require new ways to test and benchmark
their performance. Measuring the capabilities of these new multi-
purpose systems will help government track the technology's progress
and respond accordingly.
Recommendations--
Government can create objective data about AI progress in the
following ways:
A. Modern competitions: AI systems have often been measured by
performance on a static dataset. Modern systems will act in the
real world, and their actions will influence their
surroundings, so static datasets are a poor way to measure
performance. We need competitions which capture more of the
complexity of the real world, particularly in developing areas
such as robotics, personal assistants, and language
understanding. The government can continue designing
competitions itself, as DARPA did recently with the Cyber Grand
Challenge, or support others who are doing so.
B. Government information gathering: Government should gather
information about the AI field as a whole. Researchers tend to
focus on advancing the state of the art in one area, but the
bigger picture is likely to be crucial for policymakers, and
valuable to researchers as well. The government can invest in
careful monitoring of the state of the field, forecasting its
progress, and predicting the onset of significant AI
applications.
III. Increase Coordination Between Industry and Government on Safety,
Security, and Ethics
The Internet was built with security as an afterthought, rather
than a core principle. We're still paying the cost for that today, with
companies such as Target being hacked due to using insecure
communication protocols. With AI, we should consider safety, security,
and ethics as early as possible, and bake these into the technologies
we develop.
Academic and industrial participants are starting to coordinate on
responsible development of AI. For example, we recently worked with
researchers from Stanford, Berkeley, and Google to lay out a roadmap
for safety research in our paper ``Concrete Problems in AI Safety.''
\3\ Non-profit groups like the Partnership on AI and OpenAI are forming
to ensure that research is done responsibly and beneficially.
---------------------------------------------------------------------------
\3\ Amodei, Dario et al., 2016. ``Concrete Problems in AI Safety''
research paper: https://arxiv.org/abs/1606.06565
---------------------------------------------------------------------------
Recommendations--
Industry dialog: Government can help the AI community by giving
feedback about the what aspects of progress it needs to understand in
preparing policy. As the OSTP recommended in its report, Preparing for
the future of Artificial Intelligence,\4\ the NSTC Subcommittee on
Machine Learning and Artificial Intelligence should meet with industry
participants to track the progression of AI. OpenAI and our peers can
use these meetings to understand what we should monitor in our own work
to give government the telemetry needed to calibrate policy responses.
---------------------------------------------------------------------------
\4\ Executive Office of the President, National Science and
Technology Council Committee on Technology. 2016. ``Preparing for the
future of artificial intelligence'' report: https://www.whitehouse.gov/
sites/default/files/whitehouse_files/microsites/ostp/NSTC/preparing_
for_the_future_of_ai.pdf
---------------------------------------------------------------------------
Accenture recently reported that AI has the potential to double
economic growth rates by 2035, which would make it the engine for our
future economy. Having the most powerful economy in the world will
eventually require having the most AI-driven one, and the U.S.
accordingly must lead the development and application of AI
technologies along the way. The best way to ensure a good future is to
invent it.
Thank you for your time and focus on this critical topic. I am
pleased to address any questions.
Senator Cruz. Thank you, Mr. Brockman. And I was encouraged
by your testimony about the Canadian scientists coming to this
country. And I will say, as someone born in Calgary, that I
think there are colleagues of mine on both sides of the aisle
who have concerns about Canadians coming to this country.
[Laughter.]
Senator Cruz. Dr. Chien.
STATEMENT OF DR. STEVE A. CHIEN, TECHNICAL GROUP
SUPERVISOR, ARTIFICIAL INTELLIGENCE GROUP, JET
PROPULSION LABORATORY, NATIONAL AERONAUTICS AND SPACE
ADMINISTRATION
Dr. Chien. Chairman Cruz, Ranking Member Peters, and
members of the Committee, thank you for this great opportunity
to speak to you on this topic of artificial intelligence, and
specifically its relationship to space exploration.
For the record, I'm here as an employee of NASA's Jet
Propulsion Laboratory, which is a federally-funded research and
development center managed by the California Institute of
Technology for NASA.
As a Senior Research Scientist in Autonomous Systems at
JPL, I work on the development and application of artificial
intelligence to NASA's missions. I've had the privilege to lead
the deployment of AI software to NASA's Earth Observing I
mission, NASA's Mars Exploration Rovers mission and also the
European Space Agency's Rosetta mission. We focus on using AI
to improve the effectiveness of conducting science and
observation activities to--in NASA's missions.
I know of no better introduction to this topic than to
point out that, as we speak right now, there's a spacecraft,
called Earth Observing I, that's flying about 7,000 kilometers
overhead, weighs about 500 kilograms, and is flying at 7 and a
half kilometers per second, that is fully under the control of
AI software. This spacecraft has been under the control of this
AI software for over a dozen years and has successfully
acquired over 60,000 images under the control of the software,
and issued over 2.6 million commands. The AI software that's
used in the operation of this mission includes constraint-based
scheduling software to enable the spacecraft to be operated by
end users, scientists and people who monitor natural hazards,
such as volcanoes and flooding. Onboard software, including
machine-learning classifiers, enables the spacecraft to more
effectively monitor these science events--again, flooding,
volcanism, as well as cryosphere, the freeze and thaw of the
Earth's environment.
Furthermore, in a range of collaborations all around the
world, this spacecraft has been integrated into a network with
other space systems as well as ground sensor networks. And
these--the extent of this multi-agent AI system goes as far as
Thailand, Iceland, Sicily, Namibia, and even Antarctica. What
this system enables us to do is enable data from one part of
the system, such as a seismographic sensor at a volcano in
Antarctica, to trigger the observation of the system via space
assets.
Going even further afield, on Mars, autonomous navigation
software is at the heart of all of the Mars Rover exploration
missions. And this is, at its core, AI-based search software.
AI and computer-vision software form the core of the AEGIS
system, which is now operational on both the Mars Exploration
Rover mission and the Mars Science Laboratory Rover. AEGIS
enables the Rovers to automatically target science measurements
based on general science criteria, such as texture, size,
shape, and color, without the ground in the loop, dramatically
enhancing the science that the Rovers can conduct.
Machine learning has also had significant impact in dealing
with the enormous data sets that space missions produce. Just
two examples. In the very long baseline array, radio science is
being enhanced by machine learning. Machine learning is used to
identify millisecond-duration radio transients and reject radio
frequency interference events. Here, machine learning allows
the automatic triage from thousands of candidates down to tens
of candidates for manual review by highly expert scientists.
In visual astronomy, in the Intermediate Palomar Transient
Facility, machine learning is applied to identifying
transients. Point transients--point source transients are
typically supernova, and streaking transients are near-Earth
objects. Here, machine learning has been used to perform vast
daily triage of millions of candidate events down to tens of
events; again, allowing the human experts to focus on the most
likely candidates and enhance the science.
While these examples may give you the impression that AI is
commonplace in space exploration, I assure you this is not the
case. The above examples are a sampling of AI success stories
on a small fraction of the overall space missions. Because of
the high-stakes nature of space exploration, the adoption of
disruptive technologies like AI requires an extensive track
record of success as well as continuous contact with the key
stakeholders of science, operations, and engineering. However,
AI has made tremendous progress in the recent years.
Instruments in the Mars 2020 Rover will have unprecedented
ability to recognize features and retarget themselves to
enhance science. The Mars 2020 Rover mission is also
investigating other use of onboard scheduling technologies to
best use available Rover resources. And the Europa multi-flyby
mission is also investigating the use of onboard autonomy
capabilities to achieve science despite Jupiter radiation--the
Jupiter radiation environment, which causes processor resets.
In the future, AI will also have applications in the manned
program in order to best use scarce astronaut time resources.
Past efforts have placed AI in a critical position for future
space exploration to increasingly hostile and distant
destinations. What we need is sustained resources and a
commitment, support, and vision for AI to fulfill its vast
potential to revolutionize space exploration.
Thank you very much for your time.
[The prepared statement of Dr. Chien follows:]
Prepared Statement of Dr. Steve A. Chien, Technical Group Supervisor,
Artificial Intelligence Group, Jet Propulsion Laboratory, National
Aeronautics and Space Administration
Chairman Cruz, Ranking Member Peters, and Members of the Committee,
thank you for the opportunity to speak to you on this topic of
Artificial Intelligence (AI), and specifically it's relation to space
exploration.
For the record, I am here as an employee of NASA's Jet Propulsion
Laboratory, which is a Federally Funded Research & Development Center,
managed by the California Institute of Technology for NASA.
As a Senior Research Scientist specializing in Autonomous Systems
at JPL, I work on the development and application of Artificial
Intelligence to NASA missions. I have had the privilege to lead the
deployment of AI software to NASA's Earth Observing One and Mars
Exploration Rovers missions, as well as for European Space Agency's
Rosetta mission. Separately, The Artificial Intelligence Group has
deployed additional AI software to the Mars Exploration Rovers and Mars
Science Laboratory missions, as well as to NASA's Deep Space Network.
In my group and related groups at JPL, we focus on using AI to improve
the performance of space exploration assets: to conduct more science,
improve response to track science phenomena and natural hazards, and
increase the efficiency of operations.
I know of no better introduction to this topic than to point out
that as we speak, a spacecraft, Earth Observing One, weighing 500 kg,
flying at about 7.5 km/s, at about 700km altitude, is operating under
the control of Artificial Intelligence software called ``The Autonomous
Sciencecraft.'' This software, which has parts both on the spacecraft
and in the ground system, has been the primary operations system for
this mission for over a dozen years. In this time, the spacecraft has
acquired over 60,000 images and issued over 2.6 million commands.
This AI software has improved the efficiency of spacecraft
operations using AI constraint-based scheduling technology, enabling
direct tasking by end users such as scientists and natural hazard
institutions. Additionally, onboard smarts (including AI/Machine
Learning classification techniques) are used to detect and track
volcanic activity, wildfires, and flooding to enable rapid generation
of alerts and summary products. The most advanced of this software uses
imaging spectroscopy to discriminate between different substances in
images--these techniques have wide applications to environmental
monitoring.
Furthermore, in a range of collaborations, this spacecraft has been
networked together (via the ground and Internet) in a sensorweb with
other spacecraft and ground sensor networks to provide a unique
capability to track volcanism, wildfires, and flooding worldwide, with
linkages to Thailand, Iceland, Hawaii, Sicily, Namibia, and even
Antarctica to name a few. This AI multi-agent system enables detections
from one part of the system to automatically trigger targeted
observations from another part of the system, as well as enabling
autonomous retrieval, analysis, and delivery of relevant data to
interested parties.
On Mars, the autonomous navigation software used on all of the Mars
rovers has at its core AI-based search software. AI and computer vision
software form the core of the Autonomous Exploration for Gathering
Increased Science (AEGIS) system, now in operational use on both the
Mars Exploration Rover and Mars Science Laboratory Rovers. AEGIS
enables the rovers to autonomously target science measurements based on
general science criteria such as texture, size, shape, and color
without the ground in the loop, thereby improving rover science
productivity.
Machine Learning also has significant impact in dealing with the
enormous datasets generated in science observatories. Just a few
examples follow:
In the Very Long Baseline Array (VLBA) Fast Radio Transients
Experiment (V-FASTR), Machine Learning is used to identify
millisecond duration radio transients and reject radio
frequency interference in the VLBA. This Machine Learning
enables fast triage of order of 10\3\ transient candidates
daily to 10's of candidates for human review.
In the Intermediate Palomar Transient Factory (i-PTF),
Machine Learning is applied to visual imagery to identify
candidate point source (e.g., supernovae) and streaking (e.g.,
near Earth Asteroids) transients for daily fast triage from
order of 10\6\ candidates to 10's of candidates for human
review.
Significant AI technology is used in the scheduling systems for
space missions. These systems enable the operations teams to manage the
incredible complexity of spacecraft and science with often thousands to
tens of thousands of science and engineering activities and
constraints. These systems include SPIKE for Hubble Space Telescope,
Spitzer Space Telescope, as well planned use for the James Webb Space
Telescope, the MAPGEN use for the Mars Exploration Rovers and LADEE
Activity Scheduling System (LASS) for the Lunar Atmospheric Dust
Environment Explorer (LADEE) mission. In addition, NASA's Deep Space
Network, used for communications to all of the NASA missions beyond
Earth Orbit, uses AI scheduling technology.
While these examples may give you the impression that AI is
commonplace in space exploration, I assure you that this not the case.
The above examples represent a sampling of AI success stories on a
small fraction of the overall set of space missions. Because of the
high-stakes nature of space exploration, adoption of disruptive
technologies like AI requires an extensive track record of success as
well as continuous contact with the critical stakeholders of science,
operations, and engineering. However, due to both technology advances
and increased stakeholder understanding of the great promise of AI,
progress has accelerated dramatically in recent years. For example,
instruments on the Mars 2020 rover will have unprecedented ability to
recognize features and retarget to enhance science. Mars 2020 is also
investigating the use of an onboard re-scheduling capability to best
use available resources. The Europa Multiple-Flyby mission is studying
autonomy capabilities needed to achieve science in the presence of
Jovian radiation induced processor resets.
In the future, AI will likely have many applications in human
spaceflight missions where astronaut time is at a premium, as well as
in robotic missions where the technology may enable missions of
increasing complexity and autonomy. Past efforts have placed AI in
critical position for future space exploration. Sustained resources,
support, and vision are needed for AI to fulfill its vast potential to
revolutionize space exploration.
For further information see:
Autonomous Sciencecraft/Earth Observing One
S. Chien, R. Sherwood, D. Tran, B. Cichy, G. Rabideau, R.
Castano, A. Davies, D. Mandl, S. Frye, B. Trout, S. Shulman, D.
Boyer, ``Using Autonomy Flight Software to Improve Science
Return on Earth Observing One,'' Journal of Aerospace
Computing, Information, & Communication, April 2005, AIAA.
Earth Observing Sensorweb
S. Chien, B. Cichy, A. Davies, D. Tran, G. Rabideau, R.
Castano, R. Sherwood, D. Mandl, S. Frye, S. Shulman, J. Jones,
S. Grosvenor, ``An Autonomous Earth Observing Sensorweb,'' IEEE
Intelligent Systems, May-June 2005, pp. 16-24.
S. Chien, J. Doubleday, D. Mclaren, D. Tran, V. Tanpipat, R.
Chitradon, S. Boonya-aroonnet, P. Thanapakpawin, D. Mandl.
Monitoring flooding in thailand using earth observing one in a
sensorweb. IEEE Journal of Selected Topics in Applied Earth
Observations and Remote Sensing. 2013 Apr;6(2):291-7.
S. Chien, J. Doubleday, D. Mclaren, A. Davies, D. Tran, V.
Tanpipat, S. Akaakara, A. Ratanasuwan, D. Mandl. Space-based
Sensorweb monitoring of wildfires in Thailand. In Geoscience
and Remote Sensing Symposium (IGARSS), 2011 IEEE International
2011 Jul 24 (pp. 1906-1909). IEEE.
AG Davies, S Chien, R Wright, A Miklius, PR Kyle, M Welsh, JB
Johnson, D Tran, SR Schaffer, R Sherwood. Sensor web enables
rapid response to volcanic activity. EOS, Transactions American
Geophysical Union. 2006 Jan 3;87(1):1-5.
A. G. Davies, S. Chien, J. Doubleday, D. Tran, T. Thordarson,
M. Gudmundsson, A. Hoskuldsson, S. Jakobsdottir, R. Wright, D.
Mandl, ``Observing Iceland's Eyjafjallajokull 2010 Eruptions
with the Autonomous NASA Volcano Sensor Web'', Journal of
Geophysical Research--Solid Earth, v. 118, Issue 5, pp. 1936-
1956, May 2013.
AEGIS/MER
TA Estlin, BJ Bornstein, DM Gaines, RC Anderson, DR Thompson, M
Burl, R Castano, M Judd. AEGIS automated science targeting for
the MER opportunity rover. ACM Transactions on Intelligent
Systems and Technology (TIST). 2012 May 1;3(3):50.
AEGIS/MSL http://www.jpl.nasa.gov/news/news.php?feature=6575
V-FASTR
D. R. Thompson, K. L. Wagstaff, W. Brisken, A. T. Deller, W. A.
Majid, S. J. Tingay, and R. B. Wayth. ``Detection of fast radio
transients with multiple stations: A case study using the Very
Long Baseline Array.'' The Astrophysical Journal, 735(2), doi:
10.1088/0004-637X/735/2/98, 2011
Impact of ML component:
K. L. Wagstaff, B. Tang, D. R. Thompson, S. Khudikyan, J.
Wyngaard, A. T. Deller, D. Palaniswamy, S. J. Tingay, and R. B.
Wayth. ``A Machine Learning Classifier for Fast Radio Burst
Detection at the VLBA.'' Publications of the Astronomical
Society of the Pacific, 128:966(084503), 2016.
Scientific impact:
S. Burke-Spolaor, C. M. Trott, W. F. Brisken, A. T. Deller, W.
A. Majid, D. Palaniswamy, D. R. Thompson, S. J. Tingay, K. L.
Wagstaff, and R. B. Wayth. ``Limits on Fast Radio Bursts from
Four Years of the V-FASTR Experiment.'' The Astrophysical
Journal, 826(2), doi:10.3847/0004-637X/826/2/223, 2016.
Intermediate Palomar Transient Facility
F. J. Masci, R. R. Laher, U. D. Rebbapragada, G. B. Doran, A.
A. Miller, E. Bellm, M. Kasliwal, E. O. Ofek, J. Surace, D. L.
Shupe, C. J. Grillmair, E. Jackson, T. Barlow, L. Yan, Y. Cao,
S. B. Cenko, L. J. Storrie-Lombardi, G. Helou, T. A. Prince,
and S. R. Kulkarni, The IPAC Image Subtraction and Discovery
Pipeline for the intermediate Palomar Transient Factory, Draft
manuscript for Publications of the Astronomical Society of the
Pacific https://arxiv.org/abs/1608.01733
A. Waszczak, T. A. Prince, R. Laher, F. Masci, B. Bue, U.
Rebbapragada, T. Barlow, J. Surace, G. Helou, S. Kulkarni,
Small near-Earth asteroids in the Palomar Transient Factory
survey: A real-time streak-detection system, to appear,
Publications of the Astronomical Society of the Pacific.
https://arxiv.org/pdf/1609.08018v1.pdf
B.D. Bue, K. L. Wagstaff, U. D. Rebbapragada, D. R. Thompson,
and B. Tang. Astronomical Data Triage for Rapid Science Return.
Proceedings of the Big Data from Space Conference, 2014.
ASPEN-RSSC/Rosetta
S. Chien, G. Rabideau, D. Tran, J. Doubleday. M. Troesch, F.
Nespoli, M. Perez Ayucar, M. Costa Sitja, C. Vallat, B. Geiger,
N. Altobelli, M. Fernandez, F. Vallejo, R. Andres, M. Kueppers,
``Activity-based Scheduling of Science Campaigns for the
Rosetta Orbiter,'' Invited Talk, Proc. International Joint
Conference on Artificial Intelligence (IJCAI 2015), Buenos
Aires, Argentina. July 2015.
SPIKE/Hubble
MD Johnston, G Miller. Spike: Intelligent scheduling of hubble
space telescope observations. Intelligent Scheduling. 1994:391-
422.
Spike/JWST
ME Giuliano, R Hawkins, R Rager. A status report on the
development of the JWST long range planning system. In Proc.
International Workshop on Planning and Scheduling for Space,
ESOC, Darmstadt, Germany 2011.
MAPGEN/MER
JL Bresina, AK Jonsson, PH Morris, K Rajan . Activity Planning
for the Mars Exploration Rovers. In Proc International
Conference on Automated Planning and Scheduling 2005 Jun (pp.
40-49).
MEXAR/Mars Express
A Cesta, G Cortellessa, M Denis, A Donati, S Fratini, A Oddi, N
Policella, E Rabenau, J Schulster. Mexar2: AI solves mission
planner problems. IEEE Intelligent Systems. 2007 Jul;22(4):12-
9.
LASS/LADEE
JL Bresina. Activity Planning for a Lunar Orbital Mission. In
AAAI 2015 Jan 25 (pp. 3887-3895).
Spitzer Space Telescope
DS Mittman, R Hawkins. Scheduling Spitzer: The SIRPASS Story,
Proc Intl Workshop on Planning and Scheduling for Space,
Moffett Field, CA, 2013.
NI-SAR
JR Doubleday. Three petabytes or bust: planning science
observations for NISAR. In SPIE Asia-Pacific Remote Sensing
2016 May 2 (pp. 988105-988105). International Society for
Optics and Photonics.
Radio Constellation Design
S. Schaffer, A. Branch, S. Chien, S. Broschart, S. Hernandez,
K. Belov, J. Lazio, L. Clare, P. Tsao, J. Castillo-Rogez, E. J.
Wyatt, Using Operations Scheduling to Optimize Constellation
Design, Scheduling and Planning Applications Workshop (Spark),
International Conference on Automated Planning and Scheduling,
London, UK, June 2016.
Deep Space Network
MD Johnston, D Tran, B Arroyo, S Sorensen, P Tay, B Carruth, A
Coffman, M Wallace. Automated Scheduling for NASA's Deep Space
Network. AI Magazine. 2014 Dec 22;35(4):7-25.
Manned Program AI
G Aaseng, Techport: Advanced Caution and Warning System
Project, https://techport.nasa.gov/view/32946
A. Haddock, J. Frank, L. Wang, TechPort: Autonomous Systems and
Operations project, https://techport.nasa.gov/view/32946
Senator Cruz. Thank you, Dr. Chien.
And thank you, to each of you, for your testimony.
Let me start with just a broad question to the panel, which
is, What are the greatest challenges and opportunities you see
for the continued development of AI?
Dr. Moore. I do think it's very important that we grow our
AI workforce quickly. And it's interesting that, in a world
where we're actually all concerned about making sure there are
more jobs available, there's such a gap here, where we're so
short of experts. Frankly, I look at some of the other major
players around the world in this area, I see that China, India,
and other countries are really pumping out the computer
scientists who can form this cohort. So, for me, I would feel
much more comfortable if we were graduating hundreds of
thousands of AI experts every year from our universities,
instead of just thousands.
Dr. Horvitz. So, let me also complement that work by
talking about some technical directions. I mentioned human/
computer or human/AI collaboration. And we don't think enough
about the human-factor angle in AI. It's not all about
automation. Of course, there'll be some interesting automation.
We can't have people on Mars, for example, looking at those
stones and doing the digging. But, in general, there are
incredible opportunities ahead with codesigning systems so they
work really well. They're human-aware. They understand human
attention. They understand how they can complement human
intellect and what people do uniquely, and do well.
Understanding how to negotiate, to do a give-and-take, a fluid
dialogue in contributions between humans and machines. Lots to
be done there, and that includes this piece with explanation,
transparency. Many of these answers we get today out of AI
systems, the best systems we can build are black-box systems
that are opaque to human beings who need to understand to learn
how to justify those decisions and how the thinking is done,
and to understand the reasoning process, itself. Lots of work
to do there.
There's another critical direction with thinking through
opportunities to take some of the software we've done in the
intellectual cognitive space and enter into the real world of
physical innovation, to teach systems to work in physical
spaces. Our robotics today are very, very limited. Even our
best practices on Mars don't do the kinds of things that people
can do easily. And there's a coming renaissance in taking some
of our advances in AI and bringing them into the physical
space.
Mr. Brockman. So, I believe that the biggest opportunity we
have is to continue to move very quickly on the fundamental
building blocks, on the fundamental technology. And it really
feels like, today, we're kind of in the vacuum-tube era and
that the transistor is out there, and that we're building very
impressive technologies, but that, this is really just the tip
of the iceberg. And I think that the biggest thing to watch out
for--I think one of the biggest risks is--that we lose the
openness that we have. Today, we can have these conversations,
we can measure how the technology is progressing, and we can
plan for the future. And I think that we can continue to
attract the world's best talent by keeping it open.
Senator Cruz. So, in the early 2000s, I chaired a series of
hearings at the Federal Trade Commission on barriers to e-
commerce. And I'd be interested in the panel's opinion. Are
there particular legal or regulatory barriers or other barriers
to entry that are slowing down or impeding the development of
AI?
Dr. Horvitz. One comment is on--I'll make mention--is on
data. With the growth of AI and the importance of data in AI,
there has been a growth of a genuine need for innovation with
privacy to secure the privacy for individuals. At the same
time, there are massive data assets that aren't easily
available. We had to negotiate at Microsoft Research to gain
access to FAA data to help us build new weather maps for the
country based on thousands of planes in flight as we speak. We
were helped by the OSTP in getting access to that, but it was
not necessarily a simple task. But, there are many data sets
like this, and we'd love to see public-sector data sets,
especially with privacy-protected, made more available for
innovation. At the same time, we--while the NIH requires, on
NIH-funded projects, for data to be released as part of the
contracts that are made with researchers, it's very difficult
to have medical data shared as a part of the fulcrum of
innovation. And so, we need to think through HIPAA, altruistic
data approaches, where patients donate data, new kinds of
programs that let us really maintain patient privacy while
gaining access to large amounts of biomedical data.
Dr. Moore. There are some other areas, such as intelligent
braking in cars, where there are some legislative questions
which might slow us down. For example, it would be tragic if
some lifesaving technology, which would make cars safer,
couldn't be released because the legal questions about who is
responsible got in the way. What I'm talking about here is, if
I, as a software developer, invent a new collision-avoidance
system which unambiguously reduces fatalities by a factor of
three, but occasionally, unfortunately, 1 in 1,000 times, maybe
there's a disaster, there is a difficult question as to how,
legislatively, we make sure we're ready for this. So, I can
imagine a potential impasse between insurance companies,
policymakers, drivers, and car manufacturers, where no one is
willing to put lifesaving technology into these systems because
it's still ambiguous who has the responsibility for what.
Senator Cruz. So, one final question. General AI has
generated some significant fears and concerns from scientists
and innovators, such as Stephen Hawking, Bill Gates, and Elon
Musk. Stephen Hawking has stated, ``Once humans develop
artificial intelligence, it would take off on its own and
redesign itself at an ever-increasing rate. Humans, who are
limited by slow biological evolution couldn't compete and would
be superseded.'' And Elon Musk has referred to it as,
``summoning the demon.'' How concerned should we be about the
prospects of general AI? Or, to ask the question differently,
in a nod to Terminator, does anyone know when Skynet goes
online?
[Laughter.]
Mr. Brockman. So, my answer to that is that I think that,
with artificial intelligence generally, that there are a lot of
things that we should be careful about and concerned about and
think about security, safety, and ethics today. And so, I think
that the kind of general intelligence that people talk about,
my 90-percent confidence interval on when that kind of
technology could arrive, is between 10 to 100 years. It's not
something that we're at all capable of building today. And
today we know that there are concrete safety problems that we
can be working on. And so, I think that we should be investing
in those kinds of questions. And I think that that will help us
figure out the right answers for the short, medium, and long
term.
Dr. Horvitz. So, there has been a lot of hyperbole, as you
know, stimulated in no small part by Hollywood. Great--these
are great themes, and they keep us enamored with interesting
possibilities. At the same time, we don't scoff at those kinds
of long-term outcomes, and want to seriously reflect and review
possibilities, push to the limit some of these proposals about
what's possible, and, in advance, proactively work to thwart
them, to stay on a healthy, secure path.
My own sense is, these are very, very long-term issues, but
that the things we're doing today are actually relevant and
interesting, in terms of thinking about how AI systems can
grapple with unknown unknowns, how it could secure systems
from, for example, modifying themselves, their own objective
functions, which is one of the concerns that comes up at times.
In some ways, I am happy to see the growth of interest in the
long-term future questions, because it raises my confidence
that we will track closely and do the best we can when it comes
to harnessing AI for the greatest benefits.
Dr. Moore. I would just add that, at the moment, everything
that's going on in the current AI revolution is using AIs which
are like idiot savants. They are able to search a space that
we've prescribed really efficiently. And it is a matter for
future researchers, not something immediate, to imagine these
kinds of self-aware, self-reasoning systems. Those would be
really, really important to get right. At the moment, the AIs
we're building are all to do with immediately answering
questions about agriculture, safety, people's health. And the
students who are being drawn into it are being drawn into it
for these idealistic reasons.
One thing you will notice--and this is actually under the
influence of some of the institutions my colleagues have put
into place--is, many places, such as Carnegie Mellon, are
actively making ethics and responsibility a central part of the
curriculum for these AI experts. Because these kids today are
building the 21st century. We need them to actually understand
the human condition while they're doing it.
Dr. Horvitz. Just to get a sense for the kinds of things
that are going on, this coming spring there's going to be an
event where we're inviting--or a group is inviting out people
who are imagining the most fearful--feared long-term AI
scenarios--call them the Red Team--and we're--then we're
inviting out the Blue Team to disrupt them in advance, and
they're going to come together and battle it out.
Dr. Chien. So, I would like to take this chance to circle
back to one of your earlier questions and tie that in. You
asked, What are the areas that we need to work in? I would say
that one of the key areas that we need to work in is better
characterization and understanding the performance of AI
systems. And this is something that we have a lot of interest
in at NASA, because, in our space missions, we need to, if not
prove that they're going to actually perform within certain
bounds, we need to have very strong confidence that they will
perform in those bounds, because these are very high-stakes
missions. A lot of the applications that people have talked
about--healthcare, self-driving cars--these also are high-
stakes missions. Before AI can control our critical
infrastructure, we need to be confident that it will perform as
we want it to perform. And I think this has been identified
before in the OSTP study as a key area of research.
Senator Cruz. Thank you very much.
Senator Peters.
Senator Peters. Thank you, Mr. Chairman.
Again, thank you, to our witnesses.
And you're right, Dr. Chien, we have to make sure this
performs. I've been very involved in autonomous vehicles and
the research that's going on there. And my concern is that we
have to make sure the technology gets it right with as few
errors as possible, because there's already limited consumer
acceptance for letting some machine drive your automobile
through a city street. There are all sorts of benefits, which
we've talked about, but limited acceptance now. And if you had
some sort of catastrophic event in a crash--and there will be
some crashes, certainly--it could very well set back the
industry dramatically, because of the consumer pushback and the
public pushback. So, we have to do this in thoughtful ways,
which is why, Dr. Moore, some of the regulatory aspects of
this, before you put vehicles on the road, to make sure there's
proper safety in place, or we've thought through how we ensure
that, is incredibly important.
My concern with all of this has always been that there's a
disconnect between the speed we're on with public policy versus
technology. Right now, we are at a exponential rate when it
comes to technology. And even though we have estimates of AI
reaching a singularity of some sort from 10 to 100 years--we
don't know when that is, although things seem to operate a lot
quicker than we anticipate. I believe that we didn't think we
could beat the expert player in Go for a least a decade, and I
think that just occurred a few months ago. So we can't fully
anticipate what's happening.
I don't know the speed it will go at, but it will probably
be quicker than we anticipate. The one constant in all of this
is, when it comes to public policy, that operates at a constant
speed. It's called ``snail speed.'' So, it is very slow and
cumbersome. If we are not doing it now, we have no chance of
trying to catch up to what's happening with the policies, going
forward.
I certainly appreciate the comments from several of you
that we have to be thinking about this stuff now, in a very
thoughtful, comprehensive way, because if we wait, it's going
to be too late.
I want to switch gears to the positive aspects that we want
to continue to move AI forward. You've mentioned some of the
challenges that we have: the gaps that we have to fill. I'd
like your perspective on where the Federal Government's role
is, in terms of research. Mr. Brockman, you mentioned in your
testimony some subfields that need work and some other areas.
But, I'd like to ask each of you.
Obviously, private industry is already invested. In my
opening comments, I mentioned eight and a half billion dollars
in 2015. That number is going to continue to go up. So, private
industry is doing an awful lot of this work, including basic
research, which, traditionally, has been an area where the
Federal Government has supported academic research through
grants, but some of that basic research is being done by
private industry, as well. So, that's occurring. Not
necessarily in other areas. But, are there gaps where you
believe the Federal Government--there isn't going to be a
private industry group out there investing in some of these
gaps that we need to figure out. The Federal involvement will
be critical to investing in those kinds of research programs,
first to make sure that AI moves forward in a societal
beneficial way, but also to understand the time constraints
associated with the competition that we face from the Chinese
and Koreans and other folks.
Dr. Horvitz. So, one comment I'll make is that, beyond
industry, looking at private sector and public sector,
academia, there are groups coming together, so I'll just make a
few comments about the new Partnership on AI. The full name is
Partnership on AI to Benefit People and Society. And this is a
nonprofit organization that was formed by Facebook and Amazon,
Google, IBM, and Microsoft coming together, working with
nonprofit teams to--with Balance Board and so on, focused
around sets of these long-term challenges and shorter-term
challenges, with safety-critical systems, ethics, and society,
notions of how people and machines work together, and even
working to stimulate new kinds of challenges and catalyzing new
efforts in AI that might not be done naturally by industry.
That's one direction. I'm happy to answer questions about that
effort, which is ongoing.
Another couple of comments is that there are places and
opportunities where we don't necessarily see industry making
deep investments. I would call these application areas that are
rich and ripe for AI innovation. How can we solve homelessness,
or address homelessness, addiction, related problems in the
social science sphere and social challenges sphere? There are
some teams in academia right now working hard at applications
in this space. Recently at USC, the engineering department
joined with the social work department. The social work
department, looking at innovative applications of AI and
optimization and decisionmaking to new kinds of policies that
can address these long-term, hard, insidious problems.
Dr. Moore. Very good. I could not agree more with what
you're describing.
Another example of this phenomenon is, I have two brilliant
faculty in the Human-Computer Interaction Institute at Carnegie
Mellon who are looking at AI to help people make prosthetic
hands easily for folks who have lost their limbs. And, they're
struggling to find $50,000 or $100,000 here or there to build
these things. At the same time, frankly, my friends from
industry will be offering these same faculty $2 million or $3
million startup packages to move into industry. So, I do want
to make sure that the folks in academia who are building these
things are successful.
Another example is, in the defense world, tools for helping
our warfighters or other folks domestically who are putting
themselves into danger to save other people. There is so much
opportunity to use sensing, robotics, and artificial
intelligence planning to save lives there. That's an area where
it will take a very long time to grow naturally in the private
sector. And we have faculty, and especially students, champing
at the bit to work on these kinds of problems.
There's another area, which may sound too theoretical, but
I've got to tell you about it, because it's so exciting. The
big Internet companies' big search engines are powered by
things called knowledge graphs, the underlying set of facts
about the world which you can chain together to make
inferences. A large group of us from academia and industry, and
from some government agencies, want to work to create a public,
open, large knowledge graph, which will permit small AI
entrepreneurs to tap into the same kind of knowledge of the
world that the big Internet companies have at the moment. So,
in a manner equivalent to how lots of individual networking
systems came together to form the TCIP protocol for the
Internet, there's something we can do there.
Finally--and this one is going to sound really abstract--
the really good ideas at the moment in machine learning and
deep learning came out of mathematics and statistics. Without
the fundamental work going on by the mathematicians and
statisticians around the world, we wouldn't be where we are.
So, statisticians, who are often the heroes in AI, need help to
progress their field forward as well.
Mr. Brockman. I have three suggestions. The first of these
is basic research. And you mentioned that basic research is
happening in industry. But, I think that we just cannot do too
much of it, in that we really are at the tip of the iceberg
here, and I think that we're just going to find so much value.
And that's why the big companies are investing, because they
realize that, as many dollars that are being made today, that
there's 100X or maybe more increase in the future. And I think
that it's really important that the technology is not owned by
just one or a few big companies. I think it's really important
that the benefits and the technology are owned by us all, as
Americans and as the world. And so, I think that the government
can really help to ensure that this technology is democratized
and that we move faster.
The second is measurement and contests. I think that, for
the reasons I mentioned earlier, that it's really important
that we track how it's progressing so we can have a good
debate. And I think that the government has actually been
extremely successful in the past with investing in contests.
And so, I think you're creating new measurements or supporting
people in industry and academia who are doing the same.
And then the third is safety, security, ethics. I think
that's going to take everyone. And I think that we all need to
work together. I think that that's going to require making sure
that there is funding available for people who want to be
thinking about these issues. And I think that's going to feed
back into all of the questions of--that everyone's been raising
here today.
Senator Peters. Mr. Brockman, I think I saw that you
thought philosophers should be part of that. So--in addition to
technologists--I appreciated that. As someone with a Master's
in Philosophy, that's good. So, I appreciate that.
Mr. Brockman. It's going to take everyone.
Senator Peters. Dr. Chien.
Dr. Chien. Yes. So, I would echo some of the statements
that the other panelists made. They've identified a lot of
great topics for the--that really require government--a
government role. One that I would emphasize is very basic
science questions that relate to NASA's mission. So, how did
the universe form? How did the solar system form? How did life
come into existence on this planet and other planets? These are
actually fundamental questions of science and exploration that
we really need to leverage AI to go and explore all these nooks
and crannies in the solar system. And if you really want to
think far out, in order to embark on an interstellar mission to
see if there's extant life at other solar systems. These are
different questions that there's no clear financial motive, so
there's a clear role for the government, to be able to answer
these kinds of basic science questions.
Mr. Brockman. And if I could just add one last thing. So, I
believe that the statistic for the amount of government
unclassified dollars that went into AI R&D in 2015 was $1.1
billion. And that--as has been mentioned several times--that
industry investment is $8 billion. And if this is a technology
that's really going to be affecting every American in such a
fundamental way, I think that that disparity, I think, is going
to be something that we should act to correct.
Senator Peters. Great. Thank you for your answers.
Appreciate it.
Senator Cruz. Thank you.
Senator Schatz.
STATEMENT OF HON. BRIAN SCHATZ,
U.S. SENATOR FROM HAWAII
Senator Schatz. Thank you.
Dr. Moore, you talked mostly about the unambiguously
positive potential applications of AI. And we've sort of
briefly touched upon the terrifying science fiction
possibilities, which I think we're, you know, joking aside,
keeping an eyeball on, but that is from 10 to 100 years from
now. What I'm interested in is, as Senator Peters mentioned,
What are the tough, thorny, short-term public policy and
ethical challenges that we're facing right now? Not the
possibility that machines will overtake us. Not even the sort
of question of long-term unemployment. But, I think about
doctrine of war, I think about blackbox algorithms that help
with policing, or social work or healthcare. And I'm wondering
if you could, maybe just going down the line, starting with Dr.
Horvitz, give me an example of a short-term ethical, moral,
public policy quandary that is upon us now.
Dr. Horvitz. Well, for one, I think that we'll be seeing
interesting legal tests and precedents set up that define new
kinds of frameworks for dealing with things like liability. Who
or what is responsible? Manufacturers? The drivers of cars? The
people who have signed various documents when cars were
purchased? I think that we haven't--things are unsettled in
that space, and we'll be seeing lots of interesting work there.
And there are some very interesting focused workshops and
conferences where people ask these questions.
When it comes to using various AI technologies, going from
machine learning for building classifiers that do predictions
and that are used to reason about interesting problems like
criminal justice challenges. Should this person charged with a
crime have to stay in jail in advance of their court date, or
can they get out early if they can't pay their bail? They're
the systems out there that have been used and critiqued, and
it's pretty clear that there is opportunity for looking very
carefully at systems that are used in high-stakes situations
like this to ensure that there are not implicit biases in those
systems, to assure that there's accountability and fairness.
And----
Senator Schatz. So, long as it's not a government contract,
where you're working with a subcontractor, which says, ``Our
algorithm is proprietary. You're not allowed to--we just spit
out our recommendation. That's what you pay us for.''
Dr. Horvitz. Well, that's exactly where I'm going. That's
exactly where I'm going. So, the question would be, ``What are
best practices?'' for example, and do we need them when it
comes to these kinds of applications? For example, potentially
with protecting privacy, should datasets used in these
applications be disclosed and disclosable for study and
investigation and interrogation by people who want to make sure
that they're fair and that there can be trust in these systems?
The basic idea here is that many of our datasets have been
collected in advance, with assumptions we may not deeply
understand, and we don't want our machine-learned applications
used in high-stakes applications to be amplifying cultural
biases or any kind of biases that was part of the collection
process.
Senator Schatz. Right.
Why don't we go, very quickly, because I have one final
question, but I'd be interested to hear each one of you quickly
answer this question.
Dr. Moore. Very briefly. This AI technology is available to
the bad guys, too. It is possible to cheaply set up homemade
drones in a bad way. A repressive regime can now use face
recognition in a way that they couldn't last year. We need to
actually stay ahead. We can't just sit where we are.
Mr. Brockman. So, I'd like to actually build on the bias
answer and just say that one thing that concerns me is the lack
of diversity in the field, especially as we try to think about,
How can we ensure that these systems are going to do the right
things for all of us? And if you look at this panel, we're
actually, I think, pretty representative of what the current
field of AI looks like. And I think that we, the government and
industry and academia, need to work together in order to
correct that.
Dr. Chien. I would echo Eric's comments on--we need to
further understand how to characterize the performance of AI
systems. Oddly enough, there are analogues, from social science
to space science, where we work very heavily. We need to show
that the datasets collected by our (NASA) autonomous systems
are representative samplings of what you would get if you were
not smartly collecting the data. Otherwise, you'll actually
come up with different scientific theories and mechanisms for
explaining things. These same kinds of techniques apply to
making sure that your algorithms are not biased in performing
as you wish.
Senator Schatz. So, let me just wrap up with this. And I'll
ask a question for the record. My question is sort of
mechanical. Dr. Horvitz and many of the other testifiers have
made at least a brief reference to the ethical quandaries that
we are facing, a Blue Team/Red Team. I noted, Mr. Brockman, you
made reference to safety, security, and ethics. And it's--it
occurs to me that, as this accelerates so fast, that, as you do
your conferences, as you have your conversations, you may not
be--you may not have fully articulated what kind of system
among the AI community you really want to wrestle with these
questions, whether it's a public-private partnership, whether
it's led by the Federal Government or convened by the Federal
Government, but primarily driven by private-sector actors. I
don't know. But, it occurs to me, lots of good thinking is
occurring. It also occurs to me that maybe it hasn't been
fleshed out from a process standpoint. And we can't take it for
granted that it's all going to happen organically. But, I will
take that question for the record, in the interest of time.
Thank you.
Senator Cruz. Thank you.
Chairman Thune.
STATEMENT OF HON. JOHN THUNE,
U.S. SENATOR FROM SOUTH DAKOTA
The Chairman. Thank you, Mr. Chairman, for convening
today's Subcommittee hearing on artificial intelligence. This
topic complements our last full committee hearing, which
explored another nascent technological field: augmented
reality.
I'm excited by this topic, because AI has the potential to
catapult the United States economy and competitiveness in both
the near-and the long-term future. AI presents promising
applications in the areas of healthcare, transportation, and
agriculture, among others. And I want to thank our witnesses
for sharing and highlighting some of those applications today.
The recent report and strategy on AI released by the White
House Office of Science and Technology Policy provide Congress
with important considerations to weigh as we think about what
the appropriate role of government is in this promising field
so that we ensure that the United States remains the preeminent
place for AI in the global economy. And so, I appreciate,
again, the witnesses sharing your insights about what the
state-of-the-art is today and where the Nation's leading
experts see the applications, moving forward.
I wanted to direct a question, Dr. Horvitz, to you. You
mentioned, in your testimony, that new kinds of automation
present new attack surfaces for cyberattacks. And I wonder if
maybe you could elaborate on what some of those new
cybersecurity vulnerabilities might be.
Dr. Horvitz. Yes. Thanks for the interesting question and
framing.
The systems we build that are doing sophisticated tasks in
the world often are assembled out of multiple modules or
components that have to talk to one another, ending, often, in
cyberphysical or astrophysical activity or affecters, like car
steering wheels and braking and so on. Every single one of
those interfaces presents an opportunity to an attacker to
intervene and influence the behavior of a system.
There are also whole new categories of attack. I would be--
would have been surprised to learn, 15 years ago, that we
were--that the community was talking now about machine-learning
attacks. What's a machine-learning attack? The careful
injection into a learning system, in a sleuthy manner,
potentially, of data that will tend to build a classifier that
will do the wrong thing in certain cases. So, that just gives
you a sense or a taste for the very different kinds of
opportunities that are being presented by the systems we're
building now.
We often think about security in classical ways, with
verification models and encryption and so on. And these
techniques often will apply, but we have to be very careful, as
we build these systems, that we're taking--that we're covering
all ground and we're thinking through possibilities.
The Chairman. And, on the flip side of that, how can the
use of machine learning enhance security analysts' ability to
catch malicious hackers?
Dr. Horvitz. Yes, it's a great follow-on question, because
it's a yes/yes. I mean, there's--look, I mean, we have to be
cautious, because the--human beings and humans plus machines
can be very creative in how they attack, so there's a long tail
of possibilities we have to, sort of, account for. But, there
are some very, very promising angles with the use of artificial
intelligence and machine learning to detect anomalous patterns
of various kinds, with low false-positive rates. That's one of
the goals, is to do this well, where you don't call everything
strange, because people are always doing different things that
are safe, but that seem to be different over time and might
seem like a fraudulent event, for example.
So, I think there's a lot of promise. I know that--I'm very
excited about some recent projects that I reviewed at Microsoft
Research in this space. So, I think it's an exciting direction,
indeed.
The Chairman. Yes.
Dr. Moore. Speaking as someone who was at an Internet
search engine before I was at Carnegie Mellon, this is an area
where I would claim that Internet search companies are well
ahead of what you're seeing happening in the public sector.
There actually are some very good technologies out there for
doing machine learning versus machine-learning warfare. So,
it's an exciting area which I would like to see grow.
On the bright side, a recent DARPA challenge was about
using artificial intelligence to discover vulnerabilities
autonomously and using machine learning in other systems, which
sounds like a kind of frightening thing. But, (a) it is
actually important for our national defense that we have these
capabilities; and (b) it is one of the ways in which we can
keep ourselves safe, by having our own AIs trying to break into
our own systems. So, this is another capability which just
wasn't there 2 years ago. Carnegie Mellon, University of
Michigan, and plenty other major computer science universities
are heavily involved now in using AIs to both try to break and
warn us about breakages in our own computer security systems.
The Chairman. Mr. Moore, just very quickly here because I'm
out of time, but could you build a little bit on your written
testimony about how the United States can win the AI talent
war? In other words, what are the best ways to sustain enough
AI talent at universities to conduct basic research and drive
innovation while also filling what is a growing demand for AI
jobs in the private sector?
Dr. Moore. I think this begins in middle school. The U.S.
Government can really help here if we just help kids in middle
school understand that one of the most important and
interesting things they can be doing with their lives right now
is to learn mathematics so that they can be building these
kinds of robots and systems in the future. This is something
which needs training. It's not that you need to be a genius.
You need to be trained in math from about the age of 13 or 14
onwards, understand that that is even cooler as a career move
than going to work in Hollywood. Once we've got the kids' minds
in the right place, we can bring them through the university
system, scale that up, and then we'll be in good shape.
What I don't want to do is keep us in our current
situation, where the talent crunch is being dealt with by this
massive bidding war for this small amount of talent. Because
that's not sustainable when the other continents are doing such
a good job of producing AI experts.
Mr. Brockman. I think one thing that's really important is
that we can continue to attract the best AI researchers in the
world by having an open basic research community that just
draws everyone in. It has been working. And I think that we can
grow that and strengthen that community.
The Chairman. Thank you.
Thank you, Mr. Chairman.
Senator Cruz. Thank you, Chairman Thune.
Senator Daines.
STATEMENT OF HON. STEVE DAINES,
U.S. SENATOR FROM MONTANA
Senator Daines. Thank you, Mr. Chairman. And thank you for
holding this hearing today. Very timely.
Before I came to the Senate, I used to have a legitimate
day job. I was in the technology sector for about a dozen
years. And our company had several patents for AI design. We
won a national award for AI innovation. That was back in the
early 2000s. So, we're not talking about something that's new.
It has been around for quite some time.
In the Senate, we often talk about what we need to do to
ensure the U.S. maintains leadership, looking at global
competitiveness and innovation technology, whether it's
broadband, smart cities, medical research. So, I'd like to
start my questioning with Mr. Brockman.
Could you expand your testimony about what other countries
are doing, in terms of encouraging AI? And a follow-on there
would be, What do you see some of the competitive disadvantages
that we face right now in our country as it relates to ensuring
that we be--maintain global leadership in this important area?
Mr. Brockman. So, in other countries, I think there's a
mix. And so for example, you see companies like China's Baidu,
who, you know, want to scale up their investment in this field.
And the way that they've been doing it is that they've actually
opened a U.S.-based research institution, and have joined our
research community and are publishing and kind of following our
lead. With South Korea, I think that sort of around the same
time as the Alpha-Go match, that they announced that they were
going to make this billion-dollar investment into AI. And
Canada recently has been talking about that they're starting to
increase their national funding. And so that's the flavor that
you're seeing--both companies and the governments stepping up
their investments and trying to make research breakthroughts--
because I think everyone sees there's so much activity
happening.
And, I'm sorry, I actually missed the second part of the
question.
Senator Daines. Well, just looking at--what do you see as
some of the headwinds relates to create competitive
disadvantage for our country?
Mr. Brockman. I see. I think that the thing we should be
aware of--and so, there's a stat mentioned about the number of
Chinese AI papers that are published. And I think that that's
actually a true fact, but it's not necessarily the most
important fact. The most important fact is, Where do the
fundamental breakthroughs come from? Because the thing that
happens is, if you are the one who creates the underlying
technology, it's like discovering electricity. You're the one
who understands everything that went into that, and the papers
that get published those are in your language. You really get
to set the culture for how people build on top of it because
you're probably the one who published the underlying code that
people are using. And so, I think that the thing that we need
to watch is the question of--for the actual fundamental
advances, the capabilities that we just did not have before,
but that we have now, where do those come from? And, as long as
that's us, then I think we're in really good shape. And so, I
think that we need to continue to make sure that that's the
case.
Senator Daines. I want to go back to the point that was
brought up earlier on the cyberthreat. In 28 years in the
private sector, I never received a letter from the human
resource department that said that my information had been
hacked, until I became a Federal employee, was elected to the
U.S. Congress. And I, like millions of other Federal employees,
got a letter from OPM talking about the fact that my
information had been hacked. I spend a lot of time working and
chatting with some of these very innovative, smaller tech
companies that are doing some amazing things as it relates to
advancing at the speed of business, relates to protecting our
assets. I am concerned--and you mentioned the fact that the
Federal Government can lag, is not always leading in that area.
And I know it's frustrating, because we have solutions here. We
can't sometimes penetrate our own firewall, figuratively
speaking, as relates to trying to get--front here, to get our
government to move at the speed of business. Because I know
when we were--when I was in the cloud computing business, we
always--you wanted to make sure you were never on the front
page of The Wall Street Journal because of a hack. And what
that does to valuation of companies has been very obvious in
the last few years.
So, what do we need to do to ensure that the best
technology, as it's moving so fast right now, is in the hands
of our best people who are in the Federal Government? This is
not a critique on the people that work in the Federal
Government. This is oftentimes the barriers we look up here to
ensure that we're protecting our national assets. Who'd like to
answer that one?
Mr. Brockman. So--if I may--so, I've actually been
extremely impressed with the work that the USDS and OSTP have
been doing to solve problems like this. I think it really
starts with getting the best technologists in the door and
then, secondly, giving them the power and empowering them to
make changes within the Federal Government. And so, I think
that it really starts with the people, making sure that we're
attracting them and making sure that the structures within
government exist. And I think that, as long as there's an
attitude that's receptive within the agencies or wherever you
want to upgrade, I think that that's the best way to get this
to happen.
Dr. Chien. I'd like to jump in here, also. I think one of
the key things is--for the government to be at the forefront,
or at least participating in the forefront of technology, there
has to be an active interchange in what I would call a vibrant
ecosystem that includes multiple kinds of institutions. And I'm
very happy to say, in the AI and space arena, there's a large
amount of interplay between the commercial sector, between the
government sector, between small companies. It seems every week
there's another company being started up to do real time space
imaging of the Earth for business intelligence. I think that
all of this is indicative that there's a good structure with
this interchange of information. And I think that's the key to
making sure that the government stays in the right location and
able to understand and be smart in how it uses this technology.
Senator Daines. All right. Thank you. I'm out of time.
Thank you, Mr. Chairman.
Senator Cruz. Thank you, Senator Daines.
I'd like to thank each of our witnesses for coming to this
hearing, which I think was informative and productive and will
be just the beginning of what I expect to be an ongoing
conversation about how to deal with both the challenges and
opportunities that artificial intelligence presents.
The hearing record will remain open for 2 weeks. During
this time, Senators are asked to submit any questions they
might have for the record. And, upon receipt, the witnesses are
requested to submit their written answers to the Committee as
soon as possible.
Thank you, again, to the witnesses.
And this hearing is adjourned.
[Whereupon, at 3:57 p.m., the hearing was adjourned.]
A P P E N D I X
November 30, 2016
Hon. Ted Cruz, Chairman,
Hon. Gary Peters, Ranking Member,
U.S. Senate Committee on Commerce, Science, and Transportation,
Subcommittee on Space, Science, and Competitiveness,
Washington, DC.
RE: Hearing on ``The Dawn of Artificial Intelligence''
Dear Chairman Cruz and Ranking Member Peters:
We write to you regarding the upcoming hearing on ``The Dawn of
Artificial Intelligence.'' \1\ We appreciate your interest in this
topic. Artificial Intelligence implicates a wide range of economic,
social, and political issues in the United States. As an organization
now focused on the impact of Artificial Intelligence on American
society, we submit this statement and ask that it be entered into the
hearing record.
---------------------------------------------------------------------------
\1\ U.S. Senate Commerce, Science and Transportation Committee,
Subcommittee on Space, Science, and Competitiveness, ``The Dawn of
Artificial Intelligence,'' (Nov. 30, 2016), http://
www.commerce.senate.gov/public/index.cfm/hearings?ID=042DC718-9250-
44C0-9BFE-E0371AF
AEBAB
---------------------------------------------------------------------------
The Electronic Privacy Information Center (``EPIC'') is a public
interest research center established more than twenty years ago to
focus public attention on emerging civil liberties issues. In recent
years, EPIC has opposed government use of ``risk-based'' profiling,\2\
brought attention to the use of proprietary techniques for criminal
justice determinations, and litigated several cases on the front lines
of AI. In 2014, EPIC sued the U.S. Customs and Border Protection under
the Freedom of Information Act (``FOIA'') for documents about the use
of secret, tools to assign ``risk assessments'' to U.S. citizens \3\
EPIC also sued the Department of Homeland Security under the FPOA
seeking documents related to a program that assesses ``physiological
and behavioral signals'' to determine the probability that an
individual might commit a crime.\4\ Recently, EPIC appealed a Federal
Aviation Administration final order for failing to establish privacy
rules for commercial drones.\5\
---------------------------------------------------------------------------
\2\ EPIC et al., Comments Urging the Department of Homeland
Security To (A) Suspend the ``Automated Targeting System'' As Applied
To Individuals, Or In the Alternative, (B) Fully Apply All Privacy Act
Safeguards To Any Person Subject To the Automated Targeting System
(Dec. 4, 2006), available at http://epic.org/privacy/pdf/
ats_comments.pdf; EPIC, Comments on Automated Targeting System Notice
of Privacy Act System of Records and Notice of Proposed Rulemaking,
Docket Nos. DHS-2007-0042 and DHS-2007-0043 (Sept. 5, 2007), available
at http://epic.org/privacy/travel/ats/epic_090507.pdf. See also,
Automated Targeting System, EPIC, https://epic.org/privacy/travel/ats/.
\3\ EPIC, EPIC v. CBP (Analytical Framework for Intelligence),
https://epic.org/foia/dhs/cbp/afi/
\4\ EPIC, EPIC v. DHS--FAST Program, https://epic.org/foia/dhs/
fast/. See also the film Minority Report (2002)
\5\ EPIC, EPIC v. FAA, https://epic.org/privacy/litigation/apa/faa/
drones/.
---------------------------------------------------------------------------
EPIC has come to the conclusion that one of the primary public
policy goals for AI must be ``Algorithmic Transparency.'' \6\
---------------------------------------------------------------------------
\6\ EPIC, Algorithmic Transparency, https://epic.org/algorithmic-
transparency/ (last visited Nov. 29, 2016). The web page contains an
extensive collection of articles and commentaries by members of the
EPIC Advisory Board, leading experts in law, technology, and public
policy. More information about the EPIC Advisory Board is available at
https://www.epic.org/epic/advisory_board.html.
---------------------------------------------------------------------------
The Challenge of AI
There is understandable enthusiasm about new techniques that
promise medical breakthroughs, more efficient services, and new
scientific outcomes. But there is also reason for caution. Computer
scientist Joseph Weizenbaum famously illustrated the limitations of AI
in the 1960s with the development of the Eliza program. The program
extracted key phrases and mimicked human dialogue in the manner of non-
directional psychotherapy. The user might enter, ``I do not feel well
today,'' to which the program would respond, ``Why do you not feel well
today?'' Weizenbaum later argued in Computer Power and Human Reason
that computers would likely gain enormous computational power but
should not replace people because they lack such human qualities and
compassion and wisdom.\7\
---------------------------------------------------------------------------
\7\ Joseph Weizenbaum, Computer Power and Human Reason: From
Judgment to Calculation (1976).
---------------------------------------------------------------------------
We face a similar reality today.
The Need for Algorithmic Transparency
Democratic governance is built on principles of procedural fairness
and transparency. And accountability is key to decision making. We must
know the basis of decisions, whether right or wrong. But as decisions
are automated, and we increasingly delegate decisionmaking to
techniques we do not fully understand, processes become more opaque and
less accountable. It is therefore imperative that algorithmic process
be open, provable, and accountable. Arguments that algorithmic
transparency is impossible or ``too complex'' are not reassuring. We
must commit to this goal.
It is becoming increasingly clear that Congress must regulate AI to
ensure accountability and transparency:
Algorithms are often used to make adverse decisions about
people. Algorithms deny people educational opportunities,
employment, housing, insurance, and credit.\8\ Many of these
decisions are entirely opaque, leaving individuals to wonder
whether the decisions were accurate, fair, or even about them.
---------------------------------------------------------------------------
\8\ Danielle Keats Citron & Frank Pasquale, The Scored Society: Due
Process for Automated Predictions, 89 Wash. L. Rev. 1 (2014).
Secret algorithms are deployed in the criminal justice
system to assess forensic evidence, determine sentences, to
even decide guilt or innocence.\9\ Several states use
proprietary commercial systems, not subject to open government
laws, to determine guilt or innocence. The Model Penal Code
recommends the implementation of recidivism-based actuarial
instruments in sentencing guidelines.\10\ But these systems,
which defendants have no way to challenge are racially biased,
unaccountable, and unreliable for forecasting violent
crime.\11\
---------------------------------------------------------------------------
\9\ EPIC, Algorithms in the Criminal Justice System, https://
epic.org/algorithmic-transparency/crim-justice/ (last visited Nov. 29,
2016).
\10\ Model Penal Code: Sentencing Sec. 6B.09 (Am. Law. Inst.,
Tentative Draft No. 2, 2011).
\11\ See Julia Angwin et al., Machine Bias, ProPublica (May 23,
2016), https://www.propublica.org/article/machine-bias-risk-
assessments-in-criminal-sentencing.
Algorithms are used for social control. China's Communist
Party is deploying a ``social credit'' system that assigns to
each person government-determined favorability rating.
``Infractions such as fare cheating, jaywalking, and violating
family-planning rules'' would affect a person's rating.\12\ Low
ratings are also assigned to those who frequent disfavored
websites or socialize with others who have low ratings.
Citizens with low ratings will have trouble getting loans or
government services. Citizens with high rating, assigned by the
government, receive preferential treatment across a wide range
of programs and activities.
---------------------------------------------------------------------------
\12\ Josh Chin & Gillian Wong, China's New Tool for Social Control:
A Credit Rating for Everything, Wall Street J., Nov. 28, 2016, http://
www.wsj.com/articles/chinas-new-tool-for-social-control-a-credit-
rating-for-everything-1480351590
In the United States, U.S. Customs and Border Protection has
used secret analytic tools to assign ``risk assessments'' to
U.S. travelers.\13\ These risk assessments, assigned by the
U.S. Government to U.S. citizens, raise fundamental questions
about government accountability, due process, and fairness.
They may also be taking us closer to the Chinese system of
social control through AI.
---------------------------------------------------------------------------
\13\ EPIC, EPIC v. CBP (Analytical Framework for Intelligence),
https://epic.org/foia/dhs/cbp/afi/ (last visited Nov. 29, 2016).
EPIC believes that ``Algorithmic Transparency'' must be a
fundamental principle for all AI-related work.\14\ The phrase has both
literal and figurative dimensions. In the literal sense, it is often
necessary to determine the precise factors that contribute to a
decision. If, for example, a government agency considers a factor such
as race, gender, or religion to produce an adverse decision, then the
decision-making process should be subject to scrutiny and the relevant
factors identified.
---------------------------------------------------------------------------
\14\ EPIC, At UNESCO, Rotenberg Argues for Algorithmic Transparency
(Dec. 8, 2015), https://epic.org/2015/12/at-unesco-epics-rotenberg-
argu.html.
---------------------------------------------------------------------------
Some have argued that algorithmic transparency is simply
impossible, given the complexity and fluidity of modern processes. But
if that is true, there must be some way to recapture the purpose of
transparency without simply relying on testing inputs and outputs. We
have seen recently that it is almost trivial to design programs that
evade testing.\15\
---------------------------------------------------------------------------
\15\ See Jack Ewing, In '06 Slide Show, a Lesson in How VW Could
Cheat, N.Y. Times, Apr. 27, 2016, at A1.
---------------------------------------------------------------------------
In the formulation of European data protection law, which follows
from the U.S. Privacy Act of 1974, individuals have a right to access
``the logic of the processing'' concerning their personal
information.\16\ That principle is reflected in the transparency of the
FICO score, which for many years remained a black box for consumers,
making determinations about credit worthiness without any information
provided to the customers about how to improve the score.\17\
---------------------------------------------------------------------------
\16\ Directive 95/46/EC--The Data Protection Directive, art 15 (1),
1995, http://www.data
protection.ie/docs/EU-Directive-95-4-9EC-Chapter-2/93.htm.
\17\ See Hadley Malcom, Banks Compete on Free Credit Score Offers,
USA Today, Jan. 25, 2015, http://www.usatoday.com/story/money/2015/01/
25/banks-free-credit-scores/22011803/.
---------------------------------------------------------------------------
Building on this core belief in algorithmic transparency, EPIC has
urged public attention to four related principles to establish
accountability for AI systems:
``Stop Discrimination by Computer''
``End Secret Profiling''
``Open the Code''
``Bayesian Determinations are not Justice''
The phrases are slogans, but they are also intended to provoke a
policy debate and could provide the starting point for public policy
for AI. And we would encourage you to consider how these themes could
help frame future work by the Committee.
Amending Asimov's Laws of Robotics
In 1942, Isaac Asimov introduced the ``Three Laws of Robotics'':
1. A robot may not injure a human being or, through inaction, allow
a human being to come to harm.
2. A robot must obey the orders given it by human beings except
where such orders would conflict with the First Law.
3. A robot must protect its own existence as long as such protection
does not conflict with the First or Second Laws.\18\
---------------------------------------------------------------------------
\18\ Isaac Asimov, Runaround, Astounding Sci. Fiction, Mar. 1942,
at 94.
Asimov's Rules of Robotics remain a staple of science fiction and
ethical discourse.\19\ But they also emerged in a time when the focus
was on the physical ability of robots. In our present world, we have
become increasingly aware that it is the accountability of autonomous
devices that require the greater emphasis. For example, in seeking to
establish privacy safeguards prior to the deployment of commercial
drones in the United States,\20\ EPIC became aware that drones would
have an unprecedented ability to track and monitor individuals in
physical space while remaining almost entirely anonymous to humans.
Even the registration requirements established by the FAA would be of
little practical benefit to an individual confronted by a drone in
physical space.\21\ Does the drone belong to a hobbyist, a criminal, or
the police? Without basic identification information, it would be
impossible to make this determination, even as the drone was able to
determine the person's identity from a cell phone ID, facial
recognition, speech recognition, or gait.\22\
---------------------------------------------------------------------------
\19\ See, e.g., Michael Idato, Westworld's Producers Talk
Artificial Intelligence, Isaac Asimov's Legacy and Rebooting a
Cinematic Masterpiece for TV, Sydney Morning Herald, Sept. 29, 2016,
http://www.smh.com.au/entertainment/tv-and-radio/westworlds-producers-
talk-artificial-intelligence-asimovs-legacy-and-rebooting-a-cinematic-
masterpiece-for-tv-20160923-grn2yb.html; George Dvorsky, Why Asimov's
Three Laws of Robotics Can't Protect Us, Gizmodo (Mar. 28, 2014),
http://io9.gizmodo.com/why-asimovs-three-laws-of-robotics-cant-protect-
us-1553665410; TV Tropes, Three-Laws Compliant, http://tvtropes.org/
pmwiki/pmwiki.php/Main/ThreeLaws
Compliant (last visited Nov. 29, 2016).
\20\ EPIC, EPIC v. FAA, https://epic.org/privacy/litigation/apa/
faa/drones/ (last visited Nov. 29, 2016).
\21\ Operation and Certification of Small Unmanned Aircraft
Systems, 81 Fed. Reg. 42,064 (June 28, 2016) (to be codified at 14 CFR
Parts 21, 43, 61, 91, 101, 107, 119, 133, and 183).
\22\ See, e.g., Jim Giles, Cameras Know You by Your Walk, New
Scientist, Sept. 19, 2012, https://www.newscientist.com/article/
mg21528835-600-cameras-know-you-by-your-walk/.
---------------------------------------------------------------------------
This asymmetry poses a real threat. Along with the growing opacity
of automated decision-making, it is the reason we have urged two
amendments to Asimov's Laws of Robotics:
A robot must always reveal the basis of its decision
A robot must always reveal its actual identity
These insights also may be useful to the Committee as it explores
the implications of Artificial Intelligence.
Conclusion
The continued deployment of AI-based systems raises profound issues
for democratic countries. As Professor Frank Pasquale has said:
Black box services are often wondrous to behold, but our black
box society has become dangerously unstable, unfair, and
unproductive. Neither New York quants nor California engineers
can deliver a sound economy or a secure society. Those are the
tasks of a citizenry, which can perform its job only as well as
it understands the stakes.\23\
---------------------------------------------------------------------------
\23\ Frank Pasquale, The Black Box Society: The Secret Algorithms
that Control Money and Information 218 (Harvard University Press 2015).
We appreciate your interest in this subject and urge the Committee
to undertake a comprehensive review of this critical topic.
Sincerely,
Marc Rotenberg,
EPIC President.
James Graves,
EPIC Law and Technology Fellow.
Enclosures
EPIC, ``Algorithmic Transparency''
cc: The Honorable John Thune, Chairman, Senate Commerce Committee
The Honorable Bill Nelson, Ranking Member, Senate Commerce Committee
______
Algorithmic Transparency: End Secret Profiling
Disclose the basis of automated decisionmaking
Top News
EPIC Urges Massachusetts High Court to Protect E-mail
Privacy: EPIC has filed an amicus brief in the Massachusetts
Supreme Judicial Court regarding e-mail privacy. At issue is
Google's scanning of the e-mail of non-Gmail users. EPIC argued
that this is prohibited by the Massachusetts Wiretap Act. EPIC
described Google's complex scanning and analysis of private
communications, concluding that it was far more invasive than
the interception of a telephone communications, prohibited by
state law. A Federal court in California recently ruled that
non-Gmail users may sue Google for violation of the state
wiretap law. EPIC has filed many amicus briefs in Federal and
state courts and participated in the successful litigation of a
cellphone privacy case before the Massachusetts Judicial Court.
The EPIC State Policy Project is based in Somerville,
Massachusetts. (Oct. 24, 2016)
EPIC Promotes ``Algorithmic Transparency'' at Annual Meeting
of Privacy Commissioners: Speaking at the 38th International
Conference of the Data Protection and Privacy Commissioners in
Marrakech, EPIC President Marc Rotenberg highlighted EPIC's
recent work on algorithmic transparency and also proposed two
amendments to Asimov's Rules of Robotics. Rotenberg cautioned
that autonomous devices, such as drones, were gaining the
rights of privacy--control over identity and secrecy of
thought--that should be available only for people. Rotenberg
also highlighted EPIC's recent publication ``Privacy in the
Modern Age'', the Data Protection 2016 campaign, and the
various publications available at the EPIC Bookstore. The 2017
Privacy Commissioners conference will be held in Hong Kong.
(Oct. 20, 2016)
White House Report on the Future of Artificial Intelligence
In May 2016, the White House announced a series of workshops and a
working group devoted to studying the benefits and risks of AI. The
announcement recognized the ``array of considerations'' raised by AI,
including those ``in privacy, security, regulation, [and] law.'' The
White House established a Subcommittee on Machine Learning and
Artificial Intelligence within the National Science and Technology
Council.
Over the next three months, the White House co-hosted a series of
four workshops on AI:
Legal and Governance Implications of Artificial
Intelligence, May 24, 2016, Seattle, WA
Artificial Intelligence for Social Good, June 7, 2016, in
Washington, D.C.
Safety and Control for Artificial Intelligence, June 28,
2016, in Pittsburgh, PA
The Social and Economic Implications of Artificial
Intelligence Technologies in the Near-Term, July 7, 2016, in
New York City
EPIC Advisory Board members Jack Balkin, Danah Boyd, Ryan Calo,
Danielle Citron, Ed Felten, Ian Kerr, Helen Nissenbaum, Frank Pasquale,
and Latanya Sweeney each participated in one or more of the workshops.
The White House Office of Science and Technology issued a Request
for Information in June 2016 soliciting public input on the subject of
AI. The RFI indicated that the White House was particularly interested
in ``the legal and governance implications of AI,'' ``the safety and
control issues for AI,'' and ``the social and economic implications of
AI,'' among other issues. The White House received 161 responses.
On October 12, 2016, The White House announced two reports on the
impact of Artificial Intelligence on the U.S. economy and related
policy concerns: Preparing for the Future of Artificial Intelligence
and National Artificial Intelligence Research and Development Strategic
Plan.
Preparing for the Future of Artificial Intelligence surveys the
current state of AI, its applications, and emerging challenges for
society and public policy. As Deputy U.S. Chief Technology Officer and
EPIC Advisory Board member Ed Felten writes for the White House blog,
the report discusses ``how to adapt regulations that affect AI
technologies, such as automated vehicles, in a way that encourages
innovation while protecting the public'' and ``how to ensure that AI
applications are fair, safe, and governable.'' The report concludes
that ``practitioners must ensure that AI-enabled systems are
governable; that they are open, transparent, and understandable; that
they can work effectively with people; and that their operation will
remain consistent with human values and aspirations.''
The companion report, National Artificial Intelligence Research and
Development Strategic Plan, proposes a strategic plan for Federally-
funded research and development in AI. The plan identifies seven
priorities for federally-funded AI research, including strategies to
``understand and address the ethical, legal, and societal implications
of AI'' and ``ensure the safety and security of AI systems.''
The day after the reports were released, the White House held a
Frontiers Conference co-hosted by Carnegie Mellon University and the
University of Pittsburgh. Also in October, Wired magazine published an
interview with President Obama and EPIC Advisory Board member Joi Ito.
EPIC's Interest
EPIC has promoted Algorithmic Transparency for many years and is
has litigated several cases on the front lines of AI. EPIC's cases
include:
EPIC v. FAA, which EPIC filed against the Federal Aviation
Administration for failing to establish privacy rules for
commercial drones
EPIC v. CPB, in which EPIC successfully sued U.S. Customs
and Border Protection for documents relating to its use of
secret, analytic tools to assign ''risk assessments'' to
travelers
EPIC v. DHS, to compel the Department of Homeland Security
to produce documents related to a program that assesses
``physiological and behavioral signals'' to determine the
probability that an individual might commit a crime.
EPIC has also filed amicus briefs supporting in Cahen v. Toyota
that discusses the risks inherent in connected cars and has filed
comments on issues of big data and algorithmic transparency.
EPIC also has a strong interest in algorithmic transparency in
criminal justice. Secrecy of the algorithms used to determine guilt or
innocence undermines faith in the criminal justice system. In support
of algorithmic transparency, EPIC submitted FOIA requests to six states
to obtain the source code of ``TrueAllele,'' a software product used in
DNA forensic analysis. According to news reports, law enforcement
officials use TrueAllele test results to establish guilt, but
individuals accused of crimes are denied access to the source code that
produces the results.
Resources
Kate Crawford and Ryan Calo, There is a blind spot in AI
research (October 13, 2016).
We Robot 2017
We Robot 2016
Ryan Calo, A. Michael Froomkin, and Ian Kerr, Robot Law
(Edward Elgar 2016)
EPIC: Algorithms in the Criminal Justice System
Alessandro Acquisti, Why Privacy Matters (Jun 2013)
Alessandro Acquisti, Ralph Gross, Fred Stutzman, Faces of
Facebook: Privacy in the Age of Augmented Reality (Aug. 4,
2011)
Alessandro Acquisti, Price Discrimination, Privacy
Technologies, and User Acceptance (2006)
Steven Aftergood, ``Secret Law and the Threat to Democratic
Government,'' Testimony before the Subcommittee on the
Constitution of the Committee on the Judiciary, U.S. Senate
(Apr. 30, 2008)
Phil Agre, Your Face Is Not a Bar Code: Arguments Against
Automatic Face Recognition in Public Places
Ross Anderson, The Collection, Linking and Use of Data in
Biomedical Research and Health Care: Ethical Issues (Feb. 2015)
James Bamford, The Shadow Factory: The NSA from 9/11 to the
Eavesdropping on America (2009)
Grayson Barber, How Transparency Protects Privacy in
Government Records (May 2011) (with Frank L. Corrado)
Colin Bennett, Transparent Lives: Surveillance in Canada
Danah Boyd, Networked Privacy (2012)
David Burnham, The Rise of the Computer State (1983)
Julie E. Cohen, Power/play: Discussion of Configuring the
Networked Self, 6 Jerusalem Rev. Legal Stud. 137-149 (2012)
Julie E. Cohen, Julie E. Cohen, Configuring the Networked
Self: Law, Code, and the Play of Everyday Practice (New Haven,
Conn.: Yale University Press 2012)
Julie E. Cohen, Privacy, Visibility, Transparency, and
Exposure (2008)
Danielle Keats CItron & Frank Pasquale, The Scored Society:
Due Process for Automated Predictions, 89 Washington Law Review
(2014) 1
Cynthia Dwork & Aaron Roth, The Algorithmic Foundations of
Differential Privacy, 9(4) Theoretical Computer Science (2014)
211
David J. Farber & Gerald R Faulhaber, The Open Internet: A
Consumer-Centric Framework
Ed Felten, Algorithms can be more accountable than people,
Freedom to Tinker
Ed Felten, David G Robinson, Harlan Yu & William P Zeller,
Government Data and the Invisible Hand, 11 Yale Journal of Law
& Technology (2009) 160
Ed Felten, CITP Web Privacy and Transparency Conference
Panel 2
A Michael Froomkin, The Death of Privacy, 52 Stanford Law
Review (2000) 1461
Urs Gasser et. al., ed, Internet Monitor 2014; Reflections
on the Digital World Berkman Center for Internet and Society
Urs Gasser, Regulating Search Engines: Taking Stock and
Looking Ahead, 9 YALE J.L. & TECH. 124 (2006)
Jeff Jonas, Using Transparency as a Mask, (Aug. 4, 2010)
Jeff Jonas & Ann Cavoukian, Privacy by Design in the Age of
Big Data (Jun. 8, 2010)
Ian Kerr, Privacy, Identity and Anonymity (Sep. 1, 2011)
Dr Ian Kerr Prediction, Presumption, Preemption: The Path of
Law After the Computational Turn (Jul. 30, 2011)
Rebeca MacKinnon, Where is Microsoft Bing's Transparency
Report? The Guardian (Feb. 14, 2014)
Frank Pasquale, The Black Box Society: The Secret Algorithms
That Control Money and Information (Jan. 5, 2015)
Frank Pasquale, The Scored Society: Due Process for
Automated Predictions, 89 Washington Law Review 1 (2014) (with
Danielle Citron)
Frank Pasquale, Restoring Transparency to Automated
Authority, 9 Journal on Telecommunications & High Technology
Law 235 (2011)
Frank Pasquale, Beyond Innovation and Competition: The Need
for Qualified Transparency in Internet Intermediaries, 104
Northwestern University Law Review 105 (2010)
Frank Pasquale, Internet Nondiscrimination Principles:
Commercial Ethics for Carriers and Search Engines, 2008
University of Chicago Legal Forum 263 (2008)
Bruce Schneier, Accountable Algorithms (Sep. 21, 2012)
Latanya Sweeney, Privacy Enhanced Linking, ACM SIGKDD
Explorations 7(2) (Dec. 2005)
Tim Wu, TNR Debate: Too Much Transparency? New Republic
[all]