[House Hearing, 115 Congress]
[From the U.S. Government Publishing Office]
GAME CHANGERS: ARTIFICIAL INTELLIGENCE PART I
=======================================================================
HEARING
BEFORE THE
SUBCOMMITTEE ON
INFORMATION TECHNOLOGY
OF THE
COMMITTEE ON OVERSIGHT
AND GOVERNMENT REFORM
HOUSE OF REPRESENTATIVES
ONE HUNDRED FIFTEENTH CONGRESS
SECOND SESSION
__________
FEBRUARY 14, 2018
__________
Serial No. 115-65
__________
Printed for the use of the Committee on Oversight and Government Reform
[GRAPHIC(S) NOT AVAILABLE IN TIFF FORMAT]
Available via the World Wide Web: http://www.fdsys.gov
http://oversight.house.gov
______
U.S. GOVERNMENT PUBLISHING OFFICE
30-296 PDF WASHINGTON : 2018
Committee on Oversight and Government Reform
Trey Gowdy, South Carolina, Chairman
John J. Duncan, Jr., Tennessee Elijah E. Cummings, Maryland,
Darrell E. Issa, California Ranking Minority Member
Jim Jordan, Ohio Carolyn B. Maloney, New York
Mark Sanford, South Carolina Eleanor Holmes Norton, District of
Justin Amash, Michigan Columbia
Paul A. Gosar, Arizona Wm. Lacy Clay, Missouri
Scott DesJarlais, Tennessee Stephen F. Lynch, Massachusetts
Blake Farenthold, Texas Jim Cooper, Tennessee
Virginia Foxx, North Carolina Gerald E. Connolly, Virginia
Thomas Massie, Kentucky Robin L. Kelly, Illinois
Mark Meadows, North Carolina Brenda L. Lawrence, Michigan
Ron DeSantis, Florida Bonnie Watson Coleman, New Jersey
Dennis A. Ross, Florida Stacey E. Plaskett, Virgin Islands
Mark Walker, North Carolina Val Butler Demings, Florida
Rod Blum, Iowa Raja Krishnamoorthi, Illinois
Jody B. Hice, Georgia Jamie Raskin, Maryland
Steve Russell, Oklahoma Peter Welch, Vermont
Glenn Grothman, Wisconsin Matt Cartwright, Pennsylvania
Will Hurd, Texas Mark DeSaulnier, California
Gary J. Palmer, Alabama Jimmy Gomez,California
James Comer, Kentucky
Paul Mitchell, Michigan
Greg Gianforte, Montana
Sheria Clarke, Staff Director
William McKenna, General Counsel
Troy Stock, Technology Subcommittee Staff Director
Sarah Moxley, Senior Professional Member
Sharon Casey, Deputy Chief Clerk
David Rapallo, Minority Staff Director
------
Subcommittee on Information Technology
Will Hurd, Texas, Chairman
Paul Mitchell, Michigan, Vice Chair Robin L. Kelly, Illinois, Ranking
Darrell E. Issa, California Minority Member
Justin Amash, Michigan Jamie Raskin, Maryland
Blake Farenthold, Texas Stephen F. Lynch, Massachusetts
Steve Russell, Oklahoma Gerald E. Connolly, Virginia
Greg Gianforte, Montana Raja Krishnamoorthi, Illinois
C O N T E N T S
----------
Page
Hearing held on February 14, 2018................................ 1
WITNESSES
Dr. Amir Khosrowshahi, Vice President and Chief Technology
Officer, Artificial Intelligence Products Group, Intel
Oral Statement............................................... 4
Written Statement............................................ 7
Dr. Charles Isbell, Executive Associate Dean and Professor,
College of Computing, Georgia Institute of Technology
Oral Statement............................................... 22
Written Statement............................................ 25
Dr. Oren Etzioni, Chief Executive Officer, Allen Institute for
Artificial Intelligence
Oral Statement............................................... 31
Written Statement............................................ 33
Dr. Ian Buck, Vice President and General Manager, Tesla Data
Center Business, NVIDIA
Oral Statement............................................... 45
Written Statement............................................ 47
GAME CHANGERS: ARTIFICIAL INTELLIGENCE PART I
----------
Wednesday, February 14, 2018
House of Representatives,
Subcommittee on Information Technology,
Committee on Oversight and Government Reform,
Washington, D.C.
The subcommittee met, pursuant to call, at 2:23 p.m., in
Room 2154, Rayburn House Office Building, Hon. Will Hurd
[chairman of the subcommittee] presiding.
Present: Representatives Hurd, Amash, Kelly, Lynch,
Connolly, and Krishnamoorthi.
Also Present: Representative Massie.
Mr. Hurd. The Subcommittee on Information Technology will
come to order. And, without objection, the chair is authorized
to declare a recess at any time.
Welcome to the first hearing in a series of hearings on
artificial intelligence. This series is an opportunity for the
subcommittee to take a deep dive into artificial intelligence.
And today's hearing is an opportunity to increase Congress'
understanding of artificial intelligence, including its
development, uses, and the potential challenges and advantages
of government adoption of artificial intelligence.
We have four experts on the matter whom I look forward to
hearing from today. And in the next hearing we do, in March, I
believe, we will hear from government agencies about how they
are or should be adopting artificial intelligence into their
operations, how they will use AI to spend taxpayer dollars
wisely and make each individual's interactions with the
government more efficient, effective, and secure.
It is important that we understand both the risks and
rewards of artificial intelligence. And in the third hearing,
in April, we will discuss the appropriate roles of both the
public and private sectors as artificial intelligence matures.
Artificial intelligence is a technology that transcends
borders. We have allies and adversaries, both nation-states and
individual hackers, who are pursuing artificial intelligence
with all they have, because dominance in artificial
intelligence is a guaranteed leg up in the realm of geopolitics
and economics.
At the end of this series, it is my goal to ensure that we
have a clear idea of what it takes for the United States to
remain the world leader when it comes to artificial
intelligence. Thoughtful engagement by legislators is key to
this goal, and I believe that this committee will be leaders on
this topic.
So what is artificial intelligence? Hollywood's portrayal
of artificial intelligence is not accurate. Instead, many of us
are already using it every single day, from song
recommendations in Spotify to digital assistants that tell us
the weather.
And while these consumer applications are important, I am
most excited about the possibility of using artificial
intelligence in the government to defend our infrastructure and
have better decisionmaking because of the analytics that
artificial intelligence can run.
In an environment of tightening resources, artificial
intelligence can help us do more for less money and help to
provide better citizen-facing services.
I thank the witnesses for being here today and look forward
to hearing and learning from you so that we can all benefit
from the revolutionary opportunities AI provides us.
As always, I am honored to be exploring these issues in a
bipartisan fashion, I think the IT Subcommittee is a leader on
doing things in a bipartisan way, with my friend and ranking
member, the Honorable Robin Kelly from the great State of
Illinois.
Ms. Kelly. Thank you. Welcome to the witnesses. Thank you,
Chairman Hurd, and welcome to all of our witnesses today, and
Happy Valentine's Day.
Artificial intelligence, or AI, has the capacity to improve
how society handles some of its most difficult challenges.
In medicine, the use of AI has the potential to save lives
and detect illnesses early. One MIT study found that using
machine-learning algorithms reduced human errors by 85 percent
when analyzing the cells of lung cancer patients. And earlier
this month, Wired magazine reported hospitals have now begun
testing software that can check the images of a person's eye
for signs of diabetic eye disease, a condition that if
diagnosed too late can result in vision lost.
In some communities around the country, self-driving cars
are already operating on the road and highways. That makes me
nervous. Investment by major car companies in self-driving cars
makes it increasingly likely that they will become the norm,
not the exception on our Nation's roads.
But there is a lot of uncertainty revolving around
artificial intelligence. AI is no longer the fantasy of science
fiction and is increasingly used in everyday life. As the use
of AI expands, it is critical that this powerful technology is
implemented in an inclusive, accessible, and transparent
manner.
In its most recent report on the future of AI, the National
Science and Technology Council issued a dire assessment of the
state of diversity within the AI industry. The NSTC found that
there was a, quote, ``lack of gender and racial diversity in
the AI workforce,'' and that this, quote, ``mirrors the lack of
diversity in the technology industry and the field of computer
science generally.'' According to the NSTC, in the field of AI
improving diversity, and I quote, ``is one of the most critical
and high priority challenges.''
The existing racial and gender gaps in the tech industry
add to the challenges the AI field faces. Although women
comprise approximately 18 percent of computer science graduates
in the Nation, only 11 percent of all computer science
engineers are female. African Americans and Hispanics account
for just 11 percent of all employees in the technology sector,
despite making up 27 percent of the total population in this
country.
Lack of AI workforce diversity can have real cost on
individuals' lives. The increasing use of AI to make
consequential decisions about people's lives is spreading at a
fast rate. Currently, AI systems are being used to make
decisions by banks about who should receive loans, by
government about whether someone is eligible for public
benefits, and by courts about whether a person should be set
free.
However, research has found considerable flaws and biases
can exist in the algorithms that support AI systems, calling
into question the accuracy of such systems and its potential
for unequal treatment of some Americans. For AI to be accurate,
it requires accurate data and learning sets to draw
conclusions. If the data provided is biased, the conclusions
will likely be biased. A diverse workforce will likely account
for this and use more diverse data and learning sets.
Within the industry, the use of black box algorithms are
exacerbating the problems of bias. Two years ago, ProPublica
investigated the use of computerized risk prediction tools that
were used by some judges in criminal sentencing and bail
hearings.
The investigation revealed that the algorithm the systems
relied upon to predict recidivism was not only inaccurate, but
biased against African Americans who were, quote, ``twice as
likely as Whites to be labeled a higher risk but not actually
reoffend.''
Judges were using misinformation derived from black box
software to make life-changing decisions on whether someone is
let free or receives a harsher sentence than appropriate.
Increasing the transparency of these programs and ensuring
a diverse workforce is engaged on developing AI will help
decrease bias and make software more inclusive. Increasing
diversity among the AI workforce helps avoid the negative
outcomes that can occur when AI development is concentrated
among certain groups of individuals, including the risk of
biases in AI systems.
As we move forward in this great age of technological
modernization, I will be focused on how the private sector,
Congress, and regulators can work together to ensure that AI
technologies continue to innovate successfully and socially
responsibly.
I want to thank our witnesses for testifying today and look
forward to hearing your thoughts on how we can achieve this
goal.
And, again, thank you, Mr. Chair.
Mr. Hurd. I recognize the distinguished gentleman from
Kentucky, Mr. Massie, is here. He is not a member of the
subcommittee, so I ask unanimous consent that he is able to
fully participate in this hearing. Without objection, so
ordered.
Now I am pleased to announce and introduce our witnesses.
Our first one, Dr. Amir Khosrowshahi, is vice president and
chief technology officer of the Artificial Intelligence
Products Group at Intel.
Welcome.
Dr. Charles Isbell is executive associate dean of the
College of Computing within the Georgia Institute of
Technology.
Dr. Oren Etzioni is the chief executive officer at the
Allen Institute for Artificial Intelligence.
And Dr. Ian Buck is vice president and general manager of
Accelerated Computing at NVIDIA.
Welcome to you all.
And pursuant to committee rules, all witnesses will be
sworn in before you testify. So please rise and raise your
right hand.
Do you solemnly swear or affirm that the testimony you are
about to give is the truth, the whole truth, and nothing but
the truth, so help you God?
Thank you.
Please let the record reflect that all witnesses answered
in the affirmative.
In order to allow time for discussion, please limit your
testimony to 5 minutes. Your entire written statement will be
made part of the record.
And as a reminder, the clock in front of you shows your
remaining time. The light will turn yellow when you have 30
seconds left, and when it turns red your time is up. And please
remember to also push the button to turn on your microphone
before speaking.
And now it is a pleasure to recognize Dr. Khosrowshahi for
your initial 5 minutes.
WITNESS STATEMENTS
STATEMENT OF AMIR KHOSROWSHAHI
Mr. Khosrowshahi. Good afternoon, Chairman Hurd, Ranking
Member Kelly, and members of the House Committee on Oversight
and Government Reform, Subcommittee on Information Technology.
My name is Amir Khosrowshahi, and I am the vice president
and chief technology officer of Intel Corporation's Artificial
Intelligence Products Group.
We're here today to discuss artificial intelligence, a term
that was an aspirational concept until recently. While
definitions of artificial intelligence vary, my work at Intel
focuses on applying machine-learning algorithms to real world
scenarios to offer benefits to people and organizations.
Thanks to technological advancements, AI is now emerging as
a fixture in our daily lives. For instance, speech recognition
features, recommendation engines, and bank fraud detection
systems all utilize AI.
These features make our lives more convenient, but AI
offers society so much more. For example, AI healthcare
solutions will revolutionize patient diagnosis and treatment.
Heart disease kills one in four people in the United
States. It is difficult for doctors to accurately diagnose
disease, because different conditions present similar symptoms.
That's why doctors mainly have had to rely on experience and
instinct to make diagnoses. More experienced doctors tend to
diagnose correctly three out of four times, those with less
experience, however, just half the time, as accurate as the
flipping of a coin. Patients suffer due to this information
gap.
Recently, researchers using AI accurately spotted the
difference between the two types of heart disease 9 out of 10
times. In this regard, AI democratizes expert diagnoses for
patients and doctors everywhere in the world.
AI is also contributing positively to agriculture. The
population is growing, and by 2050 we will need to produce at
least 50 percent more food to feed everyone. This will become
increasingly challenging as societies will need to produce more
food with less land to grow crops.
Thankfully, AI applications provide tools to improve crop
yields and quality, while also reducing consumption of
resources like water and fertilizer.
These are just a few examples of how AI is helping our
communities. However, as we continue to harness the benefits of
AI for societal good, governments will play a major role. We
are in the early days of innovation of a technology that can do
tremendous good. Governments should make certain to encourage
this innovation and they should be wary of regulation that will
stifle its growth.
At the Federal level, the United States Government can play
an important role in enabling the further development of AI
technology in a few ways.
First, since data fuels AI, the U.S. Government should
embrace open data policies. To realize AI's benefits,
researchers need to have access to large datasets. Some of the
most comprehensive datasets are currently owned by the Federal
Government. This data is a taxpayer-funded resource which, if
made accessible to the public, could be utilized by researchers
to train algorithms for future AI solutions.
The OPEN Government Data Act makes all nonsensitive U.S.
Government data freely available and accessible to the public.
Intel supports this bill and calls for its swift passage.
Second, the U.S. Government can help prepare an AI
workforce. Supporting universal STEM education is a start, but
Federal funding for basic scientific research at universities
by agencies like the National Science Foundation is important
to both train graduate-level scientists and contribute to our
scientific knowledge base.
Current Federal funding levels are not keeping pace with
the rest of the industrialized world. I encourage lawmakers to
consider the tremendous returns on investment to our economy
that funding science research produces.
In addition to developing the right talent to develop AI
solutions, governments will have to confront labor
displacement. AI's emergence will displace some workers, but
too little is known about the types of jobs and industries that
would be most affected.
Bills like H.R. 4829, the AI JOBS Act, help bridge that
information gap by calling for the Labor Department to study
the issue and to work with Congress on recommendations. Intel
supports this bill as well and encourages Congress to consider
it in committee.
AI promises many societal benefits, and government and
industry should work together to harness them, and also to set
up guidelines to encourage ethical deployment of AI and to
prevent it from being used in improper ways that could harm the
public.
I cannot stress enough how important it is that lawmakers
seize the opportunity to enable AI innovation. As U.S.
lawmakers consider what to do in response to the emergence of
AI, I encourage you to use a light touch. Legislating or
regulating AI too heavily will only serve to disadvantage
Americans, especially as governments around the world are
pouring resources into tapping into AI's potential.
Thank you again for the opportunity to testify today. The
government will play an important role in enabling us to
harness AI's benefits while preparing society to participate in
an AI-fueled economy. Determining whether or how existing legal
and public policy frameworks may need to be altered will be an
iterative process. Intel stands ready to be a resource as you
consider these issues.
Thank you.
[Prepared statement of Mr. Khosrowshahi follows:]
[GRAPHIC(S) NOT AVAILABLE IN TIFF FORMAT]
Mr. Hurd. Thank you, Dr. Khosrowshahi.
Dr. Isbell, you are now recognized for 5 minutes.
STATEMENT OF CHARLES ISBELL
Mr. Isbell. Chairman Hurd, Ranking Member Kelly, and
distinguished members of the subcommittee, my name is Dr.
Charles Isbell. I am a professor and executive associate dean
for the College of Computing at Georgia Tech. I would like to
thank you for the opportunity to appear before the
subcommittee.
As requested by the subcommittee, my testimony today will
focus on the potential for artificial intelligence and machine
learning to transform the world around us and how we might
collectively best respond to this potential.
There are many definitions of AI. My favorite one is that
it is the art and science of making computers act the way they
do in the movies. In the movies, computers are often
semimagical and anthropomorphic. They do things that if humans
did them, we would say they required intelligence.
As noted by the chairman, if that is AI, then we already
see AI in our everyday lives. We use the infrastructure of AI
to search more documents than any human could possibly read in
a lifetime, to find the answers to a staggering variety of
questions, often expressed literally as questions. We use that
same infrastructure to plan optimal routes for trips, even
altering our routes on the fly in the face of changes in
traffic.
We let computers finish our sentences, sometimes
facilitating a subtle shift from prediction of our behavior to
influence over our behavior. And we take advantage of these
services by using computers on our phones or home speakers to
interpret a wide variety of spoken commands.
All of this is made possible because AI systems are
fundamentally about computing and computing methods for
automated understanding and reasoning, especially ones that
leverage data to adapt their behavior over time.
That AI is really computing is an important point to
understand. What has enabled many of the advances in AI is the
stunning increase of computational power, combined with the
ubiquity of that computing.
That AI also leverages data is equally important. The same
advances in AI are also due, in large part, to the even more
stunning increase in the availability of data, again made
possible by ubiquity, in this case of the internet, social
media, and relatively inexpensive sensors, including cameras,
GPS, microphones, all embedded in devices we carry with us,
connected to computers that are, in turn, connected to one
another.
By leveraging computing and data, we are moving from robots
that assemble our cars to cars that almost drive themselves.
One can be skeptical, as I am, that we will in the near future
create AI that is as capable as humans are in performing a wide
variety of the sort of general tasks that humans grapple with
every day simultaneously. But it does seem that we are making
strong progress toward being able to solve a lot of very hard
individual tasks as well as humans.
We may not replace all 3 million truck drivers and taxi cab
drivers, nor all 3 million cashiers in the United States, but
we will increasingly replace many of them. We may soon trust
the x-ray machine itself to tell us whether we have a tumor as
much as we trust the doctor. We may not automate away
intelligence analysts, but AI will shape and change their
analysis.
So AI exists and is getting better. It is not the AI of
science fiction, neither benevolent intelligence working with
humans as we traverse the galaxy, nor malevolent AI that seeks
humanity's destruction. Nonetheless, we are living every day
with machines that make decisions that if humans made them we
would attribute to intelligence.
As noted by the ranking member, it is worth noting that
these machines are making decisions for humans and with humans.
Many AI researchers and practitioners are engaged in what we
might call interactive AI. The fundamental goal there is to
understand how to build intelligent agents that must live and
interact with large numbers of other intelligent agents, some
of whom may be human.
Progress towards this goal means that we can build
artificial systems that work with humans to accomplish tasks
more effectively, can respond more robustly to changes in the
environment, and can better coexist with humans as long-lived
partners.
But as with any partner, it is important that we understand
what our partner is doing and why. To make the most of this
emerging technology, we will need a more informed citizenry,
something we can accomplish by requiring that our AI partners
are more transparent on the one hand and that we are more savvy
on the other.
By transparency, I mean something relatively simple. An AI
algorithm should be inspectable. The kind of data the algorithm
uses to build its model should be available. It is useful to
know that your medical AI was trained to detect heart attacks
mostly in men.
The decisions that the system makes should be explainable
and understandable. In other words, as we deploy these
algorithms, each algorithm should be able to explain its output
and its decisions: This applicant was assigned higher risk
because is not only more useful, but is less prone to abuse
than just this applicant was assigned a higher risk.
To understand such machines, much less to create them, we
have to strive for everyone to not only be literate, but to be
compurate. That is, they must understand computing and
computational thinking and how it fits into problem-solving in
their everyday lives.
I am excited by these hearings. Advances in AI are central
to our economic and social future. The issues that are being
raised here are addressable and can be managed with thoughtful
support for robust funding and basic research in artificial
intelligence, as noted by my colleague, support for ubiquitous
and equitable computing education throughout the pipeline, in
K-12 and beyond, and the developing standards for the proper
use of intelligent systems.
I thank you very much for your time and attention today,
and I look forward to working with you in your efforts to
understand how we can best develop these technologies to create
a future where we are partners with intelligent machines.
Thank you.
[Prepared statement of Mr. Isbell follows:]
[GRAPHIC(S) NOT AVAILABLE IN TIFF FORMAT]
Mr. Hurd. Thank you, sir.
Dr. Etzioni, you are now up for 5 minutes.
STATEMENT OF OREN ETZIONI
Mr. Etzioni. Good afternoon, Chairman Hurd and Ranking
Member Kelly, distinguished members of the committee. Thank you
for the opportunity to speak with you today about the nature of
AI and the role of the Federal Government.
My name is Oren Etzioni. I am the CEO of the Allen
Institute for Artificial Intelligence, which is backed by Paul
Allen. We call ourselves AI2. Founded in 2014, AI2 is a
nonprofit research institute whose mission is to improve lives
by conducting high-impact research and engineering in the field
of AI for the common good.
The goal of my brief remarks today is to help demystify AI
and cut through a lot of the hype on the subject. And I'm
delighted to talk to you in particular, Chairman, with a
computer science degree. But it's really important to me to
make sure that my remarks are understandable by everybody and
that we don't confuse science fiction with the real science and
Hollywood and hype with what's actually going on.
What we do have are these very narrow systems that are
increasingly sophisticated, but they're also extremely
difficult to build. We need to work to increase the supply of
people who can do this. And that's going to be achieved through
increased diversity, but also through immigration.
And so, so many of us are immigrants to this country. At
AI2, we have 80 people who come from literally all over the
world, from Iran, from Israel, from India, et cetera, et
cetera. We need to continue to welcome these people so we can
continue to build these systems.
I have a number of thoughts, but I actually want to address
the issue that came up just in the conversation now about
transparency and bias and certainly the concerns that we have
about these database systems generating unfairness.
Obviously, we want the systems to be fair, and obviously,
we want them to be transparent. Unfortunately, it's not as easy
as it sounds. These are complex statistical models that are
ingesting enormous amounts of data, millions and billions of
examples, and generating conclusions.
So we have to be careful. And I think the phrase ``light
touch'' is a great one here. We have to be very careful that we
don't legislate transparency, but rather that we attempt to
build algorithms that are more favored, more desired, because
they're more transparent.
I think legislating transparency or trying to do that would
actually be a mistake, because ultimately consider the
following dilemma. Let's say you have a diagnostic system
that's highly transparent and 80 percent accurate. You've got
another diagnostic system that's making a decision about a key
treatment. It's not as transparent, okay, that's very
disturbing, but it's 99 percent accurate. Which system would
you want to have diagnosing you or your child?
That's a real dilemma. So I think we need to balance these
issues and be careful not to rush to legislate what's complex
technology here.
While I'm talking about legislation and regulation and the
kinds of decisions you'll be making, I want to emphasize that I
believe that we should not be regulating and legislating about
AI as a field. It's amorphous. It's fast-moving. Where does
software stop and AI begin? Is Google an AI system? It's really
quite complicated.
Instead, I would argue we should be thinking about AI
applications. Let's say self-driving cars. That's something
that we should be regulating, if only because there's a
patchwork of municipal and State regulations that are going to
be very confusing and disjointed, and that's a great role for
the Federal Government.
The same with AI toys. If Barbie has a chip in it and it's
talking to my child, I want to be assured that there are some
guidelines and some regulations about what information Barbie
can take from my child and share publicly. So I think that if
we think about applications, that's a great role for
regulation.
And then the last point I want to make is that we need to
remember that AI is a tool. It's not something that's going to
take over. It's not something that's going to make decisions
for us, even in the context of criminal justice. It's a tool
that's working side by side with a human.
And so long as we don't just rubber stamp its decisions but
rather listen to what it has to say but make our own decisions
and realize that maybe AI ought to be thought of as augmented
intelligence rather than artificial intelligence, then I think
we're going to be in great shape.
Thank you very much.
[Prepared statement of Mr. Etzioni follows:]
[GRAPHIC(S) NOT AVAILABLE IN TIFF FORMAT]
Mr. Hurd. Dr. Buck, you're on the clock, 5 minutes.
STATEMENT OF IAN BUCK
Mr. Buck. Thank you, Chairman Hurd, Ranking Member Kelly,
and distinguished members of the committee. I appreciate your
invitation to give testimony today on this important subject of
AI.
My name is Ian Buck. I'm the vice president and general
manager of Accelerated Computing at NVIDIA. Our company is
headquartered in Silicon Valley and has over 11,000 employees.
In 1999, NVIDIA invented a new type of processor called the
graphics processing unit, or the GPU. It was designed to
accelerate computer graphics for games by processing millions
of calculations at the same time.
Today, GPUs are used for many applications, including
virtual reality, self-driving cars, AI, and high-performance
computing. In fact, America's fastest supercomputer, at Oak
Ridge National Labs, uses 18,000 NVIDIA GPUs for scientific
research.
Our involvement with AI began about 7 years ago, when
researchers started using our processors to simulate human
intelligence. Up until that time, computer programs required
domain experts to manually describe objects or features.
Those systems took years to develop and many were never
accurate enough for widespread adoption. Researchers discovered
that they could teach computers to learn with data in a process
we call training.
To put that in context, to teach a computer how to
accurately recognize vehicles, for example, you need about 100
million data points and images and an enormous amount of
computation. Without GPUs, training such a system would take
months. Today's GPU-based systems can do this in about a day.
The world's leading technology companies have aggressively
adopted AI. Google and Microsoft's algorithms now recognize
images better than humans. Facebook translates over 2 billion
language queries per day. Netflix uses AI to personalize your
movie recommendations. And all those systems rely on thousands
of GPUs.
My job is to help companies like these bring intelligent
features to billions of people.
But AI's impact isn't just limited to tech companies. Self-
driving cars, as was mentioned, surgical robots, smart cities
that can detect harmful activities, even solving fusion power,
AI holds the best promise to solve these previously unsolvable
problems.
Here's a short list of problems for which I think AI could
help.
First, cyber defense. We need to protect government data
centers and our citizens from cyber attack. The scale of the
problem is mind-boggling, and we're working with Booz Allen
Hamilton to develop faster cybersecurity systems and train
Federal employees in AI.
Second, as was mentioned, healthcare. Nearly 2 million
Americans die each year from disease. We could diagnose them
earlier and develop more personalized treatments. The National
Cancer Institute and Department of Energy are using AI to
accelerate cancer research.
Third, waste, fraud, and abuse. The GAO reported that
agencies made $144 billion in improper payments in fiscal 2016.
The commercial sector is already using AI to reduce such costs.
PayPal uses AI to cut their fraud rate in half, saving
billions. And Google used AI to lower the cost of its data
centers by 40 percent.
Fourth, defense platform sustainment costs. Maintenance
costs are a huge challenge for the DOD, typically equaling 50
percent or more of the cost of a major platform, totaling over
$150 billion annually. GE is already using AI to detect
anomalies and perform predictive maintenance on gas turbines,
saving them $5 million per plant each year.
These are complex problems that require innovative
solutions. AI can help us better achieve these results in less
time and at lower cost.
For the role of government, I have three recommendations.
First, fund AI research. The reason we have neural networks
today is because the government funded research for the first
neural network in 1950. America leads the world in autonomous
machine vehicle technology because DARPA funded self-driving
car competitions over a decade ago.
While other governments have aggressively raised their
research funding, the U.S. research has been relatively flat.
We should boost research funding through agencies like the NSF,
NIH, and DARPA. We also need faster supercomputers, which are
essential for AI research.
Second, drive agency adoption of AI. Every major Federal
agency, just like every major tech company, needs to invest in
AI. Each agency should consult with experts in the field who
understand AI and recruit or train data scientists.
Three, open access to data. Data is the fuel that drives
the AI engine. Opening access to vast sources of data available
to the Federal Government would help develop new AI
capabilities so we can eliminate more mundane tasks and enable
workers to focus on problem-solving.
In closing, AI is the biggest economic and technological
revolution to take place in our lifetime. By some estimates, AI
will add $8 trillion to the U.S. economy by 2035. The bottom
line is we cannot afford to allow other countries overtake us.
And I thank you for your consideration. I look forward to
answering your questions.
[Prepared statement of Mr. Buck follows:]
[GRAPHIC(S) NOT AVAILABLE IN TIFF FORMAT]
Mr. Hurd. I thank all of you.
Now it's a pleasure to recognize the gentleman from
Kentucky for 5 minutes for his first line of questions.
Mr. Massie. To the doctor from Intel, I don't want to try
to pronounce your name. Help me out with that.
Mr. Khosrowshahi. Khosrowshahi.
Mr. Massie. Khosrowshahi.
You said that AI was aspirational, but now it's a reality.
Where did we cross the threshold? In the '90s, I worked at the
AI lab at MIT. I worked on the hardware, because the software
problem was too hard. And it seemed like you could solve
certain engineering problems in the software, but it still
feels that way to me.
What milestone did we cross, what threshold?
Mr. Khosrowshahi. So I hear this a lot, that people studied
neural networks in the '90s and they're kind of curious what
has changed. And so let me just put it into a broader context.
The history of AI goes back to the 1930s. The individuals who
started the field, John von Neumann and Alan Turing, they were
also the first people to build computers.
So the history of AI and computing has been tightly
intertwined. So computing, as Dr. Isbell mentioned, is really
critical. Compute power has dramatically increased since your
time to today.
Another, the next change is data. And the algorithms
potentially have not changed so much. They might look very
familiar to you. But there has been actually a remarkable
amount of innovation in the space of machine learning, which is
a dominant form of AI, and in neural networks that Ian
mentioned that is the state of the art today.
And invariably, these things change with time. The state of
the art in AI changes with time. But the three things that are
different today are computing power, data, and innovation in
algorithms.
Mr. Massie. This next question I'd like to ask all four of
you.
If there were going to be an XPRIZE for AI, what is the
next big milestone? What's the sword in the stone that somebody
should try to pull out and if they do they deserve a big
reward?
Dr. Etzioni.
Mr. Etzioni. I would observe that every time we build one
of these systems, whether it's in medicine or self-driving cars
or speech recognition, we're kind of starting from scratch. We
have to train them with these millions or hundreds of millions
of examples. We have to set the architecture by hand, et
cetera, et cetera, et cetera.
If we could build, as Charles was alluding to, more general
systems, which is something that we're very far from being able
to do today, a system that can work across multiple tasks
simultaneously without being retrained by hand every time, that
would be a major breakthrough.
Mr. Massie. So, Dr. Buck, what would it be for you? Maybe
driving from New York to L.A.?
Mr. Buck. I think we've had our XPRIZE in self-driving cars
with the work that DARPA did to kick off the industry
innovation. There's a huge market for the first car company to
really come up with a mass-produced self-driving vehicle.
I think AI at this point has the opportunity to
revolutionize individual fields, and some could benefit from an
XPRIZE, certainly healthcare. I think if we can identify an
opportunity to do personalized medicine, to look at the
genomics data that we've been able to get flooded with, with
new instruments, and apply AI to understanding the NED
treatments that are going to solve diseases, many of them just
need to be detected earlier. If we could find them early, we
could treat them. If we wait until the symptoms surface with
today's technology, it's sadly too late.
And if I had to add one more, I think there are huge
opportunities for AI to improve our infrastructure,
transportation, and just apply it to real modern problems
today.
Kansas City is doing a great project right now on detecting
potholes with AI. They're actually gathering all the data from
the weather data, the traffic information, and trying to
predict when a pothole is going to form on a particular road.
They are now up to 75 percent accurate within about 5 to 10
feet. So they can go out there ahead of time and treat that
road and tar it up before they have to tear it up to fix a
pothole.
There are so many different applications of AI, I think
those XPRIZES would be fun to watch.
Mr. Massie. Dr. Isbell.
Mr. Isbell. So I think there's sort of two answers to this.
One, all of us have said in one form or another that AI is
interesting in the context of a specific domain, and so there's
an XPRIZE for every domain.
But the more general question, I think, the answer is in
the AI lab from the 1990s. I was also in the AI lab in the
1990s, and my adviser was Rod Brooks. As you might recall, at
the time he was building a system called Cog, and the goal of
Cog was to build----
Mr. Massie. I remember Cog.
Mr. Isbell. Yes. I was probably sitting in the back when he
announced it with you.
The interesting thing about Cog was the idea was that they
were going to build a 3-year-old. And I think that the general
problem of intelligence is a difficult one, and the real XPRIZE
is being able to build someone we would recognize as
sophisticated as a 3-, 4-, or 5-year-old.
Mr. Massie. Okay. Just a speed round here, if you'll
indulge me. All four of you, I'll start here on the left.
Since you mentioned the 3-year-old goal that Professor
Brooks had, how far away is AI from passing the Turing test,
the classic Turing test, where if you were talking to this
being, sentient being in the computer, you wouldn't be able to
recognize it as not a human? How many years away are we?
You go first.
Mr. Khosrowshahi. Twenty-plus.
Mr. Massie. Twenty-plus.
Dr. Isbell.
Mr. Isbell. I assume the day after I die, because that's
how these things usually work.
Mr. Massie. Or the day after your funding runs out.
Mr. Etzioni. I should caution that the Turing test as it's
set up is kind of a test of human gullibility. I'm afraid that
we'll pass it much sooner than is said. But if your question is
about true human-level intelligence, I agree it's 20, 25 years
and beyond, effectively beyond the foreseeable future.
Mr. Massie. It's definitely easier to fool somebody than it
is to convince them they've been fooled, right?
Dr. Buck.
Mr. Buck. I agree with my colleagues. It's equivalent to
worrying about the overpopulation of Mars at this moment.
Mr. Massie. But it's the question. So what's your guess?
Mr. Buck. Oh, decades.
Mr. Massie. Decades. Okay.
Thank you very much.
Mr. Hurd. The gentlelady from Illinois is recognized.
Ms. Kelly. Thank you.
A few of you talked about the investment that needs to be
made in this and made into some of the agencies. So what amount
of money per year do you think the Federal Government should
invest in some of the science agencies and foundations that you
were referring to? Because it's easy to say we should invest,
but what's your realistic----
Mr. Etzioni. None of us are a policy or budgeting expert,
as you can see from the few seconds of silence, but----
Ms. Kelly. We're silent, too, so don't worry.
Mr. Etzioni. Let me suggest that much more than China. We
have a substantially larger economy. We should be investing a
lot more.
Ms. Kelly. Do you know what China is investing?
Mr. Etzioni. I don't know the exact numbers, but it's
certainly in the billions, according to their recently released
blueprint.
Ms. Kelly. Anybody else?
Mr. Khosrowshahi. So I don't know the numbers exactly, but
funding for NSF I think is on the order of billions. And this
money is highly leveraged. And funding graduate students
studying at AI universities is a really good way to spend the
money to accelerate innovation in AI.
And we do this at our company. We invest heavily in
university programs, many grad students, many labs. And we've
seen a lot of return in this specific area. So money well
spent.
So $3 billion versus $6 billion, the extra $3 billion will
be hugely effective in spurring innovation in AI.
Ms. Kelly. I was going to ask you, since your company is
big in this area, how are you spurring on diversity, more
women, more people of color?
Mr. Khosrowshahi. It is actually a prime directive that
comes from our CEO. So it's something that he is very focused
on. We have diversity requirements in our hiring. Everyone
knows these requirements in our hiring process. We focus on it.
And in our field in particular, we've seen firsthand--I
have--that additional diversity benefits in many ways. So we
discuss bias, transparency, having diversity in the scientific
demographics within our company. We have different ideas
presented. Sometimes these issues that you brought up are
highly nuanced and they surprise me.
And so, again, that's a directive from our CEO.
Ms. Kelly. Thank you.
Dr. Isbell, you talked about increasing diversity, but
starting in K through 12. What do you think schools need to do
K through 12 to spur interest or what resources do they have to
have?
Mr. Isbell. So two short answers to that. I'll answer the
first one first.
They have to connect what AI and what computing can do to
the lives of the people who are in school. That's the single
most important thing.
One thing that you just heard is that every dollar you
spend on AI has a multiplying effect. And it's true, because it
connects to all these domains, whether it's driving or whether
it's history, whether it's medicine, whatever it is. And just
connect that what you're doing will help you to do whatever
problem you want to solve.
But the main limiting factor fundamentally is teachers. We
simply do not have enough of them. You asked me how much money
you should spend. Whatever number you come up with, it's 10
times whatever you will come up with is the right answer.
But even if you spent all of that money, we are not going
to be able to have enough teachers who are going to be able to
reach enough tenth-graders in the time that we're going to need
in order to develop the next-generation workforce. It simply
isn't possible.
What we're going to have to do is use technology to make
that happen. We're going to have to make it so that Dr. Etzioni
can reach 10,000 people instead of 40 people at a time and can
work with people who are local to the students in order to help
them to learn. That's the biggest, I think, resource for
bringing people in who are young.
Ms. Kelly. Thank you.
Mr. Etzioni. May I just add something real quick?
It's not just the number of teachers, but it's teacher
training. My kids went to fancy private schools in Seattle that
had classes called tech, and I was really disappointed to learn
that they were teaching them features of PowerPoint because the
teacher did not know how to program. So we need to have
educational programs for the teachers so that they can teach
our kids.
And believe me, 8-year-old, 10-year-old, what a great time
to learn to write computer programs. And it will also help at
least with gender diversity and other kinds of diversity,
because at that point kids are less aware of these things and
they'll figure out, hey, I can do this.
Ms. Kelly. Also, we talked about not getting the immigrant
community. I serve on the board of trustees of my college, and
that's something that we talked about. And they shared that the
amount of foreign students has gone down drastically, because
they don't feel as welcome in the country, and it's in
engineering and the STEM fields that that has happened.
So I think my time is about up. Oh, I can keep going.
One thing I wanted to ask, what are the biases you have
seen because of the lack of diversity?
Mr. Buck. I think biases are a very important topic.
Inherently, there's nothing biased about AI in itself as a
technique. The bias comes from the data that is presented to
it, and it is the job of a good data scientist to understand
and grapple with that bias.
You're always going to have more data samples from one
source than another source. It's inevitable. So you have to be
aware of those things and seek them out. And a good data
scientist never rests until they've looked at every angle to
discover that bias.
It was talked about in our panel, in our testimonies. The
think I'd add is that an important part of it, to detect bias,
is where did it come from?
Traceability is a term that's used a lot in developing AI
systems. As you're going through and learning better neural
networks, inserting more data, you're recording the process and
development.
So when you get out to a production system, you can then go
back and find out why did it make that incorrect judgment and
find out where was that bias inserted in the AI process and
recreate it.
It's very important for self-driving cars, and I think it's
going to be important for the rest of AI.
If you don't mind me going back to your previous question,
I also think it's important that the committee recognize that
AI is a remarkably open technology. Literally anyone can go
buy, on a PC, download some open source software. They can rent
an AI supercomputer in the cloud for as little as $3 and get
started learning how to use AI. There's online courses from
Coursera, Udacity. Industry, too. NVIDIA has an industry
program called the Deep Learning Institute to help teach.
So those technologies are remarkably accessible and open,
and I think that goes to your diversity, making it available.
It inspires students, kids with ideas of how they can take data
and apply these technologies. There's more and more courses
coming online. And I think that will inspire the next wave of
AI workers.
Mr. Isbell. If I can just add to that.
I think the first round of bias comes from all of our
beliefs, including myself. The sort of fundamental thing we
want to believe is that the technology is itself unbiased and
must be and that it is no more biased than a hammer or a
screwdriver. But we'll point out that both hammers and
screwdrivers are actually biased and they can only be used in
certain ways and under certain circumstances.
The second set of bias comes from the data that you choose,
which is exactly what Dr. Buck said.
I'll give you an example. When I was sitting in an AI lab
apparently across the hall from you, a lot of the original work
in vision was being done, particularly in face recognition.
A good friend of mine came up to me at one point and told
me that I was breaking all of their face recognition software,
because apparently all the pictures they were taking were of
people with significantly less melanin than I have.
And so they had to come up with ways around the problem of
me. And they did, and got their papers published, and then they
made better algorithms that didn't depend upon the assumptions
that they were making from the data that they had.
This is not a small thing. It can be quite subtle, and you
can go years and years and decades without even understanding
that you are injecting these kind of biases just in the
questions that you're asking, the data that you're given, and
the problems that you're trying to solve.
And the only way around that is to, from the very
beginning, train people to think through, in the way that Dr.
Buck said, to think about their data, where it's coming from,
and to surface the assumptions that they are making in the
development of their algorithms and their problem choices.
Mr. Etzioni. Bias is a very real issue, as you're saying,
as we're all saying. But we have to be a little bit careful not
to hold our database system to an overly high standard. So we
have to ask, what are we comparing the behavior of the systems
to? And currently, humans are making these decisions, and the
humans are often racist, they're often sexist. They're biased
in their own way.
We know, you talked about the case with a judicial
decision. We have studies that show that when the justices are
hungry, you really don't want them to rule at that point. You
want them to go to lunch.
So my perspective is let's definitely root out the bias in
our systems, but let's also think about these collaborative
systems where humans are working together with the AI systems,
and the AI system might suggest to the person, hey, maybe it's
time for a snack, or you're overlooking this factor.
If we insist on building bias-free technology or figuring
out how to build bias-free technology, we're going to fail. We
need to build technology and systems that are better than what
we have today.
Mr. Hurd. Ranking Member, we need an XPRIZE for that, you
know, to figure out when I'm hangry and make better decisions.
Ms. Kelly. My last question is, those of you representing
companies, do you have internship programs? How do you reach
out into the community?
Mr. Buck. Certainly. I think the most exciting work is
happening in our research institutions and even at the
undergrad and earlier levels.
We're a huge proponent of interns. Myself, I was an intern
at NVIDIA when I started at the company and worked my way up to
be a general manager.
So I'm a huge proponent of interns. They bring fresh ideas,
new ways of thinking, new ways of programming. They teach us a
lot about what our technology can do.
Mr. Khosrowshahi. If I'm allowed to comment on your last
question.
So we talked about bias, but this line of thinking applies
to everything. So transparency. I heard accountability. Humans
are largely not transparent in their decisionmaking. This is
something that's been studied exhaustively by people like
Daniel Kahneman.
So I think it's very interesting to hear this firsthand,
but we have to be concerned about humans as well as machines.
And when they interoperate, that's even more challenging.
But, again, humans are biased, humans are transparent. And
this is something to be cognizant of in your decisionmaking. I
just wanted to stress that.
Ms. Kelly. Thank you.
Mr. Hurd. One of the reasons we do these kinds of hearings
is to get some of the feedback from the smart people that are
doing this.
And, Dr. Buck, for example, we continue to do our FITARA
Scorecards looking at how the Federal Government implements
some of these rules. One of the questions we're going to start
asking our Federal CIOs is, what are you doing to introduce
artificial intelligence into your operations?
So, Federal CIOs, if you're watching, friends at FedScoop,
make sure you let them know that's going to be coming on the
round six, I think, of the FITARA Scorecard.
Where to start? So, yes, basic research. It is important.
What kind of basic research? Do we need basic research into
bias? Do we need basic research into some aspect of neural
networks? Like, what kind of basic research should we be
funding to start seeing that, to raise our game?
And all these questions are open to all of you all, so if
you all want to answer, just give me a sign, and I'll start.
But, Dr. Buck, do you have some opinions?
Mr. Buck. Certainly. As data science in general becomes
more important to understanding the root cause of bias and how
it is introduced and understood, I think it is a very important
basic research understanding.
A lot of this work has been done. It can be dusted off and
continued. I think it will be increasingly important as AI
becomes more of the computational tool for changing all the
things that we're doing.
Industry will tackle a lot of the neural network design.
You have some of the smartest people in the world here in the
U.S. building newer, smarter neural networks. They're largely
focused on consumer use cases: speech recognition, translation,
self-driving vehicles.
I feel like the science applications of AI, how AI can
assist in climate and weather simulations, how AI can assist in
healthcare and drug discovery, are still early. And it is an
area that has less of a commercial application but obviously
really important to this country.
You have some amazing computational scientists at the DOE
labs that are starting to look at this. I think they also
recognize the opportunity that AI can assist in simulation or
improve the accuracy or get to the next level of discovery. I
think there are some real opportunities there.
And we're starting to see that conversation happen within
the science community. Any more encouragement and, of course,
funding to help amplify it would be greatly appreciated.
Mr. Etzioni. I think you make a great point. There is the
investment from Google, Intel, and Facebook. But there is so
much basic research that they won't do.
And I also can't emphasize enough how primitive the state
of AI is. Sure, we've made a lot of strides forward, but----
Mr. Hurd. Not to interrupt, but give me some. What are
examples of basic research they won't do that we should be
doing?
Mr. Etzioni. Common sense. Something that you and I and
every person knows and AI does not. That a finger has five
hands. That people typically look to their left and their right
before they cross the street.
There's an infinite set of information that machines don't
have. As a result, they really struggle to understand natural
language. So we've seen success where the signal is very
limited, like in a game of Go or in speech recognition.
But all you have to do is turn to Alexa or Siri and realize
just how little our AI programs understand and how little can
we have a conversation with them.
So I think research into natural language processing, into
commonsense knowledge, into more efficient systems that use
less training data, all of these are very, very challenging
fundamental problems. And I could go on and on.
Mr. Hurd. Gentlemen.
Mr. Isbell. So I have very strong opinions about this, but
I will try to keep it short.
I think if I were going to pick one--I'm going to give you
two answers--and if I was going to pick one thing to focus on
that I don't think we're doing enough of, it is long-lived AI.
That is, a lot of the work that we're doing are systems
that solve a specific problem for a specific relatively short
period of time is why it ends up looking like supervised
learning as opposed to something like long-term decisionmaking.
But if you think about what makes human beings so
interesting, there are two things. One is that we depend upon
each other, and the other is that we learn and we live for a
really long time, not measured in minutes or hours but measured
in decades.
The problem of reading is hard. It takes human beings 6, 7,
8 years to learn how to read. We need to understand what it
means to build systems that are going to have to survive. Not
just figure out how to turn the car now, but have to figure out
how to live with other intelligent beings for 10, 20, or 30
years. That's, I think, a sort of truly difficult problem.
But having said that, I'll back off and say, I think the
answer is you trust your agencies who talk to the community.
NSF has a long list of things that they believe are important
to invest in AI and other things as well and the get that by
having ongoing communications and conversations with a large
community. It creates a kind of market, as it were, of what the
interesting ideas are.
And I trust them. I listen to them. I talk to them. They're
the mechanism that sort of aggregates what people are
believing.
And then, in some sense, what you can do or what government
can do or what these agencies can do is to push us a little bit
in one direction or another by giving incentives for thinking
about a problem that people aren't necessarily thinking of.
But, in general, I trust the people who are doing the work.
Mr. Hurd. Dr. Khosrowshahi.
Mr. Khosrowshahi. So we've been talking about high-level
aspects of AI, decisionmaking and so forth. But in some of our
testimonies we mentioned that there is a substrate for
computation that enables AI. You have lots of data, need a
while to compute.
We're at an interesting point in time where we're having
rapid innovation in AI, lots of successes. It's being driven by
availability of data and compute. The amount of data is
increasing really, really rapidly, and the compute has to
commensurately increase in power.
So that will require basic research and innovation at the
silicon level, at the hardware level, which is what Intel does.
We have fabs. We build the hardware from glass.
So areas such as silicon photonics, analog computing,
quantum computing, low-powered computing, all of these areas
are potentially great investment NSF funding opportunities for
you.
And I'd like to also mention the landscape for getting AI
systems to work involves so many different things. It requires
machine learning, teachers, and so forth. But it requires
things that seem prosaic but are really important, reliable
software systems that are accountable, scalable, robust, and so
forth.
Again, that comes from investing in STEM and computer
science in early stages of someone's career development.
Mr. Hurd. So we've talked about bias as a potential
challenge that we have to deal with as we explore and evolve in
the world with AI. Another way you can manipulate a learning
algorithm is by loading it up with bad data.
What are some of the other challenges and other threats to
artificial intelligence that we should be thinking about at the
same time that we think about bias and integrity of the data
that's involved in learning? Anyone.
Dr. Buck.
Mr. Buck. I'll emphasize that it's easy to say we have lots
of data. It's actually quite challenging to organize that data
in a meaningful way. The Federal Government has vast sources of
data. It is very unstructured.
Mr. Hurd. Very aware.
Mr. Buck. And that is a challenge. We just spent a decade
talking about big data. And as far as I can tell, we've largely
collected data, not really done much with it.
You now have a tool that can take all that data you've
collected and really have some meaningful insights, to make a
new discovery in healthcare, to save enormous amounts of money
by finding inefficiencies or, worst, waste or fraud. But that
data needs to be aggregated, cleaned up, labeled properly, and
identified.
I certainly would make sure that not only that the Federal
Government has an AI policy but also has a sister data policy
as well to organize and make that data actionable and
consumable by AIs, whether within the Federal Government or
make them available to the larger research community.
I am sure there are dozens, if not thousands, of Ph.D.'s
waiting to happen if they just had some of the more interesting
Federal data to really make those kinds of discoveries.
Mr. Hurd. Well, Dr. Buck, one of the first things this
committee looked at was the DATA Act. And, shocker, the Federal
Government was actually ahead of the game in trying to make
sure that we're taking on that data and adding some structure
to it. Implementation of that, as you have pointed out, is a
bit tricky. So any tools that you all have to help with that
would be great.
Other concerns?
Dr. Isbell.
Mr. Isbell. So I'll add one. I agree with everything that
Dr. Buck said and what other people have said before. Data is
the problem. But one real issue is we typically build AI
systems that don't worry about adversaries.
So this ties back into the notion of long-lived AI systems.
So we're building a system that's going to determine whether
you have a tumor, whether you have a heart attack, whether you
should get a mortgage, but we're not spending a lot of energy--
some people are thinking about this--we're not spending a lot
of energy figuring out what happens when we send these things
into the wild, we deploy them, and other people know that
they're out there and they're changing their behavior in order
to fool them.
And how do we make them change over time is an arm's race.
You can think about this security. It's easy to think of. We
could think of something even simpler, like spam. I get all
this terrible mail. I build a system that learns what my spam
is. The people who are sending spam figure out what the rules
are and what's going on there, and then they change what they
do. And it just keeps escalating.
And so this notion that you're going to have to not just
solve the problem in front of you but solve the problem as it's
going to change on the next round, the round after that, and
the round after that, I think that's a real limitation of the
kind of way that we build systems, freeze them, and then deploy
them.
And I'm not saying that that's all people do and that no
one is thinking about it. But I do think, because we tend to
think in this sort of a transactional way about AI, we
sometimes don't think through the consequences of having long-
term systems.
Mr. Khosrowshahi. I'd like take a slightly different tone.
So we have talked in our testimonies about bias, privacy,
transparency, assurances of correctness, adversarial agents
trying to take advantage of weaknesses in the system.
So one thing that I've seen in this past year that I
haven't seen in the past 10 years is these things are discussed
at academic conferences. Companies like Intel, my team,
actually these are some of the top priorities, these issues
that you raise. They're discussed. They're attracting some of
the best minds in the field.
I just introduced the idea of transparency literally months
ago. And it's a really interesting area. It's highly nuanced.
Humans are a tribal, multi-agent society. There are times when,
if people have more information, the overall performance of the
system goes down. It's very nonintuitive. Things can happen.
Academics are pouring a lot of effort into this area.
So I'm just very, very optimistic that the things we've
enumerated today are being addressed, and we should just
amplify them. So the government can play a big role in
investing in things like academic research.
It is quite different to me--I don't know if you guys
concur--but the last major machine learning conference, NIPS,
was really eye-opening to me, that there is a workshop on
transparency, there is a workshop on bias, there is a workshop
on diversity in the demographics of the AI community.
So we are definitely on a very positive and virtuous track,
and I'm asking government to just amplify this however it can.
Mr. Hurd. The distinguished gentleman from the Commonwealth
of Virginia is now recognized.
Mr. Connolly. Thank you, Mr. Chairman.
And thank you to our panel.
Dr. Etzioni, from here, I had a little trouble reading what
was underneath your name. And I thought for a minute it said
alien AI. I thought, wow, we really are getting diverse in the
panels we are putting together here. Alien AI.
Mr. Etzioni. I come in peace.
Mr. Connolly. Yeah. Thank God.
So we were reminded rather dramatically last September with
the Equifax hack that compromised information on 145 million
Americans as to the risks of devastating cyber attacks and the
absolute need for creating shields and protective measures,
both for the government and for the private sector.
According to the 2016 report from the NSTC, the National
Science and Technology Council, AI has important applications
in cybersecurity and is expected to play an increasing role for
both defensive and offensive cyber measures.
Dr. Khosrowshahi--and I'm from now on going to say the
doctor from Intel--how can AI be most useful in defending
against cyber attacks?
Mr. Khosrowshahi. So I'll suggest a few ways, and I guess
we'll have other opinions.
So cybersecurity, of course, is a major issue broadly in
computing, as well as in AI, and as well at Intel. It is one of
our primary focuses.
So in terms of addressing cyber attacks using AI, cyber
attacks are intentionally devious and nefarious, obscure. And
these kinds of actions are really well suited to the latest
state of the art in AI, machine learning.
That is algorithms can take large corpora of data--these
are inputs from whatever the type of cyber attack you're
experiencing--and they can build a model of the cyber attack
and a response, essentially.
And the response can have very low latency. It can study
the statistics of the attack, potentially it's a novel attack,
build a model, and respond very quickly.
So that's one way we can address cybersecurity, is with
better models to defend against it.
Another way--another thing that we can--it's not in answer
to your question--but when we build models, it's good to know
the set of possible attacks, because a researcher, a data
scientist, is very cognizant of building robust models that are
resistant to adversarial events.
So as we get knowledge of cybersecurity issues in this
area, AI, we build in security and defense against cyber
attacks into the models such that adversarial actions do not
perturb or give erroneous results.
Mr. Connolly. Presumably also one of the advantages of AI
would be early detection. I mean, part of the problem of cyber,
certainly from the Federal Government's point of view, but
apparently in the private sector as well, is when we finally
realize we have been compromised, it's too late.
Mr. Khosrowshahi. That's right.
Mr. Connolly. And AI has the potential for early detection
and diversion, preemption, protective walls, whatever.
Mr. Khosrowshahi. That's right. The nature of these attacks
could be so devious that the smartest human security experts
could not identify them. So can either augment our human
security experts or we can have systems that are early
detectors that can just flag this is a potential threat. And
these systems are really well suited for doing this, latency
and learning very quickly.
Mr. Connolly. Anyone else on the panel is more than welcome
to comment.
Dr. Etzioni.
Mr. Etzioni. I just wanted to add that at the root of the
Equifax hack was human error, several human errors. So
something you might want to think about is, what are the
incentives that we have in place to avoid that? What are the
consequences that people at Equifax face--and not to pick on
them--for making those mistakes with our data?
I think if we put the right incentive structure in place,
it's not a technical solution, but it'll help people to be more
watchful, and they should be.
Mr. Connolly. Yeah.
Mr. Buck. The statistics here are alarming. And the rate of
attacks are growing exponentially way faster than we can expect
a human operator, even with the tools they have today, to keep
up.
This is a very hot topic in the startup community. There
are many startups trying to apply AI to this problem. It's a
natural fit.
AI is, by nature, pattern matching. It can identify
patterns and call out when things don't match that pattern.
Malware is exactly that way. Suspicious network traffic is that
way.
One startup we work with, they're claiming the top AI
software is only able to capture about 62 percent of the
potential threats that are out there. But by applying AI, they
can shorten the time to discovery and get to 90-plus percent
accurate malware detection, and the false error rate, get it
down to less than 0.1 percent where normally it's 2 percent.
It's an opportunity to increase the throughput of our
detector systems and make them much more rapidly responsive.
Mr. Connolly. So why aren't we doing it? Is it the cost?
Mr. Buck. The AI just needs to be developed. It is in the
process of being developed by those startup companies. It's not
as talked about in application as maybe video analytics or ad
placement, but it is certainly active.
Mr. Connolly. Well, you put your finger on two things,
among others. But one is the exponential growth in the volume
of attacks. I talk to some Federal agencies, and I'm stunned at
the numbers. I mean, I know of one Federal agency, not a big
one, where the cyber attacks or attempted attacks are in the
hundreds of millions a year.
And you're absolutely right. I mean, this particular
agency, its mission isn't cyber. It's got a very human mission.
And it's trying to put together through Band-Aids and other
measures some protection. And it does raise questions about the
ability of, in this case, the Federal Government to protect
itself.
Mr. Buck. I'm seeing a sea change in that as well. Not just
are we looking to protect our firewalls and the data coming
into our firewalls, but the data traffic behind the firewall.
Assume you are attacked, for the sake of argument, and look
at the traffic that's inside your firewall to detect it.
Because as was mentioned before, in many cases you may already
be compromised and you don't know it.
So it's important to look at both, the front line as well
as behind the lines, in understanding your network traffic and
your security.
Mr. Connolly. And the second thing this conversation I
think underscores, and we had testimony yesterday from the
intelligence community, but the idea that the Russians are not
going to continue their attacks and attempts to distort our
electoral process is naive. All 17 intelligence agencies in the
United States Government testified to the fact that it is an
ongoing threat and the midterm elections will be a target.
So in a democracy, that's the very heart of how we
function. How do we protect ourselves? And I think maybe we've
got one tool, maybe a very critical tool, in terms of
artificial intelligence. But trying to get that out to the
myriad localities, over 10,000 localities in the United States,
is going to be a different kind of challenge.
I thank you, Mr. Chairman.
Mr. Hurd. Mr. Lynch, you are now recognized.
Mr. Lynch. Thank you, Mr. Chairman. I appreciate that.
Dr. Etzioni, in your written testimony you state, and I
quote here, ``We can and should regulate AI applications.''
Obviously, as more and more AI systems are used to collect more
and more sensitive and personal data on Americans, there are
palpable and real privacy concerns.
What are the ways in which you think that the regulations
that you anticipate would serve to protect the private
information of Americans?
Mr. Etzioni. So I think that there are some principles that
I can talk about. And, frankly, you and your staff are probably
better qualified to think through specific regulations.
But a principle that I would really advocate is identifying
when AI is involved. And that's something that we can regulate
so that the bots, at least the homegrown ones, state that
they're AI. We had Intel inside. We should have AI inside.
Most recently we've seen that there are examples of fake
pornography, superimposed celebrities on top of bodies and
things like that. If we can't trust the integrity of our
pornography--obviously I'm joking.
Mr. Lynch. Thanks for making that clear.
Mr. Etzioni. But the point is we should label when AI is
being used. And, likewise, we should be clear when we have AI
systems in our homes. Alexa, AI Barbie, the Roomba vacuuming
our floor, they naturally also vacuum up a huge amount of data,
some of it from our kids, if Barbie is talking to our kids. We
should have regulations about where that information can go.
Mr. Lynch. So the proliferation of AI, I just see it
proceeds at a velocity far exceeding the ability of Congress to
keep up with it, and that's true with many technologies. And
oftentimes we rely heavily on the private sector to look at
those ways that, if AI is being broadly used, how we might
develop a protocol that would prevent that private information
from just getting out there.
And we have, in a very narrow sense, the Equifax situation
where we have the names, addresses, Social Security numbers of
150 million Americans out there, just gone. So they basically
burnt the entire Social Security number system as a reliable
and secure indicia. So that's gone. And it's just because one
company was very lazy about protecting data.
And so I'm just concerned. I have similar concerns about AI
being out there and these bots. And we've got some pretty
creative hackers out there, Russians and others, that have been
able to access some very, very sensitive information. At one
point they swept every bit of data from any individual who had
applied for a high-level security clearance in this country.
And so I could just see if there are, as you say, not
necessarily household appliances, but other forms of AI
operating a higher level, if those are hacked, it just
increases the magnitude of our vulnerability exponentially.
And I'm just trying to think in advance, as this is all
happening in real-time, how do we protect the people who
elected us? We're all for innovation, but I think with the
appropriate safeguards in place.
Mr. Etzioni. The thing that I would like to highlight,
though, is that you're right, those are some scary realities.
But they are realties. They're often instigated from the
outside. So maintaining our strategic edge.
And that's why I emphasize regulating applications as
opposed to the AI field and AI research itself. If we adopt an
overly defensive, dare I even say in a reactionary posture,
we're just going to lose.
So this is a very competitive global business. And staying
ahead, which we're all trying to do in various ways for
education, et cetera, is essential.
Mr. Lynch. Okay. Thank you.
I assume my time has expired, Mr. Chairman. I yield back.
Mr. Hurd. Dr. Isbell, did you have a response to that
question?
Mr. Isbell. I just want to add something. I think it's
important to recognize here everything that you brought up are
deep concerns. But AI is a secondary problem there. The primary
problem there is that we are sharing our data constantly.
Every one of you has a cell phone, possibly two of them,
you have a watch, which is pinging all the WiFi hotspots
everywhere you go. Each one of those devices has a unique ID.
That unique ID is not you, but that unique ID is with you all
the time. I can figure out with very little effort who you are,
where you are, where you come from.
By the way, I've deployed systems myself, this is 10 or 15
years' old worth of technology, where I can predict what button
you're going to press on your remote control after just
observing you for one weekend.
We are creatures of habit. We are sharing our data in our
cars, our phones, everything that we do. The data itself, even
if it's anonymized, is giving amazing amounts of information
about us as individuals. That's the primary problem.
The secondary problem is the AI, the machine learning, the
technology, which can look at it very quickly and bring
together the obvious connections even though you've tried to
hide them.
But the first thing I think to think about is it's not the
AI, because computers are just fast, that's just going to
happen. It's the fact that we are sharing data, and we've given
very little thought to what it means to protect ourselves from
the data we are willingly giving to everyone around us. And I
don't have an answer, but that, in some sense, is the root
problem.
Mr. Lynch. Mr. Chairman, if I could.
The ability of AI to aggregate the data, make sense of it,
and give it direction and a purpose and a use, that's the magic
of AI. The data's out there. And you're right, that's a
problem. But I'm worried about weaponizing that raw data that's
out there and how do we control that.
But thank you. I think you offered a very good
clarification. Thanks.
Mr. Khosrowshahi. Let me make a short comment.
So I liked to balance the discussion and present a slightly
dissenting view to Dr. Etzioni.
Well-intentioned efforts, such as labeling robots and other
devices that employ AI, it could have unintended consequences.
You have in the State of California, my State, we now know that
asparagus and coffee cause cancer. So we are going to have
labels on every piece of food and every building that this
thing causes cancer. And these signs are becoming
uninformative.
So I would just be wary of unnecessary regulation or
imposing regulation on a very young and rapidly moving field,
because I can immediately see that it can have some adverse
consequences.
We talked about transparency. To use your example, would
you want something that is labeled and worse performing or
unlabeled and better performing, to use your example.
And just in general, our view at Intel is that legislation
should be based on principles, not on regulation that mandates
certain kinds of technology. So we are self-regulating.
This field is wonderful, that it does a lot of high-minded
academics who are now leaders in business, and there is a
strong impetus to be good stewards of this technology to do
good. And we have lots of things that we can impose on
ourselves to self-regulate to potentially address some of the
adverse conditions that you mentioned. Not all of them. Perhaps
some of them do need legislation.
Mr. Hurd. I've got some final questions. And this first
question is for everyone. And I know you all have all spent
your adult lives trying to answer this question, and so I
recognize this before I ask.
And, Dr. Buck, I've got to give some kudos to your team
that was out at the Consumer Electronics Show. They were very
helpful in helping me understand some of the nuance of
artificial intelligence. And if artificial intelligence was
based on Fortran 77, I'd be your guy. That's my background
experience.
But I understand how to introduce antivirus software into
your system. I understand how you introduce CDM into a network.
When we ask all the Federal CIOs how are you thinking about
introducing artificial intelligence into the networks, the
first question I'm probably going to get is, well, it's really
hard.
And so my question is simple. And we've all been saying
that AI is interesting because it's domain specific and I
recognize how broad this question is. But how do we introduce
AI into a network, into a system, into an agency?
Mr. Buck. That's a great question. And AI can seem like
rocket science. And first off, having this conversation is the
first step. Explaining what it is and understanding it so they
can comprehend it is, obviously, the first step.
And where I've seen it work most successfully is in
meaningful simple pilot projects. Project Maiden, which is a
project with DOD, where they're using AI to help with
reconnaissance so that airmen are not staring at TV screens for
8 hours a day waiting for something to happen. They're letting
the AI do the mundane parts of the job so our soldiers can do
the decisionmaking.
That kind of application of AI is well established. People
know how to do it. You don't need to invent a new neural
network to do it. It's the same work that's being done
elsewhere. But by creating these pilot projects inside of these
agencies, they are dramatically improving the lives of the
people that work there.
Mr. Hurd. So do we believe we're at a point now where the
agencies can be the ones that are involved in training the
algorithm. Okay, you find an algorithm, you figure out what
dataset you need to train it. And do you expect the person at
Department of Interior to be the one training that, or is it
folks that are providing that service?
Mr. Buck. You can do it both ways. I've definitely seen
public partnerships where agencies are going outside for
consulting to help apply AI technology to a specific problem.
Or in some cases the neural networks are well established.
Image recognition is where AI started. It is a well-established
technique. The networks are open source. The software is open
source and public.
So I think if you find those use cases off the bat that are
well published and, as was spoken, in these AI conferences well
shared. The beauty about AI is that it's incredibly open, it's
being done in the open source community, it's all being
published. And it takes very little work to take one of those
established workflows and apply it. And then the next step is
to share that success.
Mr. Hurd. Dr. Khosrowshahi.
Mr. Khosrowshahi. So AI has changed over the last 80 years
and it almost surely will change. We talked about neural
networks. Five years from now, almost surely--I'm on TV--but I
guarantee it's going to be something different.
But the underpinnings are you have data, you have model,
you have inferences. You have data that has statistical
distribution, whether it's images, whether it's a car driving
down the road collecting video in the U.S. or Canada or
wherever, different statistics. You build models, the models
try to understand the statistics of the data, and then you can
ask the model questions. Is this a cat or a dog? Is there a
stop sign approaching me? That's basically what AI is today.
So if you just take these simple underpinnings and then
apply them to whatever public policy or application CIOs want
to insert into their business workloads and so forth, just
understanding that basic element. There's going to be some
data, it will have some statistical properties, maybe it will
be difficult for a human to understand them. A machine could be
better and faster, more robust, more power efficient than the
brain. And then it can perform inferences.
And whether or not you choose to rely on these inferences
depends on how good the model is, how much assurances of
correctness you have. I mean, the landscape of AI is so vast
and it's touching so many different things. And it's still, I
would again stress, that it's very early on. We don't have
artificial agents making decisions for us almost anywhere.
So even in finance, you would expect automated trading
systems. It's not there yet. We're still in the very early
stages. There is not widespread adoption in the industry. It
will get there, but it's still early on.
But, again, the AI, the underpinnings and the applications,
there's this model data inference. You can stick it in anywhere
where that works.
Mr. Isbell. So in the interest of time, I'll keep this
short.
I want to distinguish between at least two different
things. One is face recognition and that class of things versus
shared decisionmaking. I think the answer for things like face
recognition, relatively straightforward. At the risk of
oversimplifying, it's like asking the question, how can we
integrate the internet? How can we integrate telephones? It's
relatively straightforward. It's well understood, it's very
clear, and you can ask yourself how to use the screw driver.
The shared decisionmaking is what's difficult. That
requires that the domain experts are part of the fundamental
conversations. The research question from my point of view is
figuring out how to be able to use humans in order to train the
systems that we have when they don't understand machine
learning and AI, but they do understand their domain. How do
you get those people to talk to one another?
I'm not worried about the deployment of face recognition.
I'm worried about how I'm going to get an intelligence analyst
to understand enough about what it is they are doing so that
they can communicate to a system that will work with them in
order to make decisions.
That's where the difficult problem is, but it's really no
different than just trying to understand what it is they
actually do. The problem is, the thing that we know, is that
people are terrible at telling you what it is that they do. You
can't ask them and they tell you. You have to watch them,
observe them, model them, and give them feedback. It's an
iterative, ongoing process.
Mr. Etzioni. I wonder if an approach would be to focus on
outcomes and metrics and grand challenges. And if you ask for
those rather than demanding AI and then they have to resort to
AI to satisfy those mandates, that might work.
Mr. Hurd. One minute for all four of you all to answer
these two questions.
What datasets in the government do you want access to or
should the AI community of people that are working on these
challenges get access to? And what skill sets should our kids
in college be getting in order to make sure that they can
handle the next phase when it comes to artificial intelligence?
Mr. Isbell. All of them. And the skills that the students
need in college, they need to understand computing. There
shouldn't be a single person who graduates with a college
degree who hasn't taken three or four classes in computing at
the upper division level. They need to understand statistics.
And they need to understand what it means to take unstructured
data and turn it into structured data that they can construct
problems around.
Mr. Khosrowshahi. So on the datasets, things like NOAA,
weather data, things that are not sensitive have private
information, those would be the first. And there's a vast trove
of this. This would be immediately useable by academics.
But on the skill set side, if I were to pick one, it would
be computer science. I would invest as much as possible in
teaching computer science K through 12, especially in high
school.
Mr. Hurd. Dr. Etzioni.
Mr. Etzioni. Research funded by NIH, by NSF, DARPA, et
cetera, is often not available under open access. Journals keep
it behind pay walls. That's changing way too slowly.
So the dataset that I would like everybody, human and
machine, to have access to is the data and the articles that
you and we as taxpayers paid for. I think that's incredibly
informant.
As far as the skill sets, I would say that everybody in
college should be able to write a simple computer program and
to do a simple analysis. And we can get there, and, remarkably,
it's not required.
Mr. Hurd. Dr. Buck, last word.
Mr. Buck. I certainly would love to see all the datasets. I
certainly also would like to see access to the problems around
healthcare. And I know those are sensitive topics, but the
problem is too important, the opportunity is too great, and it
is where I feel like AI will truly save lives. If we could
figure out to make that data available, it would be an amazing
achievement.
In terms of education, I believe that data science is
becoming a science again. And I also feel like training a
neural network is not that hard. I think it can be done at the
junior high level.
And the access to technology is available today. And I
think we should start teaching students what this tool can do.
Because it really is a tool and will inspire new applications
that will come from the interns, the undergrads, the college
students. That's what makes this fun.
Mr. Hurd. Well, gentlemen, I think my colleagues would
agree with me on this, this has been a helpful conversation.
There is a lot packed into your all's testimony that's going to
help us to continue to do our work on the Oversight Committee
and to look at opening up some of these datasets. How do we
double down on NSF funding? How do we focus on getting more? I
think every kid in middle school should have access to a coding
class. And we're working on that stuff down in the great State
of Texas.
And many of these points that you make, we're going to be
talking to folks in the government, in early March, in the
second series of this AI series. We intended to invite GSA,
NSF, DOD, DHS and to continue this conversation about how they
are introducing and looking at artificial intelligence and what
more support they need from Congress.
So, again, I want to thank you all and the witnesses for
appearing before us today.
The hearing record will remain open for 2 weeks for any
member to submit a written opening statement or questions for
the record.
And if there's no further business, without objection, the
subcommittee stands adjourned.
[Whereupon, at 3:54 p.m., the subcommittee was adjourned.]
[all]