Computing at the Crossroads: Intersections of Research and Education

Computing at the Crossroads: Intersections of Research and Education


[MUSIC PLAYING] Good morning. My name is Hashim Sarkis, dean
of the School of Architecture and Planning at MIT, and I’m
very pleased to welcome you to the next round
of speakers who will be focusing on the
intersections of learning and computation. One of the first things I
learned when I came to MIT four years ago is
that at MIT, we use the word “learning” much
more often than education. At MIT, we learn. We don’t educate. We learn from each other. We learn from our students. We learn from the
machines that we make that, in turn, learn from us. And in all this,
we learn by doing. This pedagogical approach
underpins the seal of MIT– Mens et Manus– and its mission
to use science and technology to advance society. Motto and mission
mirror each other. In 1896, pragmatist
philosopher John Dewey, the man who himself coined the
phrase “learning by doing,” founded the famous
school in Chicago called “The Lab
School” on the premise that we could bring scientific
method to bear on education. Learning, like
scientific exploration, should advance through real-life
experimentation, not by rote. When carefully examined,
lessons from life can become lessons from school. Dewey extended his use
of scientific methods from education to
other aspects of life. For example, he advocated
using scientific method to the solution of
social problems. This approach he
“called democracy.” Remarkably, his embrace
of scientific method in [INAUDIBLE] led him to
also embrace its corollary– human imagination. Learning by doing
involves action. Action generates uncertainties
to which we have to respond. And in such uncertainties,
it is our imagination that helps us
relate the outcomes of our past experiments
to anticipate future ones. Every great advance in
science, Dewey observed, has issued from a new
audacity of the imagination. Dewey worked very
hard to integrate imagination and scientific
method in learning, but for many reasons,
these two attributes have been increasingly severed
from each other over time. Dewey was writing in 1929 at the
beginning of the second machine age and the Great Depression
when new machines were robbing people from
their jobs and where their mechanical jobs
and rote education were, in turn, robbing
them from their ability to exercise their
imagination and to deal with emerging uncertainties. I don’t know which
machine age we are at now, but we certainly live in
an age of uncertainty. We are creating
new machines that are threatening our
jobs, our imagination, and our democracies
like never before. It is precisely
in our commitment to learning about
these new machines by founding the Schwarzman
College of Computing that we need to
invest in cultivating the human imagination. After all, this is where
the new jobs will always be after the old jobs are
taken over by the new machines. Yesterday, I learned that Dan
Huttenlocher, the first dean of the College of Computing,
had studied at John Dewey’s Lab School in Chicago. It bodes very well. Those of you who know
Dan have seen the work that he has done at Cordell
bringing design and creativity into science and technology. Dan, welcome back to MIT– the perpetual lab school where
we thrive on uncertainty, where science is
the eternal method, and where, with your help,
imagination will become the next endless frontier. Thank you very much. [APPLAUSE] It’s an honor to be included
in this celebration of the MIT Stephen A. Schwarz
College of Computing. I’m going to talk today not
about computing, per se. I’m going to talk about why in
life, science, and health care, we need computing and
computation and technology. As many of you
know, I’m a member of the Department
of Biology, which is not a part of the Institute
that has traditionally been associated with computing
and artificial intelligence. However, this is
changing rapidly, and the interface
represents an exciting– and perhaps the most exciting– part of the future of
biological science. For example, MIT has
recently launched a new, exciting J-Clinic focused
on machine learning and health care. There will also be, this
summer in this building, a Koch Institute symposium on
machine learning and cancer. But before focusing on this
interface of cancer and machine learning and
computing, let me talk a moment about the history
of biomedical research at MIT and why it needs transformation
by the Schwarzman College of Computing. When I arrived at MIT in 1974
in the newly-opened Center for Cancer Research, I
discovered a computer in the corner of the
laboratory of Dave Baltimore. But in the next 10 years, I do
not think it was ever touched. Many experiments were underway,
and research discoveries were recognized by
four Nobel prizes, but not one involved
the computer. In a related story that
Eric Lander told me, when the Whitehead
was being designed, there was not a room or space
set aside for a computer until Eric suggested
that it might be useful at some stage
in the future to have a computer in the building. All of this changed in the
1990s with the commencement of the Human Genome Initiative. The genome was completed
in 2003, cost $3 billion, and brought biomedical science
into the IT and computer age. The MIT community led this
genomic transformation as it also led the development
of biotechnology in the 1970s. We’re now in the age
of biotechnology, particularly here
in Kendall Square with the most concentrated
biomedical research community in the world. What is shown here
on this slide is MIT in the red buildings
along the Charles, and then the
light-blue buildings occupied by biotechnology
research labs with greater than 50 employees. There are probably
another 100 companies located in this space that
are smaller than that. This community has been
labeled “the most innovative square mile on earth.” The products of
Kendall Square are improving the lives of
patients with genetic diseases, spinal muscular atrophy, for
an example, as well as cancer. Amazing as this complex
is today, biotechnology and all of life
science, in my opinion, will be transformed by
artificial intelligence and machine learning. Amazon, Google,
Microsoft, IBM, and others are all locating
research facilities in and around Kendall
Square with a primary focus on health care. We need their expertise and
market power to flourish. But MIT and the
Schwarzman College will be an essential part of
making the transformation. A significant challenge to
the future of this innovation is the growing cost
of health care. The red-dash line on this
graph is the percent increase in the cost of medical
care from 1980 to 2010. It has increased real
purchasing power 2.5 fold. Note that the dashed
line just below that is the increase in purchasing
power of the top 5% wage earners in this country,
and the lines below that represent the percent increase
in income of other lower wage earners. Clearly, this separation in
the cost of health care from the purchasing powers of
families cannot continue. We, as a nation, are spending
18% to 19% of the GDP on health care, and greater
cost will have really significant economic
impact and social impact. Proprietary pharmaceuticals,
such as patent drugs, accounted for about 12% of the
costs of health care in 1980, and it accounts for
about 12% in 2010. So this has been part of
the increase in health care, but it does not account
for a major part of the increase in cost of
health care in this country. The cost of every segment of
health care has increased– physician time, hospital beds,
record keeping, reimbursement, et cetera. We need an
across-the-board increase in the productivity of health
care, better quality health care per dollar of cost. High-tech and computing has
delivered this transformation in every sector of society
except in health care. Hopefully the incorporation of
new technologies such as these will reduce this rate
of growth in cost while sustaining
increased quality. The promise of the
addition of data and IT extends across the
board from genomic data to access to patient behavior
in patient environments. We’re in the process
of collecting genomic data showing the
inherent risk of many diseases. It now costs $300 to sequence
a human genome, not $3 billion for the first genome. Medical records,
lifestyle data will reveal early signs of illness,
pathology, blood assays, imaging all with artificial
intelligence and machine learning will precisely diagnose
an illness at early onset. The integration of
these data will also be the basis for prescribing
treatment paradigms, when to treat, how to treat,
and when not to treat. This is the heart of what
is described as precision medicine, and the same
innovation, same information, will drive discovery
in medical research. When you know where
to look, discoveries are very likely to follow. The above data will have to
be shared with physicians and patients respecting their
privacy in trustworthy ways– in ways that’s understandable
by both physicians and patients. All of this depends
on new technology. Having employees and
patients and citizens trained are aware in biomedical
science and computer science and being able to create new
algorithms that seem mysterious to me sometimes. One outcome of
this transformation is improvement in
diagnosis of early disease. The impact of this
will be astounding. I use here the example
of a terrible disease of pancreatic cancer. If you diagnose pancreatic
cancer at stage one or two, a patient has a 55% chance
of five-year survival, and these odds are improving
as we improve our treatments Justice Ruth Bader
Ginsburg exemplifies this. If a diagnose happens
at stage three or four where 80% of all
patients are diagnosed, the chances of five-year
survival falls to less than 5%. Unfortunately, Patrick
Swayze is an example. Thus, advancing
control of cancer is best approached by devising
means for diagnosing it early, improving treatment at
this stage for cures, and an essential part of this is
the integration of patient data to predict risk, developing
better imaging technology to identify the cancer before
it gets large enough to spread, and that’s machine learning
as you will hear today. My sincere expectation
is that MIT has taken a strong
lead in this space through the establishment of the
Schwarz College of Computing. Its research and
educational effort will accelerate
the transformation of health care and
health care research to the benefit of
patients around the world. We sincerely need this college. I am so excited to be
part of this celebration. I thank Mr. Schwarzman and the
family and MIT for the vision to launch the Schwarzman
College of Computing. Thank you. [APPLAUSE] Hello, everybody. Thank you very much. It’s a very exciting
time for MIT, and I’m really glad
to be part of it. Computing is making inroads
to every academic discipline. It’s redefining
disciplines, creating very exciting intersections. And one such intersection is
computation and economics, and that’s what I’m going
to be focusing on today. So as we’re all experiencing,
advances in computing are transforming
myriad marketplaces. Instead of the brick
and mortar marketplaces that we’re all used to, we
now have online markets, which not just act as platforms
for bringing many more people together, but also
they offer a wide-array of new goods and services. The challenge is that for
these marketplaces to work, we need not just better
technology, better hardware, software algorithms,
but we also need the design of economic
incentives to move in tandem, and I’ll try to make
this point by providing two examples to you. One of the iconic examples
of digital markets is online advertising markets. Two of the top technology
companies, Google and Facebook, owe virtually all
of their revenue to their ability to
monetize their platforms through advertising. The chart here shows the
advertising revenue of Google over time since 2001. It has reached, in
2018, about $117 billion amounting for 72% of the
revenue of the company. Much of this revenue comes from
what we call the “sponsored search auctions.” So when you enter a search
keyword into a search engine– and in this case,
for some reason, we entered “used
cars in Bristol”– it results in a page
like this, which comprises a list of
organic search results that you see in the bottom. These are results that are
related to your keyword and basically ordered
in order of relevance by some underlying
algorithm such as page rank, but also a list
of sponsored links that you see here in the red
boxes paid by advertisers. Each time you enter a keyword
into the search engine, an auction is run in
real time in order to decide which advertisers
links will be shown, in what order, how much are
they going to be charged. Think about it– in real time. That’s a lot of auctions. And how do these auctions work? Advertisers– let’s
say in this example, some car dealer in Bristol– submits bids for
keywords– let’s say, used cars– and says,
I’m going to pay per click. I’m going to pay $1 if any
user clicks on my link. And then the way these
links are allocated is in the order of the
bids that are submitted. The highest bid
gets the top slots because they’re more valuable. Why? Because users scan the
page from top to bottom. So how do the payments get done? I just said $1, right? Very small amounts. And from these
very small amounts basically emerged a
very huge industry. As highlighted by the chief
economist of Google, Hal Varian, most people don’t
realize that all that money comes pennies at a time. So let’s now look at how
these small amounts are made. Sponsored search
started in mid ’90s with the so-called
“first price auction.” And this is basically
bidders exactly pay what they bid in the scheme
that I just told you about. A drawback was
quickly discovered. Why? Because if you
pay what you bid– you say you’re going to
pay $1, and you pay $1– what happens is if you lose– if you do not win– you’ll have an incentive to
keep increasing your bid. And if you win, you
have an incentive to shave off your bid next time. And this was actually observed
very clearly in data– led to a very unstable
price dynamics with huge revenue consequences. So instead, in 2002,
Google deployed a system which was inspired by the
so-called “second price auction” in which the bidders
pay the second highest bid instead of what they bid here. And let me show you how the
second price auction works. This is a very simple example. We have three links to
allocate, and the three highest bidders will get those links– in this case, $10,
$4, and $2 bidders. And let’s see what
they’re going to pay. The highest guy will
pay the second highest bid, which is $4. The second one will pay $2. The next one will pay $1. This is basically
what we mean by a generalized second-price
auction exactly deployed by Google. You immediately see a problem. If actually the top
guy bids $3 instead, he still gets a slot– the
second one in this case. But then he pays the
second highest price, which is $2 instead of the $4. So he increases his pay off. So that means that he has
an incentive, actually, to game the system. This is the problem with
the generalized second price auction. It’s open to manipulation–
another form of instability. I hope that these examples
highlight the costs of not getting incentives right. We need not just better
technology, but also better incentives. And in fact, many
companies are now moving towards
designing much more sophisticated
incentive-compatible, fast, scalable auction design. Let me turn to another
marketplace, a much bigger marketplace– finance. Finance is as old as humanity,
but in the modern economy, it’s done not through
moneylenders but very high-tech banks and computerized
platforms. What that means is a huge
increase in interconnections. Basically the exposure of
many of the global banks to each other’s
assets and liabilities skyrocketed over the
past several decades and actually reaching
an all-time high before the global
financial crisis of 2008. Accompanying this
is also a huge rise in the complex
financial products, such as derivatives contracts. So interconnections are good. They enable you to
do risk sharing, better matching of demand to
funds, but at the same time, also has unintended
consequences because they act as conduits for spreading
the financial distress as the world found out in 2008. So more interconnections
are more dominoes exposed to each other. And actually, research
shows that the extent of the spread of distress
very much depends on the underlying connections
between these financial institutions, size
of the shocks, and the incentives of these
financial institutions to insure themselves against
these financial distress. Technology not only enables us
to design better marketplaces, but also trading. Enormous amount of
resources are placed into fast algorithmic
trades in order to beat the market by
a few milliseconds. But of course, this
doesn’t mean that it will lead to efficient
allocation of resources and actually may be a very
big source of instability again because of
false probing bids as well as very much out of
control feedback effects. In fact, one of the biggest
crashes of the stock market was the flash crash
of 2010, which was caused by miscalibrated
algorithmic trading. So again, this is
another example where we see that we have to
design technology together with these incentives. This takes the form of
policy and regulations here in order to counteract
these systemic crashes due to interconnections,
complex financial products, as well as algorithmic trading. So all of this technology
and all of these developments provide great
opportunities for us, but it could also
very much backfire if we don’t design these systems
with a proper understanding of both computation
and economics. And that was one
of the incentives for us to design a
completely new major at MIT– computer science,
economics and data science, which is a true collaboration
between departments of economics and EECS. Our goal is to educate the
next generation of students who are equipped with the
foundational knowledge to address these
exciting problems. This Schwarzman college will
be a great enabler for us to be able to realize our dream. So thank you very much. [APPLAUSE] Hi everybody. How are you guys doing today? My name is Sarah Williams. I am a professor of
technology and urban planning, but I’m so much more. I’m trained as a geographer,
an architect, an urban planner, and a data scientist. And I combine these skills to
really create data for action. And at the heart of
data action is the idea that we need to build expert
teams to work together to really capture
the insights of data, and I believe this kind
of collaborative learning is what really allows us to
make data work for civic change. So because I like to
tell stories with data, I’m going to tell
you a story today. My story starts
in Nairobi, Kenya. I’ve done a lot of
work in Nairobi. This is a typical
scene on the street. Nairobi suffers from
severe congestion problems. This is similar in many
rapidly developing cities. As a transportation
planner, one of the things that we use to address this
issue is transportation models. But we need data to
make those models, and many cities, especially
rapidly developing cities, don’t have that data. I applied my skills
in machine learning to extract roadways for the
Nairobi metropolitan area, creating the first
GIS data set, which allowed us to make this
transportation model. And while my model was
successful, one of the issues is we didn’t have
data on the matatus. Matatus are the main way
people get around in Nairobi, and people really
depend on this almost like they depend on a
public transit system. If you don’t know
what a matatu is, I’m going to show
you a small clip. Imagine riding in this. [AUDIO PLAYBACK] [MUSIC PLAYING] – If you wouldn’t
know better, you could be in a club or
a bar, but you’re not. You’re in a minibus
driving through Nairobi. And with volumes like
this, you’d better not come on an empty stomach. On the outside, they’re
painted on odd designs, and their name “matatu” is
derived from the key Swahili words for “$0.03 for a ride,”
which nowadays is more $0.30. The fares are raised
by the conductors. – [SPEAKING SWAHILI] – Traveling by matatu
is a daily reality for millions of Kenyans. [END PLAYBACK] So if you want to get
around on a matatu, you ask your matatu driver. And right before we did
our research in Nairobi, the main way you understood
how to get around is just this informal
talking with the drivers who, as you can see, really
pride themselves in the designs of
their vehicles, and it’s really something
that people in Nairobi talk about all the time. And so I thought, how could
create raw data from my model, but create data that
everyone could use? If I didn’t have an
ability to find out where the matatus went, the
average person in Nairobi didn’t as well. So if you know Nairobi, you
know people use their cell phone for everything. They use their cell
phone to buy coffee. They even use their cell phone
to buy a ride on the matatu. So we decided, how
could we leverage this ubiquitous technology to
collect data on this system but open it up
for anyone to use? We developed an
application in coordination with the University
of Nairobi, and this was so that we would provide
context for the Nairobi metropolitan area, but also so
that the technology we built would retain and remain in
Nairobi so long after we left, this work could be built upon. We collected the data in GTFS. How many people
know what GTFS is? One. There’s always one person. You probably all use GTFS today. It’s what allows you to route
yourself on Google Transit. So if you’re using Google
Transit, in the background is GTFS. I note this as importance
because by making it in a open data format,
it was instantly usable by a lot of different
software companies– so not just Google. GTFS is a text file
that has basically latitude and longitude points
that gives us a schedule. And as we collected the
data, it came streaming in collecting the routes. But these routes were,
combined, hard to understand. How do we visualize it? How are we going to
communicate this information? We ultimately decided
to develop something that looks like a subway map
that you might see in New York, London, Paris, even Boston. And this map was developed
in coordination, not just with our team in the
University of Nairobi, but also the matatu drivers,
owners, the government, and this is a idea that
collaborating with data allows people to
trust the results. And here, you’re
seeing them actually noticing that there’s a lot
of missing routes in the top. They’re instantly
using our map to create new plans for the city. The maps went viral
on the internet. We got them published
in the papers so people who don’t
have cell phones could access the information. And what I like to talk about
is, how do you measure success on a open data project? And that’s when other
people leverage your data for their own change. So in Nairobi, I
was really excited when the government invited
us to a press conference and made the map an
official map of the city. So while we were
largely disinterested through the project,
they felt that they could trust the data set. And now, it’s the
official map for the city. Google has put the data
into Google Transit. It is the first informal
transit system navigable through Google Transit. The World Bank copied
our visualization to get support for
their BRT bus line. This is actually the World
Bank’s map of BRT, not our map. It looks very
similar, doesn’t it? And there are now
five apps in Nairobi that use our data as the base. Our collaborative
process really helped us make relationships with
Nairobi’s strong technology community. And we have continued to
teach classes, do research with that community. A recent class where we
had a World Bank policy expert embedded in
the class collaborated with [INAUDIBLE] Reroute, which
is a Waze-like app in Nairobi to perform semantic
analysis on text streams from Twitter to
map where crashes were happening in
real time in Nairobi, and that came from
another class. So semi-formal transit provides
mobility around the world, not just in Nairobi. The majority of countries have
this kind of transit system. The work in Nairobi
inspired Amman, Managua. In fact, 26 different
cities have used our tools and have become
part of our network in developing data sets
and maps for their systems. This has caused us to
create a global network for mapping transport data
where we can provide resources, keeping people to
the policy, but also help them scale their work–
create continued involvement. We just launched the Africa
resource center last August, Latin America in January,
and I hope the work that I showed you
today shows how a geographer, urban planner,
architect, data scientist, can teach the students
of the next generation how to combine these schools
to create civic change. Thank you. [APPLAUSE] All right. Hello, everyone. I’m Vivienne Sze. I’m a faculty member
in the EECS department, and I’m going to talk to you
about energy efficient AI. So today, most of the processing
that’s being done for AI happens in the cloud, but
there’s many compelling reasons why we want to move it out
of the cloud, into the edge, and process it locally
on your device. So the first thing
is communication. So if we really
want AI to be used by or accessible to many
people across the world, we need to reduce the
dependency on the communication infrastructure– so bring
it directly to the person. We saw this morning,
also, there’s a lot of applications of AI
in the health care space. So privacy is also really
important in terms of the type of data that we’re collecting. So again, maybe you
want to keep the data on the actual local
device– preserve privacy. And then finally, there’s
a lot of applications that involve interactions
with the real world and where you don’t want to
have a slow response time. So a typical example of this
would be self-driving cars. So imagine if your car’s
going very fast on the highway and you’re trying to
avoid a collision, you might not have time to
send the data all the way to the cloud, wait for it to
be processed, and then pushed back out to the car itself. So latency is
another reason why we want to do the
processing at the edge. But there are
challenges involving moving this computing to
the edge itself, primarily power consumption. So for example, if we take the
self-driving car as an example again– so self-driving cars consume
over 1,000 or 2,000 watts for just the computing power to
just crunched the data of all the sensors that
it’s collecting. So that’s a challenge. And then if we think about
moving this compute onto a smaller device– so let’s
say a handheld device like your phone or these
smaller robots– the power challenges
are even more stringent. So for example on
these small devices, you have very limited battery
capacity because of the size and the weight of the
device, so you can’t have too much energy there. And then also, if we take a
look at the existing embedded processors out there,
currently, they consume, in order of
magnitude, more power than what is allowed on
these handheld devices. So typically on these
handheld devices, you can only afford about a
watt of computational power. So if we take a look
and if you think about how we have dealt with
this over the past few decades, typically what we would
do is we would just wait for Moore’s Law
and Dennard Scaling to give us faster, smaller,
and more efficient transistors. But this trend has really slowed
down over the past decade, so we need to think
of something else. This is not going to
be a solution that will carry us forward. So in our group, primarily
what we’ve been looking at is, how do we deliver
energy efficient AI through cross-layer design
all the way across the stack. So what does that mean? The first thing is we want to
develop new algorithms that are energy efficient. So we really want to
think about the energy consumption of the
algorithm in addition to the accuracy of the algorithm
and how these algorithms might map onto hardware. The second thing
we need to do is we need to build more
specialized hardware and redesign the
computers from the ground up really targeting AI. So this means new compute
architectures and new circuits. And then finally,
it’s really important to think about how this computer
hardware would be integrated into an actual system. So both the sensing or
actuation, if you’re a robot, are also important. So you want a holistic
solution in terms of reducing energy consumption. So now I’ll tell you a
little bit about couple of the projects that
we’ve been working on. So the first is, we’ve worked
on building efficient hardware for deep neural networks. So if you’re familiar
with deep neural nets, it’s used for a wide
range of AI applications. Today, it delivers state
of the art accuracies. People are very
excited about that. In terms of developing
specialized hardware for it, what we actually
really focused on was reducing the cost
of data movement. So as it turns out,
it’s not really the computation like
doing the multiplies or adds that’s really
expensive, but it’s how you move the
data from the memory to the compute engines, which
is consuming a lot of energy. And so we really designed
a specialized hardware named “Iris” that focuses
on minimizing this data movement so we can drop
the energy consumption. So as a result, we can do tasks
like image classification, which is the core task
in computer vision and under a third of a watt. And in the end, if you compare
it to existing mobile GPUs, it’s, in order of
magnitude, lower in terms of energy consumption. OK. So then another project that
we’ve been working on– this is in collaboration
with Sertac Karaman who is a roboticist in
the AeroAstro department here at MIT. We’ve been looking at how do
you do autonomous navigations for these very small drones
about the size of a quarter? In an autonomous navigation,
one of the key things that you have to do before
you can actually navigate is figure out where you
actually are in the world. So that’s localization. So you can see
here in the image, we’re getting a video stream. And then on the most
right-hand side, you can see we’re trying to
estimate the position in the 3D world. And this localization is the key
step in autonomous navigation. And with this chip [INAUDIBLE]
that we developed together through the co-design
of both the algorithms and the hardware, we’re
able to do this processing in under a tenth of a watt–
so around 24 milliwatts. And so where can this
actually be used? Well actually,
there’s a whole class of low-energy
robotics out there, which take less than a
watt to do actuation. So for example, you can imagine
these lighter-than-air vehicles that can be used for
air quality monitoring or miniature satellites that
you could use for deep space exploration or origamian
and foldable robots that you can use for
medical applications. So all of these robots
take very little energy to actuate and interact
with the real world. And so it’s really important
that the computation is also very low power. Another example that
I want to talk about is some work I’m doing with
Thomas Heldt in Institute of Medical Engineering
and Sciences here at MIT. And really what
we’re focusing on is looking at the role of energy
efficient AI in the health care space. In particular, we’re
looking at the monitoring of the progression of
neurodegenerative diseases, which currently
affects more than 50 million people worldwide. One of the ways in which people
are assessed for dementia, let’s say, is that they
have to go into the clinic and talk to a
specialist and they’re asked a series of questions. And the issue with
this is, first of all, it’s very expensive to do this. It’s very time consuming,
so people can only go maybe once or
twice a year at most. And then third, it’s also a
very qualitative and subjective assessment. So for example,
different specialists might have different conclusions
in terms of their evaluation. And even a specialist
themselves, the repeatability in terms
of their testing might vary. What’s been really exciting
is that recently, it’s been shown that there is
some correlation between eye movement and these
types of diseases. And so if you can
measure the eye movement, eye movement will
give you a much more quantitative evaluation of
the state of the person’s mind or the disease
progression or regression. And this could be
useful in terms of evaluating whether or
not a drug is working. But again, the challenge right
now with these eye movement assessments is that you have
to go into the clinic to do it. It takes really
expensive cameras, and it’s quite inconvenient. And so with Thomas,
what we’re looking at is whether or not we can
integrate the eye movement tests onto a smartphone
itself so then you can bring it into the home. It’ll be very low cost, and you
can do frequent measurements. And this would be
a good complement to the specialist’s assessment. So in summary– oh,
and so also for this, it’s really important to do
the processing on the device because obviously, it’s
medical information. So in summary, I
think energy efficient AI is really important. It really allows AI to extend
its reach beyond the cloud. What it enables us to do is
you can reduce your reliance on the communication network. You can enable privacy. You have lower latency. And so you can use AI for
a broad set of applications from robotics to health care. And in order to enable
energy efficient AI, what’s really important is, you need to
have kind of cross-layer design from algorithms all the way
down to specialized hardware. And by having
specialized hardware, we really believe this will
enable the progress of AI over the next decade or so. Thank you very much. [APPLAUSE] Good morning, everybody. My name is Munther Dahleh. I’m a faculty member in EECS– also the faculty
director of the Institute for Data, Systems, and Society. And today, I want to
make a few remarks about the economics of data. To put this whole
thing in context, I think it’s important
to sort of summarize the evolution of computing. So if you think about it
back in the ’50s and ’60s, we relied on mainframes to do
centralized heavy computing that allowed us to simulate
very complex problems in terms of the weather phenomenon, or
transportation systems, energy systems, and so forth. Several decades later, I
would say ’80s and ’90s, we started talking about
mobile communication. And communication enabled
distributing the computing across many agents,
but allowed us to collect enormous amount
of data about these agents. And now, we’re in an era
where these mobile computing devices are embedded
in physical systems where people not only
are collecting data about their
surroundings, but they’re making decisions with respect
to these surroundings. And that, for example, changed
the way many infrastructures are operating. One example would be
the transportation systems were obviously
getting from bad to worse. Demand is increasing. We cannot supply more roads
to deal with these demands. And so the only hope we have
is to manage the information, to manage the data,
incentivize people to do the right thing in
order to minimize these kind of congestions. So there’s a change
in the paradigm in terms of research
and education where we used to think
about physical systems and engineered systems
on one side and people and institutions and
on the other side. And now because of this rapid
computing and decision making, these two words have connected. And research, in
terms of understanding whether you were thinking
about infrastructures, you’re thinking
about voting systems, you’re thinking
about health care, you have to think
of these things in an integrated way that– [INAUDIBLE] 614
program is one that addresses this kind
of a challenge, but also IDSS PhD program
also is structured to enable the bilingual
student to be able to think about those problems
together in solving some of the societal challenges. And you can see that this sort
of decision making over data is exploding in terms
of the number of papers, in terms of the funding
that is going there, in terms of the
revenue generated, and certainly, the skill set
that is needed from students is entirely around data. Whether it’s actually processing
the data, machine learning, AI, you name it, it’s
all about data. So if that data is
so important, how do we begin to think
of it as a commodity? How do we think about the
value that data provides? And I like this quote that
says that “personal data is the new oil of the internet
and the new currency of the digital world.” If it’s a currency,
what is its value, and how do we quantify that? I have to say that many
of the data companies are struggling
today figuring out how to get customers
because customers are not clear on the value
that the data provides. Bloomberg and Thomson Reuters
use fear tactics to say, well, your competitor has
bought this data. Why don’t you buy this data? We are very fearful of things
like 23andMe providing our data to come back and bite us in the
future in terms of insurances and God knows what if
your data turns out to have something
that is not desirable. We cannot yet figure out the
impact of Cambridge Analytica, or for example, the 150 million
financial data sets that has been made public by
the Equifax breach. How do we quantify
this kind of a breach? We still have a hard
time understanding this. So we need a marketplace–
a place where we discover the value of data. And [INAUDIBLE]
saved me a minute because she described
the whole ad market. I don’t know how many of
you would have imagined that we have over $200 to $300
billion market in advertisement 30 years ago. I think all of us would
think this is insane. The most interesting thing
about the advertising market is that it’s entirely
predicated on your data. At the same time, you are
not part of that market, and you do not have a
choice of how your data is being processed in that market. And what we want to do is
bring that privacy back to the consumer– retain that choice to the user. Creating a market in which
the data value is discovered is going to be beneficial
in many applications– transportation, energy, the
future of the integration of all the electrified
cars into the market, and understanding the
processing of all the stories that we have, logistics,
fraud data, and what have you. It’s an amazing space that will
allow this sharing of data. A few bullets to tell you that
this problem is challenging. We like to work on
challenging problems. It’s challenging because data
is not a real commodity– it’s a digital good. It has zero marginal
cost for replication. It’s very difficult
to authenticate. If a buyer comes
into to buy data, they have no idea
what they’re buying. Typically, a buyer should
come in to buy some value. They want to do something with
the data– not to buy data. How does a market
match the two together and keep the privacy of the
different data distributors? How do you
authenticate the data? And finally, how do we deal
with externality because if I sell my data to two different
competing companies, the value of my data drops. Not only we’re working
on understanding the mathematics that go with
this, which require economics, algorithmic game theory,
optimization, machine learning, and a lot
of information theory, but also, we’re trying to apply
it in places where it matters. One project that I will
tell you that in IDSS that we’re taking
on is empowering farmers in sub-Saharan Africa. The problem that farmers– right now, they are very
worried about getting credit to upgrade their operations
because they’re poor, and they’re worrying about
mortgaging their land. What we want to do by
creating a data market– a sharing market
between farmers– is to be able to extrapolate the
value of introducing technology and the gain that
they can have so that they can go
and negotiate better terms with these creditors. We are setting up
this whole ecosystem around this platform for data. My feeling is that creating
these kind of platforms where data is shared incentivize
people to contribute good data because they can
retain the value themselves but also can benefit
from everybody else’s, and allows us to address
critical and societal challenges. Thank you for your attention. [APPLAUSE] Thank you. Over the last few
months, a number of organizations
and conferences have marked the 10th anniversary
of the worst financial crisis in our lifetimes. And in the aftermath
of that crisis, we passed the
Dodd-Frank Act of 2010– a sweeping piece of
legislation that completely reconfigured the
landscape of regulations and financial ecosystem. This act was 2,319 pages long,
contained hundreds of new rules and regulations, thousands
of conditions and exceptions, and it was written
by many authors– a number of whom had no idea
what the others were doing. It’s just too complex for
any one person to comprehend. And as a financial economist
interested in regulation, I really had a hard time
making heads or tails of it. Did the financial reforms
deal with the issues of the financial
crisis, or did we go too far or maybe not enough? What do these questions
have to do with computing? Well, computer
scientists and engineers understand complexity deeply. For example, the Google search
engine, or the Linux operating system, or the Mozilla
Firefox search engine are all examples of complex
rules and regulations that contain hundreds of various
different laws, thousands of conditions and
exceptions, and were written, in many cases, by
authors working entirely independently. What if we use the principles
of good software design to analyze the
financial regulations? What would we learn
from that process? Well, in order to
tell you about that, it’s not as crazy an idea as you
might think because, after all, the laws of the land are,
in fact, the operating system of our society. So if you take a look at all
of the laws of the US Code, you print them out, it would
be many volumes on a law school library shelf. But the fact is that
over the past 80 years, if you took the annual
versions of the entire US legal code, they would
actually fit on a thumb drive. Not really that big
a data set at all. In fact, the 2014 vintage
of the US legal code is over 200,000
pages long, but it’s only 1.8 million sentences– about 41.4 million words. And that actually
compares pretty favorably with the kind of complex
code that computer scientists and engineers use all the time. So let me tell you a little
bit about the legal code. First of all, like most
pieces of complex software, it’s actually divided up into
discrete units called “titles.” The current US legal code has
53 titles, such as Title 12– Banks and Banking. That’s the part
of a US legal code that’s most relevant for
financial regulation. Title 26 is the Internal
Revenue Code– something we’ll all be dealing with
in a couple of months. Title 35 is the Patent System,
and so on, and so forth. And each of these
titles is divided up into even smaller sections. Now is this collection of titles
and sections well-designed? Well, if you talk to computer
scientists and engineers about what makes a good
piece of software– and I’ve done this. A number of my colleagues
here in this room have helped out with this– they point to the following– the five Cs of
software development– conciseness, cohesion, change,
coupling, and complexity. In the interest
of time, I’m only going to focus on one of these– coupling. Coupling is an
idea that has to do with how different parts of
a complex piece of software interconnect. And the best way
to understand it is through a pretty
simple example. Imagine you have a
piece of software that’s comprised of different
discrete components labeled A, B, C, all the way
through I, each of which engages in certain kinds of
computations based upon inputs that it takes in and
outputs that it provides. And the arrows between
these different components indicate whether or not there
are any interdependencies. For example, the
arrow between A and B indicates that B
calls or references A. And the reason the
arrow goes from A to B is that if you
make a change in A, that could affect the
proper operation of B. So when you make changes
in A, you better go check to see what’s happening with B. Now once you define
these nodes and arrows, you can actually define
subsets of these nodes that have special properties. And one way to do that
is using the property of strong connectedness. A strongly connected
subset of nodes has a property that you
can go from each node to any other node in
that subset by following a particular path of arrows. They’re all connected sort
of like the New York subway system. So as an example, in
this particular graph, nodes B and E are a strongly
connected subset because you can go from B to E,
and E to B. Notice that A is not part
of that subset because while you can get
from A to B and A to E, there is no way to get
from B to A or E to A. It’s sort of like the
Boston subway system, or as they say in Maine, you
can’t get there from here. Now once you define all of
these strongly connected subsets as I have here, you
can then come up with a measure for how
efficient this code is. What you do is, you
look at the largest strongly connected subset– in this case D, G, F, and H,
which we call “the core”– and you ask the question,
how big is the core? Because the bigger the
core, the more difficult it is to manage this code. Why? Because by definition, the core
is a strongly connected subset. So any change you make
in any part of that core, you need to worry
about what it’s going to do to every
other element in that set. And so by that measure, this
is a pretty badly written piece of code because the core is 44%
of the total number of pieces of code. By comparison, if you look at
Firefox across the versions one through 20, the median size
of the core is only 23%. Now across the 53 titles
of the US legal code, there are 36,670
discrete sections. And if we treat each
section as a node, and we look at cross
references from one section to the other as the
arrows, we can actually calculate the core
of the legal code. And what is that? It turns out that the core
is 6,947 sections, or 19%, which is pretty good
except for the fact that not all parts of the
legal code are created equal, and let me give
you a few examples. In 2009, US Congress passed
the Omnibus Appropriations Act, which is basically a list of
appropriations for various line items– a very, very simple
piece of legislation. Not a lot of interdependencies. So when we develop a graph
of those interconnections, we see that it’s actually
pretty straightforward. The nodes that are
colored red are the core, and you can see that the
core is a really small part of the overall set. Now the US Patent
System Title 35 is quite a bit more
complicated– many more interconnections,
not surprisingly. But actually, when
you look at the core, it’s still a pretty small
part of the overall system. Now what about Title 12– Banks and Banking– the
part that interests me the most as a
financial economist? [LAUGHTER] Yeah. This is a problem. Highly complex with
interdependencies that most people couldn’t
even begin to calculate, but that’s not the worst. Title 26– the
Internal Revenue Code– [LAUGHTER] Talk about your spaghetti code. We really need tax reform
sooner rather than later. Using this and other
measures, we actually can analyze the legal code
to understand whether or not it’s well-written or
whether there are going to be problems that arise. And in a paper that I wrote with
some students and a practicing attorney, we
developed these ideas and applied it to
the legal code, and we published it
in a law journal. At the time we
started our research, the complexity of financial
regulation and design was not a subject that
involved computing. Now, it is. This is why we
need the Schwarzman College of Computing, and I
want to join my MIT colleagues in thanking Mr.
Schwarzman and his family for this transformative gift. And I want to
congratulate and thank Dean Huttenlocher for
taking on this incredibly exciting challenge
and opportunity to launch the college. Thank you. [APPLAUSE]

3 thoughts to “Computing at the Crossroads: Intersections of Research and Education”

  1. I want to get into MIT
    currently in class 11
    and hey I know about Arduino, java, ide, raspberry, kali linux
    But in India none cares 😕😕

Leave a Reply

Your email address will not be published. Required fields are marked *