Keynote Talk: AI and Technology in Service of Society

Keynote Talk: AI and Technology in Service of Society


>>It’s my pleasure to introduce Professor Raj Reddy as the first keynote speaker
to this summit. Proffesor Raj Reddy, of course, it would be preposterous
for me to introduce him because I’m sure all of you
know and read about him. His contributions to AI and speech recognition
research is famous, and that has received a Turing Award what are
the one talked about in 1994. But I would like
to say a few things about this other engagements
that he has been. He’s the chairman of our Governing Council to play the head of
Governing Council. So, I watched him recently closely for
the last 10 to 15 years. He is engaged with lots
of education efforts later to taking the education to people through whom they don’t reach very easily
and ordinarily. He started a program called MSIT right here on our campus
and a few other campuses. This is electronic-based courseware primarily
mentor-based training of people who otherwise
won’t get into an M.Tech or program in any
of the established places. How do you bring
them and train them? That’s been going successfully, and in 2008, due to his efforts, the state government, AP government established
a massive system called Rajiv Gandhi University of Knowledge
Technologies, RGUKT. This is a rural university. They take students after the 10th class for
a six-year program. The selection criteria in the early years that are out
there and researched these, take three students from
every mandal of Andhra Pradesh. That is 6,000 students in total. There were three campuses
in three parts of the state, one in Telangana, one in Nuzvid in
a place called Basar, a little far from here, and two in what is
now Andhra Pradesh, one in a place called Nuzvid, one in a place called RK Valley. The students are
almost all rural and very deserving
of complete support, and they’ve been
doing very well. After the state split, the Andhra Pradesh has doubled
the number of campuses, there are four campuses now, and they take thousand students in each of the campus per year. So, it’s a massive outreach, I mean reaching out
to the students, and he tells that
the statistics, when we talk about [inaudible]
variable the gender ratio. These students do
their normal admission process. There are 58 percent
women students in these universities, 58 percent. And maybe 60 or 75 percent
are all rural, and I’ve seen him
spend a lot of time, his passion for reaching
or taking education or technologies to the people who are really
underserved is unmatched, and that’s one of
the signs of his humanity. And his talk here
also is related to AI in the service
of humanity. So, we are all eager to
listen to you, Dr. Reddy.>>Good morning. It’s a pleasure
for me to be here. I come here regularly, but this is the first time
I’m in this auditorium, I didn’t even know it existed. And there are many things
I would have talked about, but today, because this
is an AI conference, and AI academic research summit, future with AI, I
thought I’ll talk about the role of AI
in service of society. And I use that phrase
AI and technology. It turns out, anything
we try to do in AI really needs a lot of support from the rest
of computer science. Many of the breakthroughs
you’ve seen in AI in the last 10 years, 15 years, I’ll talk
about them in a minute, would not have happened
if we had not had a million times
more computing power today than 30 years ago. And many of
the AI ideas including deep learning and
back-propagation were invented in early ’80s. In fact, I’ll come back to them in a minute. I think almost everything we do in AI needs
a lot of technology, and a little bit of AI
then goes a long way. Without it, and that’s what
happened with Deep Blue, and that’s what happened
to a lot of other things, you need an integrated approach
to systems and solutions. The first question I was asking, in the last 15 years, I’ve been spending
a lot of time on what can technology do to help people that are least
able to help themselves? People at the bottom
of the pyramid. In that space, we had
a conference sponsored by United Nations and World Bank in 2002 in Bangalore
called ICTD, ICT for Development in
developing countries. There’s a proceedings
that came out, I’m sure it’s available
on the web somewhere, but what I want to
kind of ask a question, I’ve been asking
this question from my own understanding
of what we mean, how we can help society. There are many different,
if I ask each of you to go and say spend
the next five years thinking about how you
would serve society, technology and AI
would serve society, you would come up with
different answers. And I’m sure all of
them are relevant. The question is, are
there one things that you can do that will
be truly dramatic, that will kind of scale to the entire population
of the planet. Everybody on the planet
will benefit from it. Scalability was one
of the issues that is being a concern to me. First, I started early on
looking at human rights. The United Nations in 1948 published this Universal
Declaration of Human Rights. There were 30 different rights that most of the
countries signed onto, almost everybody including
Russia and China and everyone, except not all of them, even
though they signed them, they don’t follow them, right? The question I was asking was, can I use technology, AI, to detect violations in society anywhere in the world of any
one of these things. The answer is yes. Hopefully, by
the end of the talk, it will become clear
what I mean by yes, but this was too big a list, too kind of defocused. Sometimes it’s not
clear what it means, right to a cultural
life, number 27. I don’t know what I
would do if I had that. But I guess if we
investigated it further, because the United Nations under the chairmanship of
Eleanor Roosevelt, created this in 1947-48 and they debated it for several years before they finally
announced it. In particular,
the constitution of India also kind of took
three years to create, but the constituent assembly debated a lot of these things. So, you’ll find
many of these things have gone into
the constitution of India. I was not quite
happy with the list. I was at the United
Nations Summit in Johannesburg called
Sustainability Summit, and they came up with
a smaller list of 15 or so-called Sustainable
Development Goals. If you go to any UN conference
these days, SDG, everybody talks
about SDG because all the agencies
have to push SDGs. The problem is some of these things like end poverty
has been a major goal. And many of the other things, food security, health, education,
gender equality, water, sanitation for all, gets
closer to the kinds of things I’ve been more
comfortable with. Even here, there is a book by Jeffrey Sachs called End
of Poverty or something, and I was very hopeful
when I bought the book, thinking maybe
I’ll see something. What I found is, both the previous one
and this one are created by rich people who don’t really understand
what it means to be poor. And they create their version of what the whole rights
of human beings are. But if you’re really poor, and you really
don’t have anything, equality may not mean much and lots of other things
may not mean much. You want some food and water
and a few other things, what I call basic
necessities of society. Before anything, you need water. If you look at
the first human right, it doesn’t say anything
about water, right to water. I’m saying we need to look at the most basic things
every person needs in the world. So, this is my final list. We may need to cut
it down even further. Some of them are not in the previous two lists
like housing. I don’t know how
many of you heard about Roti Kapda aur Makaan
as a slogan, right? And those are the basic needs. Everybody needs some clothing. Nobody talks about clothing. This is the list I started with. The question was, can AI help? Can technology help
in all of these? The answer is absolutely, and this is what I hope to
convince you to think about. So, basically, I just wanted to take a few minutes to revisit
AI in the last 50 years. And before I say that, I want to kind of motivate
why I’m going there. Because in order to help society on the basic things
like water security, or food security, and so on, we need to know what is happening and we need to be able to monitor
what’s going on. And the question is,
do we know how to do that for all the
seven billion people? And I’m going to try
and convince you. We have the beginnings of technology where
we can monitor it. And you’ll say, “No. That’s too big, and
it’s going to be violation of my privacy
and all kinds of things.” We’ll come and visit
that topic also. But if we can monitor what
everybody in the planet is doing every minute or
every five minutes, it’s a huge amount of data. You might collect as much as a gigabyte per person per day, and you multiply that by seven billion people
or 10 billion people, we are talking about
exabytes and maybe more. And we shouldn’t be scared
about it because we have million times more storage than 30 years ago
for the same cost. I paid two million
dollars to Dark in 1972 to buy 40
megabytes of memory. You can buy that for
a few pennies today. So that’s what the
change that has happened in the
last 40, 50 years. So, to revisit quickly, we have AI, and people have used many different terms like
intelligence and so on. But to a large extent, when we look back at it today, it is expanding the scope of what can be automated
using computers. That’s another definition
you can get namely. Normally, computers
were designed for number of crunching mathematical
computations and so on, but there are other things like playing games,
or proving theorems, and later, in understanding language
as understanding speech. And so, there is
a hierarchy of AI stuff. There are sensory
perception type things, perception, speech, vision,
speaking, and language. Then there is learning. We all learn,
human beings learn. The question is, can
computers learn by themselves? Can they program by themselves? And then comes, can they reason? Can they plan, and
reason, and think? And finally, can
they do, or act, sense, think, and act, mentor, monitor, diagnose, and repair? That is the mantra of AI. If you accept
that mantra, and say, I am going to be able to
monitor every person in the world of
what good and bad things that are happening
to that person, then I can build
intelligent agents, what I call Guardian Angels, which will at least
warn them, saying, “You’re going to
have water scarcity in the next month or so. You better get ready and
practice conservation. Do various other things, and for the following reasons.” Not only that, it would say, “If you were living in a desert, this is what you would do because there’s no water there.” People living in a desert. They’ve practiced
a very different kind of thing than those of us who
are living next to rivers. For us, water is plentiful. We never even think about it. But if you’re in larger sun
or in Sahara desert, water becomes the most
precious resource. And so, the goal of AI in helping society broadly
for me has turned out to be, “Can I monitor
everybody to figure out what the hell is going
wrong, may go wrong?” Because they’re not in
a position to understand that something might happen because these are
unpredictable events, and warn them upfront. And these are what I call Intelligent Agents,
Guardian Angel technologies. And so, the issue of
how do we build those, where do they come from, how do we, what are
the architectures. So, basically, going back, I think we skipped one. Basically, what I’m saying
is, not just intelligence, it would be intuition, it could be creativity. There are people that
have created programs that will synthesize music
like Bach or Bath Elwell. There are people that
have created paintings. Harold Cohen was
an early adopter of AI technology for painting, computerized painting,
sketches and painting, and innovation, and
emotion, and empathy. If you haven’t read the book by Minsky called
Emotional Intelligence, where he kind of goes
through and says, “Emotion is not that much
different than other things, and you just have
to think about under what conditions do you
trigger emotional response. What are the causes of it, and how do you kind
of reason about it?” So, all these things, not just into AI, really encompasses of
many different aspects of how human beings act
in solving problems. And all of this is, I’m saying is, everything
that we do can be automated. That is the thesis of
the founding fathers of AI, Newell, Simon,
Marvin Minsky, and McCarthy. And I subscribe to it also. That is, if you
come back and say, “No, you cannot do this.” Then it becomes
a scientific experiment to me. Basically, I would say, “Okay, let’s kind of lay down
an experimental design of where we can prove or disprove the hypothesis
that this cannot be done.” And if you have to
go back and read the original paper by
Alan Turing about intelligence, he’s very clear about it. When he talks through
each of those things, he tries to explain why he thinks intelligence
can be automated. And it’s very interesting
to go back and read papers that were written
50, 60 years ago. Earlier, we tried to
prove theorems or play chess and understand
language speech and so on. And later systems, we were
dealing with most recent one, that’s very exciting
in the last 10, 15 years, is automated trading
of stock market trade. And these systems by
definition are not perfect, just like human beings
are not perfect. We make mistakes. And
so, automated trading, most of the time works well and make a lot of money
for a lot of the people, but once in a while, it gets into trouble. And I think it was
April or May of 2010, there was what is
called a flash crash, where the stock market went down by a thousand points
in 36 minutes, it never happened before, and then it recovered
by the end of the day. And it turns out there was
a day trader of Indian origin sitting in London that kind of created literally
thousands of agents, each which was placing
different kinds of orders. One would place
an order for this, another one would place an order competing against
it, back and forth. It gets into race condition
and which led to that. It took them five years to
figure out what happened. Finally, he was arrested, and I don’t know what
happened to him more recently. But if you Google 2010
flash crash, you’ll find all the
information about it. So, automated trading,
things like that, composing music and so on. But most of these solutions
are not perfect. Remember, OCR, Optical
Character Recognition, has been a basic component
of research for 30, 40 years. And when I came
here in 1989 or 90, the Department of Electronics, IT at that time, handed me a disk saying, “We have done OCR of
all Indian languages.” They were extremely
happy that they could get 90 percent correct
recognition and accuracy. I said, “This won’t work.” 90 percent recognition
of OCR in a typical page, you might have, let’s say, 100 words, 500 characters. That means 50
characters are wrong. That means every
other word is wrong. And you would put up with it. In a publisher’s, in real books, they expect 99.999
percent correct. So, there is a lack of
understanding of the level of accuracy that’s needed
before we can begin to use it. I’m sorry to say, even today, Indian OCR is not
working government. But there are other OCR systems, the best OCR for English and other Roman characters
used by Russians. It’s called Abbyy FineReader or Abbyy something, A-B-B-Y-Y. And you can buy them. And it almost never makes any mistakes. And the Google and others also have similar OCR systems
that are pretty good. But what they actually
will notify you, that I’m not sure
about this word. And then, at CMU one of our faculty members invented
the Recaptcha Paradigm, where rather than giving you some robot things to type in, captcha, it would simply say, here is a sentence translated or give me the correct version
of the sentence. And if you give
the correct version, then you get through otherwise. The definition of
correct version is, at least two other people
have typed the same thing. If you’re the first one,
you go to them get through or some such thing. So they used crowdsourcing type thing to get
better results. So that’s kind of
an attempt to kind of give you a version of what AI might be or what you
might want to think about AI, rather than somehow it’s
replacing our intelligence, which is not, okay? And so basically,
in the 20th century, we kind of started
without a definition. We said AI system. This is from Newell,
Ellen Newell, learns from experience, uses vast amounts of knowledge, tolerate error and ambiguity, respond in real time and communicates with humans
in human language. There’s a couple of
more which had left out, but these are obvious
useful things. Unfortunately, we don’t
have a single system even today that does all of
them simultaneously. I have things that do
one or two of them, maybe three of them are,
but not all of them. And that still remains
the grand challenge, right. And so systems that
learn and use knowledge, and tolerate error
and ambiguity, and respond in real time. One of the things I was
delighted in 1999 and 2000 when, Google made it a requirement for every search response time
in 200 milliseconds or less. Until that point, nobody cared. I complained to IBM, when I was not with Microsoft, I was on the advisory board. I said, my god
whenever I turn it on, I have to wait for a minute. When I turn it off, I have
to wait another minute. And my quotation, my complaint appeared in the Wall Street
Journal they said, did you really say
that? They mistake. I said, I’m sorry, I said that. They didn’t like it, but now
after 20 years,that’s one of their main requirements
at Microsoft Windows 10, because they’re using
smartphones and other things, people won’t wait that long, so they’re kind of giving that the highest priority
on the on time. So the other concepts that
came out of the 2000 is, we understood the role
of search and knowledge. If you have knowledge, you
don’t have to do search. If you don’t have any knowledge, then you can solve the same
problem by using search. As simple as that, it’s a dual. And so, knowledge compensates
for a lack of search, search compensates
for lack of knowledge. And so we looked at
formal knowledge of different kinds and types
to make them into rules. And if you see
this pattern to do this. That’s what Feigenbaum
[inaudible] program did, when they discovered
molecular structure by looking at the patterns
of mass spectra data, that was in the 60s and 70s. And so, the idea
of using knowledge of biology or molecular
chemistry but known to people, and people seem to
be able to use it to understand how to discover
molecular structure. Can a computer do
the same thing? And it took them several people, a few many years to build it in, so that the system they
were using Dendral was routinely making
molecular structure discoveries, predicting as good as
a PhD in Molecular Biology. And that was what was exciting in the 60s
and 70s and 80s. And out of that
came expert systems, knowledge based systems,
and rule based systems. So, if you now go
to the next thing. So if you look at
the major breakthroughs that we had in
the twentieth century. And we had the world
Champion Chess Machine. And we had
mathematical discovery. We have proof checkers. We have accident avoiding car. In 95 we had a car that drove from one end of
the country to the other end. And that’s the origins of autonomous vehicles
started in at CMU in 1984, in the roboticists institute. And it’s taken that long. And then in robotics and speech, the dictation machine is
not what you see today, but at that time, real time dictation was a big challenge and accuracy
was a big challenge. Computer vision and
image processing, all of these were going on and the significant
advances were made. But they were also
making a lot of errors. It was also not as
good as it could be. Then, came the deep machine learning technologies and
in particular Geoff Hinton, who was at CMU from 80 to 86, which is where he invented the back
propagation algorithm, and went to Toronto from there, has been working on
saying human beings use this neural network and they do all these things that
we think are intelligent. If only I can build
an artificial neural network, that would learn the same way as human beings think,
I’ll be there. I don’t have to worry
about handwriting the code and everything else. So he was passionate about
not writing anything, he wanted the system
to learn by itself. And it took him 30 years. He was dedicated to
continuously keep on working, and it would work
a little better, but it would not get
as good an accuracy, as other systems like in speech and image
and other things. It was much worse
than the other system. It took a million time more computing power
and which we now had. And the first major
breakthrough came in 2010. And then Deep Learning has
taken off since that time. And that’s what I want
to kind of tell you. So AI in- the main
difference is, in the 20th century, human beings rule based
rules and heuristics or whatever into some programs
and executed them to see if it work and if didn’t
work and corrected it. The challenge now is, to get the human out of the whole problem of
creating AI systems. We’ll give you data, we’ll give you maybe some explanation
of what the data is, but you hope to
learn by yourself. And we are not quite there yet, namely, we have
a lot of systems. It’s called ‘The Fourth
Paradigm’ data driven science. There are books if you go to Google and type
‘Fourth Paradigm’ you’ll find the whole book
that Jim Gray, another Turing Award winner made popular in the late 90’s. The challenge there is to go to data-driven AI systems to solve previously
unsolved problems. The interesting thing to me is, many things that we are
now able to do routinely, I used to say as
early as 15 years ago, 2000, it will not be
solved in my lifetime. What are those? Basically, I said class language translation won’t happen anytime
in my lifetime. From any language, even
any pair of languages. Now, Google produces pretty
respectable translation from any language to any language. If you ever followed “IBM Jeopardy Competition” where the IBM, what is it that system?>>IBM Watson.>>Watson, IBM Watson beat the reigning champions
in Jeopardy, right? When you look at what they did, it goes to the item number
two on newest list. It learns, but it also uses vast amounts of
knowledge by itself, not because I
programmed it, right? The interesting challenge there, is the fact it has to
work in real-time like 200 to 300 milliseconds
before the other guys can, if they may know
the answer they can press. So, at that time it has to
not only press the button, but also it has to say how confident is it that it
knows the right answer. There are many challenges of AI that came in when
they were building Watson that were
never seriously taken into account before hand. And that fear for other things, I’ll come to that, the one
that’s kind of close to me is, we used to talk about speech
to speech translation, and speaking to a computer
like an assistant, and those of you who have been following Alexa smart speaker, and they’re now, every company is producing
a smart speaker system. Apple is producing one, Google is producing one, Microsoft is producing one, because they suddenly saw Amazon is stealing
their thunder because they were the ones that pioneered
in speech input and output. And all that Amazon did is, came and hired people that
were already trained there. They did to Microsoft and
others what they did to CMU. They came and hired all of our people and
that’s how they got started in IBM and so on. So, the idea is that a lot of
knowledge that’s available, data that’s available from many different sources
that we couldn’t even seriously think
about using suddenly becomes possible to
think about today. You have to go back and revisit any idea you might
have had saying, “If I gave you
unlimited computation, if I gave you unlimited memory, and if I gave you
unlimited bandwidth, then can you solve the problem?” That’s the trick here. So, in the 21st century
here are the, my list of what happened in the first 15 years
of the 21st century. Language translation,
speech to speech dialogue, Siri, Cortana,
autonomous vehicles, Watson, Robo Soccer,
World Champion Poker, and all of these have one
common underlying technology. That’s machine learning. Machine learning was
never seriously looked at. People were working on it. There are volumes of books by Tom Mitchell and
Jaime Carbonell and others, but they were all kind
of, when I looked at, I said they’re
toy learning systems, and never could get
very excited about it. Suddenly, given million
times more computing power, a million times more memory, and million times
more bandwidth, you need all of that by the way. If you just had
million times more computation then the other two
become the bottlenecks. You can have all that, I don’t know if you ever tried
to copy 100 gigabytes from one USB flash drive to
another one, it takes forever. Mainly because the bottleneck is the data bandwidth
between the devices, and it takes hours to transfer. So, all of these have
this common property, that they’re using
large amounts of data, and using machine learning,
in particular, deep learning to
solve the problems. But there’s a
fundamental flaw that is yet to be solved and maybe
one of you will solve it, is, if you ask
these deep learning systems, “How did you get there?” they can’t answer. This is a requirement
in many legal systems. Basically, you say, “I
made this decision,” Why? Because the computer told me
to do it. That doesn’t work. You have to be able to explain. That was one of the things about expert systems another thing. They could actually explain
the reasoning chain. How because of this I went here, and it’s a kind of a mark
of shame, mark of shame. And that gives you
the exact reason of how you got to the answer. That was not possible
and in this system, it’s not yet possible. There’s another flaw
in Deep Learning, which is, if you
just take an image, which is working perfectly recognizing and simply
add random noise to it, just random noise,
when you look at it, it looks just like
the same image. Now, you feed the same
random noise image, it goes completely, the system breaks down.
So it’s very brittle. So there are still
many things we don’t understand about
Deep Learning technology. And I’m expecting
in the next five, 10 years we’ll solve them. It takes time, and effort, and people, and we’ll get there. So, all of these, as I said, are able to do massive big data analysis
to discover patterns. One of the things, you know, if you’re a data scientist
or data analytics. I think this is
the data analytics floor. You can very quickly tell what are the topics
that you want. There are things like, are there clusters of
things that are, what we call k-nearest
neighbor clustering algorithms. Can you group, take this data and do they fall into
individual clusters, and what do the clusters mean
in unsupervised learning? It turns out, if you
were given the same data, you might ask a completely
different question. Is that an outlier, an anomaly? And if there’s an outlier,
what does that mean? To me that’s the important thing for some tasks that
I’m thinking about. If I’m trying to understand
the problem of hunger, how many people are hungry, how many people are not
eating three meals a day, I want to find the outliers
in the village. Maybe everybody is eating okay, but maybe one percent of
the people are so poor, that they’re not
getting any food then, yes, and I believe
that given the data, I don’t even need
very much data, I’ll come to that, I’ll
be able to predict. Somebody’s having
water scarcity, food scarcity, energy scarcity,
don’t have education, any of the things. It’s a question
of classification, saying this group of
people seem to have food, this group of people
don’t seem to have food, okay, and it’s exactly the same
classification you have to do, and deep learning
will do that for you. It’s essentially
a binary classification, yes or no, and works very well. And so, I’m very comfortable
to guarantee to you any of these societal
problems that we talked about can be identified. That’s the first step.
Monitor diagnosis [inaudible] , can be identified using AI technology, people
learning technology. Then we come to the second tool, and we’ll discuss that too, but let me then take you there. So the question is, we are now entering a realm, if I told you eliminate hunger, you say where do I start? I have no clue. Especially,
I don’t know who is hungry, where they are, how to find
them, or any of the things. Now, all of a sudden,
we have a technology and a system and
possible solutions where computers can solve this problem, and
that’s the beauty of it. Now, we are entering
a realm where computers will be able to solve problems that human beings
can never solve. They don’t even
know where to begin, and this is what Herb Simon
got the Nobel Prize for. Namely, one of the problems
with human beings, is they cannot
handle a lot of data, and when given a lot of data, they simply throw
away most of it, and then use one or two facts,
and do the best they can. This is called the satisficing, we call it the
satisficing problem, and the same problem
occurs in economics also, and he got a Nobel
Prize in Economics, but it is mainly when human beings are faced
with a huge amount of data, they get confused and they
don’t use it properly. So, that’s where I
think we’re entering, and this is very
fascinating to me, that given these basic things, and if I can take the entire population of
the world, population of India, a population of a state, or a population of a district, or even a village, continue
to collect all data I can. I don’t even know
what data I need to say, is this person having water scarcity or is
not water scarcity? I just collect a lot of data, and I throw it into
my Deep Learning System, and all I have to
do is label a few, saying these people have water, these people don’t have water. You figure out what are the properties that
you need to look at. You can do it,
hence, if you know, and then it will do
the clustering, you know, supervised learning, and so that is basically
the starting point. Namely, I think we understand
how given a community, whether it is one village,
or one district, or one state, or a whole
country, the more the better. If you have the entire country
where you know people in Rajasthan have
water scarcity all the time, and people in, maybe here in Hyderabad and Andhra Pradesh have no problem, and you can see whether they fall into two separate clusters. That would tell you, whether that’s a way of distinguishing people with water scarcity
of [inaudible]. Then if you find a few people in Andhra Pradesh that also
fall into that cluster, then you know there’s
something going on, that’s the anomaly of detection. So, given these things, now you can say,
I’m going to use AI. So, what does that
mean? “Smart Water Security”, “Smart
Energy Security”. That’s by just adding that word, it forces you to think
about what does that mean, and how I approach to
solve that problem. And I believe, you know, each of us will be able to
work out different things, and I’ll give you
my version of a solution. Okay. The first one is what Dr Anan then will talk about
tomorrow. AI for good. It’s doing well by doing good. You know, don’t worry about not having a million
times more computing. Don’t worry about not
having a lot of data. Even without all of that, you can do something even today, and that is essentially, you know, supposing I need
any one of these things. I have solutions for them, are products I
need in both ways. For each basic need, there’s an opportunity
for rural business. If you were to create
one as an entrepreneur, we have a T-Hub here, then you may be
able to immediately provide that solution using
“Amazon Market Place” today, and you would both provide
jobs and wealth creation. And I strongly recommend
you’d pick any one of them, if you’re looking for
an entrepreneurial things to do, but, the solutions that I
want to talk about, have some necessary conditions. That’s why I’m saying that
right at the beginning, technology is more
important than just AI. You need both of them, and absolutely, you
must have connectivity. Until 5-6 years ago, maybe even 2 years ago, not everybody had connectivity. Now at least, there’s
one company Geo, that has a national license
for wireless broadband, which is doing pretty well. I used to be very negative
about them last year, think they keep
promising, not delivering. But this year, I have
a Geo-Fi hot spot, I use all over the country when I
travel and it works fine. And so, in that sense,
we need connectivity. There are other initiatives in the country under
Modi’s nine digital pillars, one of the pillars
is connectivity. And there’s a thing
called BharatNet, which is trying to connect. They’ve been promising
it for years. Before them,
the previous government, there was a minister called
Dayanidhi Maran in 2006 said, “By next year, I’ll have all villages connected by
wireless broadband, Wi-Fi.” And I said “that’s wonderful”
and never happened, nothing happened. They
spent a lot of money. And the same, you need a tablet
of some sort, a device. Again, India we have
all kinds of sad stories. You might have read about
the hype around Aakash tablets. We’re going to make one for $35. I said, who the hell
cares about $35? Where did you come
up with the number. Give me a working
tablet for $200. But why are you making it? You can buy it from
international markets. The same thing
happened 15 years ago. We should learn
from our mistakes. India Ministry of Science had a startup that came out
of them, called Simputer. How many of you
heard about Simputer? Anyway, again it’s
the brilliant idea I’ve had lots of things
that we need in India, except the volumes and
the learning curve are so bad, they could never compete. The parts cost for them was
more than what you could get the whole fully built device from international
manufacturers, because they were making them in millions and these people
were making them in thousands, and it was obvious. I knew the founders. I said why are you
wasting your time? And they said, you
don’t understand, India needs India centered
solution. I agreed with that. And so, the India centered on all smartphones that
you buy from anybody, you can do Camel data entry,
Telego data entry, any other kind of
any data entry, and they’ve come up with
a reasonable entry solution. So you can send text messages in Telego if you want, right? And what was missing, is that once the technology
becomes available, so connectivity and computing
access I think are solved. If you were reading the
newspaper yesterday, news item, I’ve been telling
Microsoft for a long time, there’s something
called a Chromebook. You know about the Chromebook. Basically it’s the low end
netbook, which you can buy, which assumes
you’re connected to the internet and youl do all the computation
on the Cloud, all this is a data entry
back and forth in a display. And I said to Microsoft, why don’t you produce Internet Explorer
Book instead of calling it Chromebook?
Call it Explorer Book. Another Netbook, but
I want it under $200. If you can sell
Chromebook for $200, you can certainly do
the other one just as easily. Finally, yesterday
they announced that they’re going to
have something called Edgebook or Netbook and it’s $189 and it comes with Windows 10, especially
for schools. And I thought that was
fantastic. I’m waiting to see. So regular laptop made by Lenovo and there’s
a keyboard and a display. And if you’re trying
to learn and use, it’s not enough to have
a tablet you have to kind of [inaudible] it would
be nice to have a keyboard. So, then you also need, we had a country where
almost 40% of the people, maybe more, in different parts
of the country, the illiterate and they not only don’t know
how to read English, they don’t know how
to read any language. And so it’s not enough to say, we’re going to produce
technology and services society, AI and services society
will solve their problems, because you have to
think about how you’re going to solve the problem
of the illiterate. That adds to a whole series
of other things? And I have given them
that talk at the place in Heidelberg on wise computing? After seeing Alexa I
said, we have arrived. Basically we have
come to a point where we used to talk
about it 20 years ago, where you talk to the computer just like you talk to
your assistant, secretary. Say do this, do that,
whatever you need. You don’t type it at
them, there’s no display, they sometimes get confused, but the most human interaction
happens only by voice. If we want to work together and I want you to do something, I tell you and hopefully
you understand. And occasionally I need to see a display with lots of data, but most of the time
it was just voice. And so, these are the four Cs, connectivity, computing,
capacity building, and content. So, one of the things
you need is, for people to be able to use
your technology solutions, is all of those four. Assuming that we have them, then I’m going to
propose that we create two intelligent agents. One I call Guardian Angels. That’s the more important one. The other one is
simply Cognition Amplifiers. The Cognition Amplifiers. Okay. Two families have intelligent agents that may be able to help
suicidal problem. Cognition Amplifier is a personal enduring
autonomic intelligent agent that anticipate what you want to do and makes you do
it with less effort. The most common thing
I can think of, is paying your bills. You know, it takes you every
month some amount of time. And if a system is monitoring what’s happening in your life and knows
what’s going on, you can say, oh, this bill
came and I paid it. It was in the range of
a usual monthly bill and as you authorized me in the beginning in
the settings and I paid it. So, there are lots
of things like that. I’ll give you some examples
in the next slide, that makes it useful to have
these Cognition Amplifiers, we use to call them in
the old days, what AI was, was Intelligence Amplifier
or Augmented Intelligence. There are many different
phrases we used. Cognition Amplifiers,
just another way I use it, is to say normally you
know the Word autocomplete, and when you’re using
a computer you type some string and it comes out as
an auto-complete. How are we doing
for time? Am I okay? Five more minutes?>>10 more minutes.>>Yeah, okay. Now, I
can go on, but basically, let me speed up a little bit and then you can ask me questions. Basically, buying and
email and news and banking, would autocomplete using
a cognition amplifier. So, they would do it. Or occasionally, they might
ask you, but that’s a no, no. They should never
ask you because they’re interrupting
whatever else you’re doing. So, a Guardian Angel
is a person around and like agent that discovers and warns you
about unanticipated events. Things that you
cannot possibly know. And will try to protect you from harm when these things happen. And all I’m saying
is, given enough data, given the capability to
publish and subscribe, I can create Guardian
Angel Technologies that will monitor everything that’s going on in the world. We are talking about
a huge amount of computation, huge amounts of
memory and then figure out how it might affect me
and then provide solutions. With that, I’ll leave
the rest of it as assignments, there are another 40
slides or something. And if I can advance it, I’ll quickly go through. These are the steps
on how you go about creating one
of these things. Smart X. What are
the necessary conditions. Technologies. There
are other things here. How to create a Guardian Angel. Architecture. Always on, always
working, always learning. This is very important. If you just think about it, you have apps on
your smartphone. They only come on
when you tap on it. And then, as soon as you
turn it off, it goes away. What you want are set of
agents that are always on and are always working, thinking day and night
even when you’re sleeping, saying ‘”Oh, there’s
a tornado coming your way. “, and wake up your phone and make you
go into the basement. That’s what a Guardian Angel
should do, right. And for that, it has to
be always on and always working and always learning. It’s not only learn from
your own experiences, but also learns from the
experiences of your neighbors, your village and your state and your country
and the whole world. If a tsunami happens because of an earthquake
in Indonesia, which is what happened in 2004, 300,000 people died,
100,000 from India, in 2004, you want to be able to- somebody to know
what’s going on. You can say, “How do I create a Guardian Angel that can
track all these things?” That’s where the publish/subscribe
paradigm comes in. So, one of the things
you want is if you are a Guardian Angel for
your safety or security, may really be a hundred
different Guardian Angels. One that only tracks
earthquakes around the world and subscribes to all the websites that
know about earthquakes. What does that mean?
The websites that you create must be
machine-readable. Most existing websites are not. And one of our countrymen, Guha, has created a whole set of technologies for
machine-understandable websites. If you go to a schema.org, he works for Google and he’s
actually now left Google, and he works for all the IT companies,
Microsoft, everybody, to convince everyone to create websites that are not only
readable by human beings who use conventional
natural language but machines can also read the same thing and
interpret what is being said and use that information in the way I’m
kind of proposing. It’s very interesting set
of challenges. And Guha has been talking about it and thinking about
it for 10 years. And it’s always learning, and it’s learning because now it’s collecting
huge amounts of data. And if it can cluster
and understand the space, when a new event happens, it falls into one
of those clusters. Then you immediately
know what has to be done. Very fascinating area. So anyway, Cloud-based
architectures, how you personalize
them and goes on. And publish/subscribe. Okay, there’s one whole set of things about how
we are going to create a whole new industry of people that create
these Guardian Angels, just like you buy apps from Apple Store or
Play Store or something. And then they’re
already ready to go. You have to personalize it, just like you go to the
“Settings” and set up your email address and so
on. You have to personalize. And I say that’s not acceptable because I’m
an illiterate person. I don’t know how to personalize. So, how are you going
to solve that problem? It turns out there are
solutions to that too. But all of them require first identifying
we have a problem. Then we’re all pretty good at coming up
with various designs. What we now have
is a tool which, even if you don’t
know how to solve it, it will figure out and give you the solution saying such and such is going to
happen, do such and such. So monitor,
diagnose, and repair. And so, if there’s
a water scarcity, you predict it and then you understand it
and then you say, there’s two or three
things you can do. Conserve water, there’s water
next door or you can go buy one of these reverse
osmosis machines for home and take whatever
garbage water you have, put it through and you
have drinkable water. So, that kind of intelligent assistant
that will tell you what to do when you have scarcity, turns out to be
something we can do. And all of this assume
you have access to knowledge that most of
us cannot and do not have the time or cannot have
the capacity to absorb. But computers can
and that’s what the opportunity and
excitement is. Thank you.>>Professor you are ready?>>Yeah.>>Yeah. So, you mentioned about explainable deep learning and I hear that it’s some consortium being formed for explainable AI or
something like that. Are you aware of what
is happening or is that?>>I am not but I
am aware of DARPA. DARPA is funding
a big research project on explainable deep
learning because a lot of people want to know, and I think we’ll make
some breakthroughs. And we used to do that
for expert systems. We say, “How did
you reach this?”, and it will do the explanation. But for deep learning, we don’t have the solution. If you are working on
deep learning explainable AI and if you make a
breakthrough, that’d be fantastic. And not only fantastic, you might win a major
award or some sort. And so these days, I don’t know if
you’ve been following DARPA grand challenges like the autonomous
vehicles and so on, they’ve been issuing
other things. More and more people are issuing these
goal-directed outcomes. And Google had said the first group that
can send a robot to moon and discover the water
or something, $20 million. And I just read they
withdrew it because there nobody succeeded so far. They will probably come back
with a new version of it. These are called X Prizes, and it’s also called
Push research. Rather than where you
say, “I want this. If anybody can solve it, I’ll give you a
million dollars.” Bill Gates says malaria research is of that kind, where he says, “If anyone can produce malaria vaccine that at
this cost can be used easily, we will buy 100 million units
of whatever you produce.”>>You have explained
about something like two issues
of deep learning. That is there is no formal proof for
anything what is going on. And the second one you told that images then you will add noise. So, could you tell
that what is that means image if you
will add noise.>>So basically, deep learning
has been used for image understanding
where they were able to kind of say this is
a cat and this is a car, there’s a driveway and
various other things. And then they went and added
white noise to the image. And when you add white noise, it was very light intensity. When the human beings look at it just looks like
the same picture as before. When a computer looks
at the same data, it gets completely bonkers. It gets confused.>>It is something like Generative Adversarial
Network from the noise, it can generate
a complete EMS, is that.>>No. This is
very simple additive noise. I give you two images, one without noise
and one with noise. They’ll look the same
to a human being. But a machine trained on one kind of image
doesn’t know how to recognize this one,
that’s all. Yeah?>>Thank you so much
Professor [inaudible] for such a wonderful talk,
highly informative. I want to comment on
that talk that you paid two billion dollars in
1972 to buy a memory of 40.>>Two million not billion.>>Yeah two million dollars
you paid to get the memory of [inaudible]
for yourself in 1972 That is that basically shows the kind of growth
in electronics segment, microelectronics what has
happened at the memories of much higher size are available at much cheaper price nowadays. So, I needed your
comment here because whatever we are doing in say Artificial Intelligence,
or Deep Learning, or Big Data everything is ultimately related to analyzing
too huge amount of data, which all of us are
producing as humanity. But simultaneously
we know very well that since last decade or so, there have been conversations Moore’s Law is now giving up, we’ll not be able to scale down the technology
even further, we already, researchers are working on technology like 30 nanometer or seven nanometer. So, I just needed
your comment here owing to your huge amount
of experience. What do you feel on one hand
we are able to produce computer technologies
using which we can, we are developing
capabilities to analyze that kind of data, right? Using deep learning or
big data technologies. But simultaneously
the platform on which all these technologies have
to run those are giving up. So just needed your comment on your experience what
do you have to say?>>Yeah, I’ll be
happy to answer that.>>Thank you so much.>>First of all, so that those of you who
have not been following, computation as for
the Moore’s Law has been doubling
every 18 months. Therefore, you get a factor
of 100 every 10 years. In 30 years, we get a million. Memory, this hard disk memory technology
has been actually doubling every 12 months and bandwidth has been
doubling every nine months. Not recently in
the 80s and 90s and in the early 21 century and those are tapering
off right now. And what is happening
is we’ve all been wondering where
this exponential growth cannot go on forever. There comes a point where the physics gets in
the way as he was saying the feature sizes of circuits right now that are being made are
about 10 nanometers, and they may go
down a little bit then the current
lithographic techniques won’t work using
conventional light. But they’re looking at you know other lower wavelength
technologies. And they’re looking at
other kinds of things but none of them have the property
of mass production, which is one of
the reasons why we can get all these things cheap. But every time we thought
we were predicting this exponential growth
that will come to an end by 2020, it
doesn’t look like it. Maybe another 10
more years maybe 20 more years and
the exponential growth doesn’t have to come because of feature sizes, but parallelism. You can essentially
say things are so small I can make a bigger chip, one centimeters square will hold a thousand times more memory and therefore I’ll use parallelism rather than a single chip
with a little thing. And at some point
that will also give up looking at 3D
fabrication techniques. So there are lots of
things going on because that whole industry
has a huge momentum. There’s a company called
SEMATECH in Austin, Texas which was which is a co-operative research arm of all the semiconductor
manufacturers. They make a road-map of
saying if we’re going to get to here what must exist. Then they go and stimulate
manufacturers of equipment. Saying can you make this to
do at this nanometre level? And that has been going on so far now that I’m
looking at can you use ultraviolet or some other lower wavelength
fabrication techniques for the lithography that will actually get you the additional. It has to come to an end. Exponential growth just cannot
go on forever and there’s the famous story of the king and the minister saying he was so pleased to
say, what do you want? He said, give me one grain for the first piece of chessboard, then double it each one. Two raised to
the power of 64 was more rain than the whole kingdom
could hold, right? And so the king gave
up at some point. So just the the power of exponential growth is
just amazing to see. So I think right
now we’re okay for another 10 years and
something else will come along and we need to
fact all the technologies, but it doesn’t matter. Basically, all I’m saying
is even with what we have we will have a billion times more
computational power and memory and bandwidth and they all have to be balanced otherwise they
don’t work together. One of them becomes
the bottleneck. And so we’ll have to kind
of wait and see, basically.>>Sir, could you please comment on how AI will
help to create more jobs. Our basic understanding
is that AI is curtailing onto the jobs. Could you please
comment on that?>>Absolutely. Yeah, I’ve
been thinking about this. So my highest priority
for helping society, especially rural people
is to create jobs. And I came up with a solution 15 years ago that I wasn’t thinking about AI, but
technology broadly. So the idea is the following, if you take a rural head
or someone there, ideally they should
be able to stay at home in the village
and not move into the city and earn a living. At that time, I came up with some job but unfortunately it required an infrastructure that did not exist or
could not be created. I’ll give you an example.
I went to Mahindra, Alan Mahindra is the chairman. I told him, you’re making all these cars why
don’t you set up a driver training center in some villages and you’ll
sell cars and then they’ll learn how
to repair and drive. And the country needs
a lot of drivers and then the same thing
for setting up made so on. I said, you don’t have
to send people there all we need is facetime or video conferencing by
Skype or something where the lecturer stays in the city tells them
exactly what to do. And then they’re able to, so I wanted them to teach
a village leveled person. All kinds of jobs like
floriculture, horticulture. Various kinds of agricultural
things that only if you went to agricultural university
you get to learn. But I’m an illiterate villager, I’m never going to
get into university but I still need that knowledge. So if you could
train me and so we coined the term
Eighth Grade MBA. Eighth Grade Medical MD or
MBBS, Eighth Grade Lawyer. And so it turns out no matter
what profession you take, you can teach that
to fifth grade people if all that they’re
going to do is learn only that narrow thing
and execute it. So anyway. Fast
forward 15 years. Today at that time I needed
[inaudible] or [inaudible] or ITC to set up what I call the McDonald’s
like franchises saying, I’m going to you grow me the flowers and I’ll market it for you I’ll get you
the money and so on. This is how tobacco was grown by ITC you know and that’s where the childmen are and
all the cigarettes came from. And what they did
was proactively went and trained the farmer
on how to grow tobacco. You could do
the same thing but it requires somebody for whom that’s the major business model. And creating jobs
in villages was not a major business model for
most people but it was for me. And today I see an opportunity, which is because I can create
a storefront on Amazon, I can market
everything I make in my village to anybody in the world actually and
certainly anywhere. All that they have
to do is learn how. So the question is, I’m
an illiterate person, how am I going to create a storefront to sell my stuff
and how do I modify it? Again, I’m saying using
YouTube like video tutorials. I can train an
illiterate person to do simple modification
to his website on prices and
features or whatever, to be able to set
up the storefront. But still you need
lot more technology, you need to understand
various other things, and all of that can be taught. And also there will
emerge a set of consultants like you
who might actually say, “For five percent of
the profits you make, I will set up and
maintain your storefront.” That will work out perfectly
for both, it’s a win win. So I’m still working through
this because only Amazon and Flipkart are just coming online to serve rural communities,
they’re not yet there. They’re setting up some
storefront, but it will be there. It’s almost it’s a constant
six months of a year of that. And anybody that can make anything that is in
demand now can be sold. You’re no longer have to go
to the nearby town and go to your Mandi and set
up a little shop and hope somebody will buy it. You do it and you
take the order for the flowers and then
you cut the flowers, and then you ship it, and then the whole
system is set up because the same guy
that delivers stuff into the village will also take the stuff and then it gets into Amazon
distribution system. That whole logistics technology
that has evolved in the last five to
10 years is amazing. They started with
three-four day delivery then they went to guarantee
two day delivery and now they’re doing
one day delivery. Here if you’re in the city, if you order from Amazon in day, you get it within the next day. And now they’re talking
about four hour delivery. And even Walmart
is getting into it. So it’s a very interesting
ecosystem that is evolving that will wipe
out a whole bunch of small retail stores because
anything you want to buy you can get
much broader selection online. So if you have a sophisticated to be able
to go online and order, you’ll get better product
at a lower cost. And that’s going to happen and everybody will learn
that’s the beauty of it. I may be illiterate
but I’m not stupid. And all I need is an example somebody to say, “Do
this, touch this, touch that,” and I
can learn from there and I’ll be there.
It’s very interesting.>>I wanted to know what
your thoughts were on Singularity and what
its ramifications might be.>>Those of you
who have not read the book on Singularity
by Kurzweil. Ray Kurzweil predicted,
he’s one of the gurus of AI, predicted by 2045 computers will have more memory
than human brain and will be able to have more
computational power and will exceed the human capabilities and then they will
take over the world. That’s a big leap [inaudible]. And I don’t believe it. It will not happen.
All you ever do is study evolution for
the last billion years. We do evolve. There will be a next species
of human successors. But they’re not going to be
silicon or mechanical robots. Is more likely to be they’ll just look just like you and me. Like if you remember
10,000 years ago there were two species
that used to co-exist. Homo-sapiens and Neanderthals. And over a period they were competing for
the same resources, and one species got wiped
out because they did not have the reasoning capability
that Homo-sapiens had. And so supposing you and I
with my 1,000 intelligent assistants can do things
at superhuman rates faster, better, cheaper, 1,000 times faster than the rest
of you guys. Now, I can take over the world. This is the complaint
of Bill Gates. He says, “My God, I can’t live in a world where a small number of people
can take over the world.” So the singularity as predicted by Ray Kurzweil
may not happen, but there will come
a point where some of us will have
superhuman capabilities. Now, just because we have superhuman capabilities
doesn’t mean we’re going to go around
killing everybody else. You know we don’t go around
killing all these monkeys and chimpanzees and snakes and maybe we do kill snakes
if they’re threatening us. But in general species have learned to
co-exist on this planet. So if there is another species
next generation evolution that might not happen for thousands of years
or millions of years. And that’s one of
the predictions the discussion. Some people say
the next evolution will happen much faster. That could be
because it would be extra genetic evolution is not because of mutation in the body, it would be extra
genetic evolution that could make
the next species come faster. But my expectation is that something I’m
not scared about. Basically, I don’t expect
them to go around killing everyone just
because they maybe, there will always
be some malcontent. There will always be some people that will be bad actors
on the planet. But there will be
a lot more good actors who will suppress
them [inaudible]. If there are 100 people all of whom are
super human capability one or two of them don’t
want to do some damage, they’ll be wiped
out by the others because by pure science. So in that sense I’m
very hopeful about the future that we will be able to produce and
do things faster, better, cheaper, but it will be for the good
of all mankind.>>This is Barter
from ICAI Bangalore First of all thanks a lot for
a very thoughtful in talk. So my question is
about the state of AI research and activities
in this country. Increasingly when you go to AI and machine
learning conferences, traditionally you’d
have seen like you know European and US based companies are being represented there. And increasingly
Chinese companies are a significant presence there and their governments
are also putting in lots of efforts in there. So when I was in NIPS this recent NIPS there
was like single, zero Indian companies
that are getting represented there
while there are tons of companies from China. So in general, I guess people all over the world are
seeing AI as an opportunity. And we in this country are taking it as
business as usual. And the amount of I guess
like urgency that we should have shown we are
not showing it as much. I mean that’s a general feeling
here in a country of 1.2 billion people most of the AI research are probably
sitting in this room here. So, since you have
a different vantage point compared to most of us, my question is whether you share this concern and if
you do then like what we could do [inaudible] AI and
benefit from an Indian context?>>I do share the concern. Basically there’s
been a long time coming on how to
create more PhD’s, Dr. [inaudible] and
Dr. P.J talked about the issue of not producing enough PhD’s in
computer science and so on. And I also share
the concern that there’s not enough understanding that we have to sponsor
going to conferences. One of the first things I do at CMU is any student that’s
working on some area, I say, “Go to at least
one or two conferences, I will pay for it.” Because that’s where
they get energized. That’s where they get motivated. They listen to all the other
people doing the same kinds of things and they
learn from each other. And that’s one of
the first things. And China has been very
good in doing that. They spend a huge amount of money sending people
abroad and many of this, if you go to any US
university there are a lot of Chinese PhD’s
right now studying. And conversely they also invite people from all over
the world that are doing interesting work to come
and talk to larger groups. To give you an idea
I was in China four times last year,
four times. I was in India only twice. And I’m an exception for India, but in general and but that’s the difference
between I wouldn’t call it dictatorship but it’s kind
of the economy are led by a group of I wouldn’t call them intellectuals or intelligentsia, a group of. Leaders that have
been self-selected, not self-selected,
basically it’s a subgroup. If you’re a member of the party, then you have an election but it is not
the entire population. We have a democracy and I think when all things
are said and done, I would prefer this solution
than that solution. That means we may not get there faster and we will get there. I remember not having
any other cars that we see on the road today
until more recently. Sooner or later, everywhere, it will breakup and that’s exactly what happened
to television. They said, “Nobody needs colored televisions,
nobody needs television.” Mr. Nehru and all the Prime
Ministers of 50’s and 60’s, and I used to watch them, and say and somebody wanted
to make microwave oven. They said, “Nobody
needs microwave oven.” So, everything was
a problem and slowly, we are kind of getting there. So, we’ll be 20 years behind and I don’t think that should be a concern. I am concerned. But what I’m saying
is basically, I can make my prediction. In another 30 to 50 years, India will be
the second largest economy after China, ahead of USA. Right now, US is $16 trillion, China has $8 trillion,
we are three. And Mr. Modi said
at Davos yesterday, a day before, he expects to
be at $5 trillion by 2025. And I think that will happen. It’s another seven more years and you can grow by 50 percent. And the question
is when will he get to $20 trillion or $25 trillion? When will the country get
there? We will get there. This of course,
whether it’s 20 or 50. And at that point, we will be
the second largest economy. Then we can afford
all these things. This is the problem.
We still have poor boy mentality and that causes many of
the decisions that people make. But I think, overall, I am very proud of what this country has accomplished
I know and so on. We could have done
better but that’s okay. You know that is the price
you pay for democracy.>>One more question. But before I return the mic, just to follow up
on that question. I think if I may had
a question of my own. So, the other part of
that question was that, you know especially you said
about the number of ML PHDs being in this room representing all of
the country almost, any comment from
you on what can be done from education,
higher education research. I mean, the problems
where those things.>>Dr. Pidyan and I have
talked about it many times. And the single word is
grow your own, right? If we want, if somebody in
the country were to make a decision that we want to
be one of the leaders of AI, then it’s easy to do basically their NF up us better understand what
the basic technologies of AI. I was very pleased to see
Dr. Jawhar up here and from here before I even knew
he was working on it. He has made significant advances in deep learning from here. And so, I’m sure there
are people in IIT Bombay and Chennai and everywhere
that are doing great stuff. And together, if the way China mobilizes
people is they say, “We are going to have
a million people trained.” And an order goes to all the governors and all that
kind of thing and they are supposed to pick up the best people and then
send them to the training. This is the way they
do Olympic training. If you want to be
the best swimmers, best various kinds
of Olympic Games, they have a special program. They literally go to the whole country and
pick up the people and then put them into an intensive training
program just for that. So, if I had a million people, that’s not a very many people
out of a billion, 1.3 billion population,
and we can train them. And then, all that will
happen is they will be trained in AI but
with nothing to do. And that’s where you now need to go to
the next step namely, once you train them, exactly what is it that you
want them to do? There, I will give them my list saying let’s not do what China is doing
or what others are doing. Let’s take our national problems and see how AI can be used
to solve these problems. But given the data,
the question is I need to create an environment
where every person that has a smartphone and
every person has opted in to provide me with data
one percent of the time, anonymized appropriately so that I’m not violating
anybody’s privacy. But I know what’s going
on in the entire country. Once I have that data, I need those people for different task groups
to be analyzing that data continuously and they will make
a lot of progress. It’s purely a question
of time and money. And somebody up top
deciding that has be done, tomorrow we’ll have Mr. Ranjit
Shahani with the Secretary of Defense and ask him
that question and say, “When will you mobilize
the country for AI?”>>Yes, so I think we’re running out of time
but we’ll probably take one more question here.>>Thank you sir for
your illuminating talk. I’m Abilop from IT Corp.
You mentioned in your talk about Guardian Angel
technologies, which will act, actually monitor the environment
for threats and will forewarn the user
about imminent threats. Now, I have a few questions. Do you envision these
Guardian Angels as a form of Cloud services which will somehow continuously
take inputs from sensors, and cue some kind of
prediction framework to generate predictions of say, a threat or bad situations, and forewarn the users? And furthermore, how do
you code these systems? What kind of algorithms
do you all want us to put into these
behind the systems? And more importantly,
how do we test them? Because if we are actually proposing such a system and we are claiming
that this will work, then we need to have a solid performance statistics
to back up our claim, which will actually
convince people to actually even start implementing
them on a wide scale. So, how do we go about this?>>I’ll give that to
you as an assignment. It was part of my talk on architecture and how
to design it and so on. But you already envision
some of the things. It has to be Cloud-based. It is always on
and always working. And that one of
the first things we had to do was to create
an app and put it in the Cloud and then have it continuously monitor
one source of data. And therefore, we
are envisioning literally hundreds of
Guardian Angels for each person. If you read the book by Minsky called
the Society of Minds, in the old days we used to have this debate about whether there’s one grand
intelligence that are doing everything
or whether they are highly specialized
intelligences which do one thing, but they
talk to each other. So, each knows what
others are going on. And so it’s that Society
of Minds idea here. If you have all these
intelligent agents and you built one purely for being paying bills or one purely for monitoring what’s going on and earthquakes
around the world, and then if it’s
going to affect you, it needs to figure out who I am, where I am at
that point in time. Most of the time it’s
not going to affect me so it will be quiescent. But occasionally, if
there’s some serious threat, then it warn me, right? And so the issue of how to design one of those things,
we have done it. We know how to do it, right? But how to propagate it to every person requires
that infrastructure, that means I need everybody
to have a smart phone. And I need to create a marketing
ecosystem where somebody makes an earthquake detector
Guardian Angel app, and which will self install
itself into the thing, and will restart itself when the power goes
off or something else. It’s what we call
self-healing systems. The whole class of system
where no matter what happens, you turn it off,
it’ll turn back on. And so, the whole issue of
how you design such systems, self-healing system is not part of AI but
other computer science. This is especially important
for stock markets and so on. You can’t say, “Sorry, I’m dead and we can’t
do any trading.” Because they need
quadro duplicated systems, so that no matter what, you can actually do it. So the architecture, if
you just think through it, is doable, but there are
so many aspects to it. Always learning,
what does that mean? Where I’m collecting the data, where do I’m
collecting the data? Not only from me but everybody else that is worried
about earthquakes. There’s some earthquake
somewhere else, how did we detect
what happened there? What are the solutions
those people used? Can I use them here
when it happen? So, all of that requires significant external
entrepreneurial thing, just like happened in
the app space for smartphones. There are a million
apps that you can download. Many
of them are free. Some you have to
pay, and people pay because it’s very
useful. Thank you.

Leave a Response

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

1 thought on “Keynote Talk: AI and Technology in Service of Society”

  • Prof. Raj Reddy is not only a legendary computer scientist but also a great visionary for the betterment of the country. Ideas from people like Raj Reddy should be implemented in collaboration with the government and the industry. My heartiest regards to Prof. Raj Reddy and his thoughts for the country. Salute to you Sir!