[SQUEAKING] [RUSTLING] [CLICKING] GARY GENSLER: On
Monday, we talked about financial
technology as a whole, and what we're going to do
in these next four classes is really talk about
three major technologies. The broad subject of artificial
intelligence, machine learning, and deep learning,
we'll talk about today, and next Monday, we'll move on
and talk about the marketing channels, what I'll broadly
call OpenAPI and some of those conversational agents
and the relationship that is happening there. And then next Wednesday's
class, I might be off. It's the following
Monday's class, we'll talk about blockchain
technology and cryptocurrency. So we're taking three
big slices of technology, and then we're going
to go into the sectors. And so what I thought of about
in structuring this class, as we've already discussed,
is I take a broad view of what the subject of fintech
is, that it's technologies of our time that are potentially
materially affecting finance. So in any decade,
any five year, it might be different technologies. It's technologies of a certain
time that are materially affecting finance,
and it's the broad-- any competitor can use fintech. Now, I recognize
that a lot of people use the subject or fintech,
and they narrow it down, and they just use those
terms for the disruptors, the startups. And there were
really good questions of why I think of
it more broadly. I just think that the incumbents
are very much involved in this fintech
wave, and to ignore that would be at your peril if
you're starting a new business, if you're an entrepreneur. And of course, if you're Jamie
Dimon running JPMorgan Chase, or you're running a big
bank in China or in Europe, you would be at your peril
if you didn't think about it. And big tech, of course,
has found their way in here. The Alibaba example's in China,
but all around the globe, whether it's Brazil,
whether it's India, whether it's-- your big tech
has found their way in here. And so I've also decided
to sort of structure it around these three big
thematic technologies, and then we'll start to dive
into the sectors themselves and take a look at
four or five sectors before we call it a day at
the end of the half semester. So I'm about to pull up
some slides, but, Romain, are there any broad questions? ROMAIN: Nothing so far, Gary. GARY GENSLER: OK, thank you. So just give me a
second to make sure I pull up the right
set of slides. I think that-- all right,
so two readings today. I don't know if you were
able to take a look at them. I hope you had the time
to take a look at them. And these were
really just a chance to sort of grab hold of a broad
discussion of what's going on. Again, I went to the
Financial Stability Board. It's a little dated
because it's 2017, but I thought the
executive summaries and these various
sections were helpful. And then a shorter medium
post really on six examples, and today, we're
going to talk about what is artificial intelligence,
what is machine learning and deep learning, and then what
are the eight or 10 major areas in finance that it's being used. And then next
Monday, we're going to talk about a lot
of the challenges and go a little deeper
with regard to this. Now, I said that I
had study questions. We're going to see if
we can get this to work, and I'm going to ask for
volunteers to speak up. If not, Romain might
just unmute everybody, and then we'll see who I can
cold call on or something. But who would like
to take a crack? And remember, we're all
pass emergency and no credit emergency, so this
is just about trying to get the conversation going. But who would want to
answer the question, what is artificial intelligence,
machine learning, deep learning? And you don't need to
be a computer scientist, but these terms are
really important if you're going to be an
entrepreneur and do a startup, or if you're going to be in
a big incumbent or big tech company, just to have
this conceptual framework and understanding of artificial
intelligence, machine learning, and deep learning. So I'm going to wait for
Romain to either find somebody who's raised their
blue hand, or, Romain, you get to help me cold
call if you wish. But hopefully, somebody
wants to just dive into this. ROMAIN: I'm still waiting
let's see who will be-- ah, we have our first
volunteer of the day. Thank you, Michael. GARY GENSLER: Michael. STUDENT: Sorry, I
forgot to unmute. So my understanding-- artificial
intelligence just is more of an over encompassing
term, being-- just computer is kind of
mimicking human behavior and thought, so that's more-- GARY GENSLER: Let's pause there. And a very good
answer, so computers mimicking human behavior. And, Michael, just a sense-- do you have a sense of how--
when did this come about? Was this in the last
five years, or was it a longer time ago that
somebody came up with this term "artificial intelligence?" STUDENT: More like the
early to mid 1900s. It's been a while. GARY GENSLER: So
it's been a while. It's actually way
back in the 1950s, "artificial
intelligence--" the concept of a computer mimicking humans. In fact, at MIT, we have the
Computer Science and Artificial Intelligence Lab. It's not a creature
that was just invented in the 20-teens or 20-naughts. It goes back decades. It was a merger of
two earlier labs, but we've had an artificial
intelligence lab at MIT for multiple decades. So who wants to say what
machine learning is? Romain? I hand it off you
to find another-- ROMAIN: Who will be the
next volunteer for today? STUDENT: I think the
machine [INAUDIBLE],, there's very limited or even
no intervention with human, and learning are you
solve problems step by step in a sequential manner. So that is like
the machine solve problems step by step without
much intervention by humans. GARY GENSLER: So I
think you're raising two points, is the
machine solving a problem without
intervention, which is good. And then you said
briefly, learning. Do you want to say
something more, or anybody want to
say, what does it mean that the machine is learning? ROMAIN: We have Rakan
who raised his hand. STUDENT: Yes, well,
essentially, it means that you're
feeding the machine data, and as you feed
the machine data, the machine learns not
to do specific tasks. And you get better
results as you feed it more data and more and more. GARY GENSLER: All right, so the
concept of machine learning, again, is not that new. It was first written about
in the 1980s and 1990s. It had a little bit of a wave-- you might call it a boomlet,
a little bit of a hype cycle-- and then it sort of tamped down. But the conceptual framework
is that the computer, machines, with data, are actually
learning, that they actually-- whatever algorithms or decision
making or pattern recognition that they have gets better. So machine learning is a subset
of artificial intelligence. Artificial intelligence
is this concept of computers, a
form of machines, mimicking human intelligence,
but it could mimic it without learning. It could just replicate or
automate what we are doing. Machine learning
was a new concept-- it didn't take off. It didn't dramatically
change our lives at first-- where, actually, the
computers could adapt and change based upon
analysis of the data, and that their algorithms
and their pattern recognition could shift. Does anybody want to take
a crack at deep learning? ROMAIN: Anyone? Albert raised his hand. STUDENT: So deep
learning involves using large neural networks
to perform machine learning, so it's sort of a subset
of machine learning. But it can be very powerful,
and most of the time, it just gets better and better
as you feed it more data. Other machine
learning algorithms tend to have sort of a plateau
and can't improve no matter how much data you feed them. GARY GENSLER: So important
concept there in a phrase was "neural networks." Think of our brain's-- neural, neurology. It's about our brains. Early computer
scientists started to think can we learn something
from how the human brain works, which is in essence, a
biological computer that takes electrical pulses and
stores data, analyzes data, recognizes patterns. Even all of us right
now are recognizing voice and visual patterns. That's in our human brain. So when looking at the brain,
the conceptual framework is could we build a network
similar to the brain, and thus, using these
words, neural networks. In deep learning and
machine learning, there's pattern recognition,
but deep learning is a subset of machine
learning that has multiple layers of connections. And in machine learning, you
can take a base layer of data and try to find patterns,
but deep learning finds patterns in the patterns. And you can think of it as
putting it through layers. Now, if you're in Sloan,
and you're deeply involved in computer science, and
you also enjoy the topic, you can go much further. But in this class, we're
not trying to go there. The importance of deep learning
is that it can find and extract patterns even better
than machine learning, but it takes more computational
power and often more data. Deep learning is more an
innovation of the 20-teens, and by 2011 and 2012, it really
started to change things. And in the last
eight years, we've seen dramatic advancements
even in deep learning. It's a conceptual framework
of taking a pool of data, looking at its patterns,
and going a little higher. I'm going to talk about
an example a little bit later in this
discussion, but please bring me back to talk about deep
learning in facial recognition, deep learning in
autonomous vehicles and just thinking
about it there, and then we'll pull
it back to finance. ROMAIN: Gary? GARY GENSLER: Yes, questions? ROMAIN: We have two questions--
one from Rosana, who's asking, what is the difference
between representation learning and deep learning? And then we'll give
the floor to Pablo. GARY GENSLER: So
very good question. Representation learning,
you can think almost as in between machine
learning and deep learning. If we said, this is almost
like those Russian dolls. The AI is the big vessel. Machine learning is a
subset, and deep learning was a subset of
machine learning. Representational learning is
a subset, in this context, of machine learning, and it's
basically extracting features. So think of machine learning
that's looking for patterns. It's extracting
features out of data. So photo recognition, our
standard Facebook photo recognition that might
recognize Kelly versus Romain is extracting certain
features of Kelly's face. I'm sorry if I'm
picking on you Kelly. You just happen to
be on my screen. But it's extracting
features, so some people call it feature learning or
representational learning, but extracting features as
opposed to specific data. And then deep
learning is generally thought of as a subset
even though there is some debate as to how
you would categorize these. And you said there
was another question? STUDENT: Yeah, hi, Gary. This is Pablo. So I just wanted to verify,
because my understanding is that, also, one of the big
differences between machine learning and deep learning,
or artificial intelligence in general besides
deep learning, is that deep learning
used unstructured data, whereas typically, for
machine learning models, you need basically tables of
data that you have organized and labeled and everything. And just the check
whether that's the difference besides
the additional complexity and multiple layers. GARY GENSLER: So let
me step back and share the important vocabulary,
and again this is important vocabulary
well beyond being a computer scientist. It's important vocabulary if
you're running a business, and you're trying to get the
most out of data and your data analytics team. So I'm going to assume we're
talking as if you are now going to be in a C-suite, and you
want to get the most out of your data analytics
teams and so forth-- is this concept of structured
data versus unstructured data, and then I'll go to
that specific question. Data that we see all the time
with our eyes, that we read, that comes into
us, you can think of as unstructured
sometimes because it doesn't have a label on it. But if it has a label
on it, all of a sudden, people call it structured data. So machine learning,
conceptually, is that the machines are getting
better at recognizing patterns. It's primarily about
pattern recognition, that the machines are getting
better recognizing patterns off of the data. And the question here
is, is machine learning always structured learning? And structured learning
means that the data is labeled, that you
have a whole data set, and it's labeled. An example of labeling
I will give you, that we all live with in our
daily lives-- how many people-- just use the blue
hands if you wish. How many people have ever
been asked in a computer, will you please
look at this picture and tell us whether
there are any traffic lights in the picture? We want to make sure
you're not a robot. All right, so here,
I'm just going to-- Romain, just cold call somebody
that put their hand up. I want to ask them a question. ROMAIN: Sophia, please? GARY GENSLER: All right,
Sophia, are you are unmuted? STUDENT: Yes I'm on. GARY GENSLER: All right,
so, Sophia, when you go in, and you tell the
computer that you're not a machine, why do you think
they ask you that question? STUDENT: To make sure that
there aren't any bots who are trying to take
advantage of any service that the machine
is trying to offer. GARY GENSLER:
Really good answer. You're correct, but you're
not completely correct. What's the other
reason that they're asking you whether that's
a traffic light or not? STUDENT: And also
to collect data so that they can use
that data for labeling for future purposes as well. GARY GENSLER: So they
are labeling data. They're using Sophia, if I
might say, as free labor. Sophia is training the data
that Google or whomever is putting together there. And thank you, Sofia. You're labeling that data
so that autonomous vehicles will work better in the future. You're also frankly
labeling data to put millions of people
out of jobs as truckers, but I don't want
you to get sort of wrapped up into those sort
of social and public policy debates, but that's
what's happening. So back to the earlier question. Data can be labeled by us. An earlier form of
labeling was labeling what is an A, what is
a B, what is a C, what are all the letters
of the alphabet so that our postal
services now can use a form of machine learning
to read all of our written scratch on envelopes. If we address an
envelope, it can all be read by computers
rather than humans. To your earlier question-- I went a long way around
this, but machine learning can be both unstructured
and structured. The question was, is machine
learning always labeled? Is machine learning
always structured? And the answer's no. Machine learning can also be
unstructured and unlabeled. Deep learning can be
both labeled, which is structured, or unlabeled. Some of the economics and
some of the computer science are worthwhile to understand. Labeled data can
be trained faster. Label data can, in many regards,
lower your error rates faster and have a certain-- extract correlations better,
but it comes with a cost. You need to label the
data, and so there's some tradeoff of getting Sophia
to label data or other humans to label data versus
unlabeled data. Think about radiology in
the practice of medicine, and looking at body
scans or mammograms or any form of
radiology to identify whether there's an anomaly. There's something that needs
to have further investigation to see whether it's
a tumor or not. Radiology is dramatically
changing in the last three or five years based upon machine
learning and deep learning. Remember, deep
learning just means there's multiple layers
of pattern recognition in these neural networks. Labeled radiology, labeled
mammograms, or labeled MRIs will train the
machines faster, but unstructured, unlabeled
data can also be used. You need bigger data sets. So I went I went
off a little bit, but I hope that that's helpful. Other questions, Romain? ROMAIN: Yes, we have
one from Victor. GARY GENSLER: Please. STUDENT: Hi, professor. I just wanted to
double click on the-- take a step back in the initial
definitions between machine learning and deep learning,
because we were discussing that deep learning had
the feature that it keeps learning despite the
data growing exponentially. It doesn't plateau. But I don't fully understood
the differentiation between two concepts beyond that. GARY GENSLER: So I
didn't disagree or agree with that comment,
and I apologize, I can't remember who said
that deep learning keeps to grow exponentially. I think both machine
learning and deep learning-- both machine learning and
deep learning learn from data, and this word "learn" should
be explored a little bit more. What machine learning
and deep learning can do is extract correlations. I hope nearly
everybody in this class has taken some
form of statistics at some point in time. You might have hated
statistics, but we all took it, and some of us took
more advanced statistics where you use linear
algebra and the like. But just thinking about a
standard regression analysis-- a standard regression
analysis finds a pattern, generally a linear pattern,
or a quadratic pattern if you move on. Machine learning and deep
learning find pattern, and they're really remarkable
tools to extract correlations. And one of the features
of both machine learning and deep learning is
they look at error rates, particularly versus data
sets that have been labeled. And traditionally, what you do
is you have a big data set-- maybe it's millions of pieces
of data that you're training on, and you take a
random sample of it and put it to the side, a
random sample on the side that you label. And then you compare
what comes out of the machine learning with
the test data on the side and see what's the error rate. And this labeled set on this
side might say, these are men, these are women,
this is a stoplight, this is a traffic light,
whatever the labeled data is, and you see the predictive
model, what's the error rate. Both machine learning
and deep learning continue to do quite well. And I apologize I cannot
remember who said it earlier, which student said, deep
learning continues to grow further than machine learning. It can, but I wouldn't accept
that machine learning can't get better and lower error rates. Now, once you get down
to very low error rates, that's another
circumstance altogether. The difference between deep
learning and machine learning is that-- I'm going to use photo
recognition software. If you put a photograph into a
computer, where does it start? Does anybody-- what does it
see at the very beginning, it's base level of data? STUDENT: The pixel? GARY GENSLER: What did I hear? STUDENT: I said, the pixels. GARY GENSLER: Pixels. So the only thing a
computer can read is pixels. It has to start with
the pixels and build up, and the next layer at most-- and again, I once
pretended to know something about computer science. But I programmed in
Fortran and APL years ago, and that was before
many of you were born. But I guess I used to know
how to program something. But if you read the
pixels, the next layer up to find a pattern
in the pixels is just small changes
of shade, and then you can think of the next
layer up from those little-- you can see edges. So the computer has to
sort of go through layers from the pixels up to this is
a traffic light versus a stop sign, and so deep learning
interposes multiple layers of pattern recognition. Some would say that
you need many layers, and then other research
shows that, no, once you get to
about three layers, there's less and
less return on this. So let me sort of-- unless, Romain, is there
other questions, or can I-- ROMAIN: No, we're all good. GARY GENSLER: All
right, so again, this is just a broad
thing, and we're not going to spend as much time. But what's natural
language processing? Does anybody want to-- just what are these
words broadly mean? ROMAIN: Any volunteers? GARY GENSLER: I'll
volunteer, then. So natural language
processing is just simply taking human
language, natural language, and processing it down
to computer language, all the way down to
machine readable code, or going the other direction. So you can almost think
of it as input and output to the computer, or we
have many, many languages represented on this call right
here with 99 participants. You can think about
it as translating French to German and
German to French, but instead it's
natural language, what we do, down to the computer. And this is really important
in terms of user interface and user experiences, and we'll
get to that a little bit more. And then we're going to talk
a lot about which sectors in financial services
are being most affected at this point in time. So I won't call on the class
right now because I want to keep moving to go forward. So we're in to talk about the
financial world and fintech, and then-- oops, I
didn't change this. This slide will-- I'll
shift, because that's from the other day. So we looked at this
slide the other day, and it just helps us-- what is AI machine learning? Extracting useful
patterns from the data, using neural networks
that we talked about, optimizing to lower error rates. Optimizing so that you actually
say with 99% or 99 1/2% of the time, this is a traffic
light, this is a stop sign. Actually optimizing. There's lots of programs
that you can use. Google has TensorFlow, and-- I don't know. Have many of you
ever used TensorFlow? I don't know all of
your backgrounds, but any show blue hands? I'm not going to call
on you to describe it. I'm just kind of curious
if there's many people that have used TensorFlow or not. ROMAIN: How about
we go with Devin? GARY GENSLER: So Devin's
actually used it. I wasn't going to
pick on them up, but if you wanted to say
anything about it, Devin. STUDENT: Yeah, I
can very briefly. So it essentially gives
like a plug-and-play method to do machine learning and
build neural nets in Python. You don't necessarily have
to have a full understanding of how under the hood works. You can just add bits as
you want, take bits away as you want, and it speeds
up the whole process. GARY GENSLER: And so what's
important about that is, just as somebody that
came of age in the 1970s and '80s didn't have to learn
how to computer code all the way down to
machine readable code, they could learn how to use-- by the 1990s, C++, or C
And C++, and later, Python. And many people in this
class know how to use Python. In the machine
learning area, there's been plug and play
programs like TensorFlow, where you don't need
to actually know how to build the data
and things like that. Most importantly,
and this is if you're thinking about
being a data analyst or actually building
a business around it, it's the data and the questions
you train on the data. And most studies have
shown, as of 2019, that 90%, 95% of the
cost of data analytics is what some people might
call cleaning up the data, making sure the
data's well labeled. We talked about structured
versus unstructured data earlier. Really important about
that labeling, the cost. If you have 1,000 people working
in a machine learning shop, or 500 people or five,
it is quite likely that a big bulk of their time
is standardizing the data, so to speak cleaning up the data,
making sure that the fields are filled, and ensuring that you
can then train on this data, meaning it's labeled, or
enough of it's labeled. And then thinking about what
the questions you're really trying to achieve, what
you're trying to extract. Why is it happening now? This is off of Lex Freeman's
slide again, but why? Because the hardware, the
tools, the analytics-- a lot has shifted in the last
five or eight years. To give you a sense of what it's
being used for, all of these, we know. We know already. It's dramatically
changing our lives. When we're sitting at
home, sheltering at a home, and you're thinking
about the next movie, and you're on Netflix,
Netflix is telling us what they think the next
thing we should watch. That's training off of
not just the knowledge of what each of us
has been watching, but it's about what
others are watching. It's the Postal
Service that no longer has to have a human
reading the text, our scrawl on the envelope. It's Facebook with the facial
recognition programs and so forth, and autonomous
vehicles that are now being
tested on the roads but are very likely
part of our future. Now, will they be rolled out in
a dramatic way in five years, or will it be 15 years? But I would feel
comfortable that we will have autonomous
vehicles on the road sometime at least by the
2030, but maybe others would be more optimistic. So it's changing a lot
in many, many fields. The question is now, how is it
shifting this field of finance? Why do I put it at the center
of what we're doing here? So we talked about this. This was just my little
attempt, and so now you have a slide that does what
we chatted about before. One important thing also is
happening is, in finance-- [CLEARING HIS THROAT]
excuse me-- people are grabbing alternative
data, using alternative data. I said earlier that the
most important questions are what's the good data? So then we think about, in
finance, what type of data do we want to grab? Data analytics in finance
goes back centuries. That is not new. The Medicis, when
they had to figure out to whom to land
in Renaissance era had to figure out who
is a good credit or not. And two data scientists from
Stanford started a company called Fair Isaac-- those were their last
names, Fair and Isaac-- and that became the FICO
company, the Fair Isaac Company. So the data analytics in
finance and the consumer side has certainly been around
since the 1950s and 1960s, but where we are
now is to say what is the additional types of
data that we might take, and not just banking and
checking and so forth? But Alibaba can look at a
company, look at a company very closely, and do a full
cash flow underwriting. Alibaba, because
they have AliPay, can see what that small
business is spending and what that small
business is receiving. Amazon Prime can't
quite do it as much, but they can do a
bit of it as well. And even Toast, which
is a fintech company in the restaurant business
until this Corona shut down, Toast could see a
lot about what's happening restaurant by
restaurant in their cash flows. They had the revenue side more
than the expenditure side. Alibaba, much better
data sets than Toast, but I wouldn't put at rest
a Toast started earlier in the last year to do credit
extension to the restaurant business. Now, do you need deep
learning and machine learning to do that? Not necessarily. You can still use
plain old regression and linear statistics,
linear regression analysis, but machine learning and deep
learning help you go further. And then there's, of
course, everything about our usage, our browser
history, our email receipts, and so forth. If we look at China,
they've stood up a broader social credit
system, and in that system, they can tap into
data about users in many different platforms. Romain, do I see--
are you waving at me? Is there a question? ROMAIN: No, I'm not. Sorry for that. GARY GENSLER: That's all right. So natural language
processing, I mentioned. I just want to say a
few more words about it. Think of it as computers input
and output interpretation. This sort of going
from German to French, or going from computer language
to human language and back again. That's this important
back and forth, and so it's natural
language understanding, meaning a computer understands
something, and also natural language generation. So it can be audio,
image, video, any form of communication--
even a gesture. This hand wave can
be interpreted-- if not now in 2020, within a
few years will be interpreted. A movement of your face
will be interpreted as well. And so how it's being used
is really quite interesting, but we all know about chat bots
and voice assistance already. That's shifting our worlds. Yeah? ROMAIN: We have a
question from Nadia. GARY GENSLER: Nadia, please. STUDENT: I have one question
related to the chat bots. What kind of
factors do you think will encourage people
to use chat bots? Because now, I do
think people prefer to talk to a person
rather than chat bots. GARY GENSLER: Well,
I think, Nadia, we might still prefer
to talk to a person, but there's a certain
efficiency in-- and I'm just going to stay
in finance for a minute. But there's a certain efficiency
that financial service firms find that they can use chat
bots instead of putting a human on the phone. So even when you and I call
up to a Bank of America, and we want to check in on
something on our credit cards, we're put through a
various series of push one if you want this, push
two if you want that, and we're pushing buttons. That's not high
technology, by the way, but that's an efficiency
that Bank of America has interposed into the
system instead of having a call center of humans. And so if they can move
from a cost center of humans to an automated call
center of chat bots, they can provide services at a
lower cost and to more people. Now, you and I might still
want a human on the other side, but business is
interposing an automation, and that automation means,
often, quicker response time. So many of us, you go
into a website today, and there's a little bot
window that comes up. And the first thing that
comes up on so many websites-- and this is true if
it's a financial site. It's true if it's a
commercial website where you're buying something online. It's probably true
of dating websites, that there's some little bot
window that comes up and says, can we help you. That's not a human, but it
does give us greater service. It gives us an immediate
recognition somebody's answering a question. Nadia, do I sense you'd
prefer not to have the chat bots interposed? Is Nadia still there? STUDENT: Oh, yeah. Yeah, because I do
think sometimes, chat bots, we ask a
question, but their answer is not really related
to our questions. GARY GENSLER: So you
would prefer a human because you think the human
will interpret your question and be able to answer it better? STUDENT: Yes. GARY GENSLER: So if the chat
bot could answer as well as Kelly could answer-- again, I'm sorry, Kelly. You're on my screen. But if the chat bot could answer
as well as Kelly or Camillo or others in this
class could answer, you'd be all right with that? STUDENT: Yeah, it's faster. GARY GENSLER: See,
if it's faster, and it can answer as well. So those are some of the
commercial challenges. ROMAIN: I think Ivy would
like to contribute as well. GARY GENSLER: Sure. STUDENT: Yeah, I just wanted
to offer a little bit of some of the consumer
studies that we did when I was working
for a startup where we were building chat bot. And interesting enough, I
think most people are actually pretty-- and we did a pretty
large survey, and most people were
pretty open to the idea of working with a chat
bot because I think that's become so pervasive. But then there's
this idea, they want to know that it is a chat
bot, that the company is very transparent about that,
because people change their behavior when they
are speaking to the chat bot or some kind of
virtual assistant, as long as they know,
just to build that trust. And also, I guess
we try to be more explicit in our
wording, both verbally as well as written texts. And then secondly,
I think because-- I guess I wanted to pose
this as a question, too. When I think about chat
bots and things like that, the technology is not
necessarily there, or it takes a long time. And so I'm just curious what
your thoughts are in terms of-- for me, I see it as like AI
gets us 80% of the way there, but we need the human
touch 20% of the way there. And so I actually see a lot
of companies either having that human at the end of the-- you ultimately still need a
human at the end of the day. So I mean, I just wanted
to explore that a little. GARY GENSLER: I think what
Nadia and Ivy are raising, and I'm sure we're all
grappling with this. We are living in a
very exciting time, and I'm not talking
about this corona crisis. That's a different
type of challenge. But we're living
in an exciting time where we can automate a lot of
things that humans have done. We've automated so
many things that humans have done for
centuries, but we're now automating this
interface, through chat bots and conversational
interfaces, voice assistance. I would dare say that, of the
90-plus people in this class now, that most of us, if not all
of us, at some point in time, have used Siri. I mean, if we're
driving along a highway, and we're supposed
to be hands free, we might talk and start up
an app or something legally, legally. And there's a lot of
automation that's going on. How many of us have called
to arrange a reservation at a restaurant, and
we're not quite sure if we're talking to a human
or a conversational agent? But I think that
what Ivy's saying is that there might always need
to be a human somewhere there. I don't know, Ivy,
if that's correct. That's where we are in 2020. Let's think about
autonomous vehicles. Right now, we're not
comfortable enough. The manufacturers aren't
comfortable enough. The computer scientist
aren't comfortable enough. The regulators aren't
comfortable enough. The public's not
comfortable enough to have autonomous vehicles
on the road with no humans whatsoever, but that's
not really necessarily where we'll be in 2030. Or take radiology. Right now, at least
in advanced economies like in Europe and
the US and elsewhere, in advanced economies,
we say we still want a doctor's eyes on
a radiologist report. So the mammogram might be read
by some artificial intelligence machine learning trained data,
but we still have a human. But is that really the tradeoff
we'll make in a few years? And is it the right
tradeoff to be made in less developed
countries, where they don't have the resources
to have the doctors? And now, we even look in
the middle of this crisis, the corona crisis, if-- this is sort of God willing. If the Baidus of China
and the Googles of the US and others sharp
analytic AI shops come up with a way to
extract patterns and develop some recognition as to
who's most vulnerable, are we going to rely
on that, or are we going to say a human has to also
interpret it and be involved in it? And I don't know. So I think we're
at an exciting time where we're automating
more and more. I do agree with you,
Ivy, there's always going to be a role for humans. I'm not terribly worried that
we'll all be put out of a job. 200 years ago, our ancestors,
all of our ancestors, were, by and large,
working on farms. That's the economies. And we have found other
things to fill those roles and those needs. I think we'll still have the
humans, but not in every task. So let me go through
a little bit-- so the Financial
Stability Board, this is their definitions
I'm going to pass on, but this was in that
paper that you all read about what big data
and machine learning was. When I show this page
to computer scientists, when I show it to
colleagues of mine at MIT from the College of
Computing, they look at it, and they say, jeez,
that's funny that a bunch of financial
treasury secretaries and central bankers
and their staffs define big data machine
learning this way. So I partly put
it up because this is kind of what the regulators
define as to what it is. Machine learning may be defined
as a method of designing sequence of actions to solve a
problem, known as algorithms, and so forth. Computer scientists would name
it a little bit differently. So I said to you
the other day that I think of financial technology
as history is building on these things, but machine
learning and deep learning is at this top level. In the customer interface,
it's the chat box we were just talking about, and
on the risk management side, it's extracting patterns to
make better risk decisions. So it's in these
two broad fields, I sort of think of it is
the customer interface and then lowering risk
and extracting patterns. And sometimes, it's
not just lowering risk. It's enhancing returns. And so I was going to
go through and chat about each of these
eight areas, and not all of the areas we're going
to talk about are as robust. Asset management, right now-- asset management from
hedge funds all the way to the BlackRock and Fidelity
are exploring the use of machine learning and AI. By and large, most high
frequency trading shops, most hedge funds today,
most asset managers today, are not using much
machine learning and AI. I view that as an
opportunity in the 2020s. I view that as a real
possibility of a shift, a very significant shift. But where is it
being used so far? So BlackRock has been
announcing and saying that they're already
using BlackRock as one of the world's, if
not the world's, largest asset manager. Before this corona crisis,
probably $6 or $7 trillion of assets. It's, of course, a
lower number now. And BlackRock and
others have been saying we're using machine
learning to actually listen to all of the audio files,
all of the audio files of the major companies
when they announce their quarterly earnings. And they're also putting in
news articles, digital news or articles about
those announcements, and also feeding in some of
the actual financial statements that are released. And that takes a little bit of
natural language processing. You need some form of
taking the audio files, taking the written files,
and interpreting that. But with that data they're
looking for sentiment. They're trying to
interpret the sentiments and see if the stock
price are moving based on all of that data. Now, that's BlackRock with
$6 or $7 trillion of assets. They're deeply resourced,
fidelity, and so forth. But if you go down
the value chain, if you go down to
smaller asset managers, they're not doing a
lot, I would say, yet. But there are hedge funds
that are specifically saying, we are data analytic
hedge funds. We want to move
a little further. We want to try to use this
machine learning because it's a better way to
extract correlations, and that's what
it's looking for. It's pattern recognition. I think that you're going
to see more and more high frequency trading shops and
hedge funds exploring this. But one conversation I had
in the last couple of months with the high frequency
trading shop-- and it was a shop that
had about 100 employees. It was not big, but
it was big enough. It was certainly making
money in those days. I don't know how it's doing now. But they said, look, we feel
pretty good about what we do, and when we look for
algorithms, when we look-- our algorithmic trading doesn't
need all of that expenditure and all that resource
intensive thing of machine learning and cleaning up the
data and finding the patterns. And in fact, we think that is
not flexible enough yet for us. We're looking for short
term opportunities, and we think that regression
analysis and our classic linear and correlation analyses
are enough at this moment. What's going to come in
2023 is a different thing, but what they're doing now,
[INAUDIBLE] not really needed. Questions about asset
management just before I go to a couple other fields? ROMAIN: Now is the time to
raise your hand if you have any. I don't see any, Gary. GARY GENSLER: So cost-- STUDENT: Sorry, Gary, I
raised it in the last second, so Romain didn't see it. GARY GENSLER: [INAUDIBLE]. STUDENT: My question is-- so I fully understand
their claims that this is not the claims
of the asset manager, you were talking of
high speed trading. I don't see why machine
learning or any algorithm besides linear regression
doesn't have enough flexibility to actually replicate
what they're doing, but with a bit more of
accuracy or computing performance, et cetera,
because in the end, my understanding is
that you need the same-- except when you go
to deep learning. But you need the
same data sources that you currently have, and
the only thing that you're going to do is, instead of just
having linear models predicting the different traits
that you want to do, you have other
models that can find alternative basically
trends and et cetera. So I don't see the
limitation there of using it. GARY GENSLER: So I think you
raise a really good point. We're in a period of transition. I personally don't think
that machine learning and deep learning is the
answer to all these pattern recognition
challenges in finance. But you will find I'm more to
the sort of center maximalist than center minimalist,
and those of you that know me, that when we talk
about blockchain technology, you'll find I'm more to
the center minimalist side. But what I mean by that is
I think that the pattern recognition out of deep
learning, machine learning-- and by pattern recognition I
mean the ability to extract, with remarkable ability,
correlations and then create certain decision sets based
on those correlations-- is better than classic linear
algebra, classic regression analysis. But it comes with a cost, and
that's the tradeoff in 2020. Maybe in a handful of
years, it'll be less cost, but I'm going to use
an example, which is not about asset management,
but it's about lending. I had this conversation just
a few weeks ago with the CEO of a major peer to
peer lending company, and as this is
being recorded, I'm just going to maybe
not say his name. And I said, do you use
machine learning and deep learning to do your
credit decisions, all your credit decisions? And he said, yes, we use
a lot of alternative data. We run it into
the decision sets, and we see what patterns emerge. And I said, so you extend
credit then based on that? He said, well, not exactly. He said, what we do is,
we look for patterns, and then when we find
them, we then just use classic algebra and linear
flagging when we actually extend the credit. So we use it to
look for patterns, but then we sort of use the
traditional way, and I ask why. He said, well,
there's two reasons. It's less costly than running
the whole thing all the time, and two, they can explain
it better to regulators, and they can explain it
better to the public. And at least in
consumer finance, there were laws passed about
50 years ago-- in the US, it's called the Fair
Credit Reporting Act. These laws were
passed and said if you deny somebody credit you
have to be able to explain why you're denying them credit. And of course, there are other
laws in many other countries similar to that,
and there are also laws about avoiding biases. In our country, we call it
the Equal Credit Opportunity Act or ECOA, and
in other countries, similar things about
avoiding biases for gender and race and
background and the like. So I'm just saying that,
whether it's asset management, whether it's consumer
credit, there's some tradeoffs of these
new data analytic tools. And those tradeoffs, I think,
in the next handful of years, will keep tipping in the way
towards using deep learning. But they don't come-- they're not cost free
is what I'm saying. ROMAIN: José has his hand up. GARY GENSLER: Please. STUDENT: So you
talked a bit about how this is impacting the high
frequency trading shops. Is there something
similar happening on more like value investing
long term or in the hedge funds? So I heard some of them
are using satellite image to predict the number of cars-- to see the number
of cars [INAUDIBLE] and predict the sales for
that year, things like that. But I don't know, do you see a
lot of headroom in this area? GARY GENSLER: I think there
is headroom where I'm saying, I've mentioned two areas,
but I think you're right. There's a third area. The two areas are
sentiment analysis, just sentiment analysis around
either an individual company or the overall
markets and seeing what the sentiment, the
mood, the sense of the crowd is off of words,
images, or the like. And then also the
high frequency traders and so forth just looking for
the patterns in the short term trading. You're talking about more on
the broader value orientation, and I absolutely share
your view that we'll see more of that develop. I haven't heard a lot of it,
but you're right about sector after sector that you
might be able to analyze. But again, it has
to have some ability to extract a pattern better
than classic linear regressions in analysis. So let me just try to hit
a couple more of these because we're going
to run out of time, but we are going to talk
Monday more about these. We talked about call centers,
chat bots, robo-advising, and so forth, so I
think you've got that. A lot of that is
not just efficiency, but it's also inclusion. You can cover many,
many more people by automating some
of these tasks. It's just reality,
it's the tradeoff of efficiency and inclusion,
that you cover far more people. You can also be more targeted. You can be more targeted
with advice and so forth as these are automated. It comes with the tradeoff
that Ivy and Nadia were talking about earlier. Credit and insurance--
this is the concept of basically how you
allocate or extend or price, either alone
or price insurance. To date, the insurance companies
and insurance underwriting are starting to grapple with
this, starting to move on, and we'll talk about
some fintech companies in this space. I think that insurance companies
have been a little bit slower to do it than, let's
say, the credit card companies, but the allocation-- I think this will be
dramatically shifting. I think if you look at
what's going on, again, in consumer credit and small
business credit in China, through WeChat Pay
and AliPay, they're much further along than we
are, frankly, here in the US. We're still largely reliant on
a 30 or 40-year-old architecture around the Fair Isaac
Company, the FICO scores that are used in about 30
countries around the globe. These are still quite
limited, but FICO itself, FICO itself rolls out
new versions of FICO every few years. I think they're rolling
out FICO 10.0 this summer, or were before
the corona crisis. I think, if you look at
the end of the 2020s, either FICO will
not exist at all, or FICO 14.0 or 15.0
will look a lot more like a machine learning,
deep learning type of model. What it is right now
is pretty rudimentary compared to what it could
be in 5 or 10 years. This is an area that's
being used a lot-- fraud detection and prevention. The credit card
companies, if you look at Cap 1 and Bank of
America, Discover, American Express, they're deeply now
using machine learning tools for fraud detection. Many of us probably remember
just a year or two ago, you would still call up
your credit card company, and you would say I'm
traveling to France, I'm traveling to Italy,
wherever I'm traveling. I want to put a
flag on there that I might be using my cell phone-- using my credit card in
one of these countries. Well, most companies
now don't ask you to put a travel alert on it. Now, part of that
is because they know where we're traveling
because we're walking around with these location devices. The banks no longer need us
to call them to tell them we're going to be in
Paris because they know we're in Paris-- this location tracking
company device. But in addition to
that location device, they are also using
machine learning to do fraud detection and
prevention in the credit card space. Similarly, they're
using it to track and try to comply with laws
called anti money laundering. These two areas, fraud detection
and any money laundering, which I might call
compliance broadly, are two of the areas most
developed right now in 2020. That doesn't mean they'll
be the most developed later. I think a lot more will happen
in the underwriting space. A lot more will happen in
the asset management space. Robotic process
automation-- I want to just pause for a minute. Does anybody have a sense of
what these three words together mean-- robotic
process automation? Romain, you get to see if
there's any blue hands up. ROMAIN: Andrea. STUDENT: Hi. So robotic process
automation is very simple. You have a lot of
manual processes or manual work done
in, for example, back office of the banks. And the idea here
is, instead of people doing that, low skilled
work or workforce, you can actually teach robots
or the algorithms with the PC to do it instead of you. So for example,
it can be anything as simple as just
going through the forms and copying or
overwriting and rewriting some of the words or parts of
the forms to some other place. GARY GENSLER: Right. So robotic process
automation can be as simple as you're giving your
permission to a startup company, a fintech company,
to access your bank account. And one of us gives a fintech
company-- maybe Credit Karma. We give Credit Karma the
right to go in and look at a bank account of ours. Credit Karma might not have
permission from Bank of America go in, but they have my-- I'm permissioned them. I've given them my password
and my user ID, and they go in, and they want to automate. Credit Karma wants to automate. They don't want to have a human
actually have to type that all in. It can be as simple as just
automating inserting the user name and the password
and the like, but then you go further than
it can navigate the web page. It can navigate and
click the right buttons and get the right
data and so forth. So robotic process
automation can be helping a startup
company say Gary Gensler, we want you as a
client of Credit Karma. We, Credit Karma,
will figure out how to interface
with Bank of America and with Chase and the others. But also, the banks are using
robotic process automation to automate so much of their
both back office and their data entry. And one form, just
to say, is many of us now feel very
comfortable to deposit a check from our cell phone. So you can take your cell phone,
take a picture of a check, and somehow, that
gives an instruction, a digital instruction,
to move money. Well, part of that's
natural language processing, that the cell phone could take
a picture, read all that data, put it into computer language,
and actually move a digital. Part of that is robotic
process automation. Romain, was there a question? I think I saw some
flashing chat rooms. ROMAIN: Yes, sorry. You have 15 minutes left. GARY GENSLER: Oh, OK. Then was there a question or no? ROMAIN: No question
at this point. GARY GENSLER: So now, trading-- trading is an area I spent a
lot of time with Goldman Sachs. And in that trading
of the day, we were automating everything
we could automate, and this is in the 1990s. And ever since, anything is-- a trading floor in 2020 looks
very different than a trading floor in the 1990s in terms
of the day to day trading, and this is trading at
the center of the markets, the platforms themselves, and
of course, the high frequency traders on the other
end of the market is basically just
like asset management. What patterns can you see? Now, this is less
about value investing. This is the patterns
right in the nanoseconds and milliseconds and so forth. I'll make one note in terms
of trading, which is not related to machine
learning, but just related to the corona crisis that
we are all living in. I have an overall belief
that this coronavirus crisis will accentuate trends
that we've already seen. In industry after industry,
if we're locked down for two, three, or four months,
or God forbid, for 18 or 24 months-- if we're locked
down for that long, we're going to find
new ways to engage with each other in
economic activity and social activity
and the like. And we've already had
some trends, deep trends-- we talked about them Monday-- that we're unlikely to be using
many paper money and coinage money. Three months from now,
nearly 70% or 80% the world will have forgotten how to
use paper money and coinage. In fact, it will even be viewed
as a disease delivery device. It might be dirty. It might be something
we don't want to use because it could be a problem. Well, let me talk about
trading for a second. The New York Stock Exchange
and the world's largest stock exchanges are now electronic. They could have done
that two years ago. When the Intercontinental
Exchange, which is a big public company,
bought the New York Stock Exchange a handful of years
ago, five or so years ago, Jeff Sprecher, the
entrepreneur who started the Intercontinental
Exchange in 1998 or '99, he's always been an
entrepreneur, an innovator, to do electronic trading. They could have taken the
New York Stock Exchange fully electronic, but guess what? That's what happened in
the last three weeks. So after we get out of
this lockdown period, will we bring back the floor
of these London and New York and Shanghai and
Mumbai and so forth? Will we bring back the floors? I think quite possibly not. Not sure, but there's a lot
that's shifting on, I think. Natural language
processing-- we talked about in customer service,
process automation, and sentiment analysis. These are sort of the
slices that I think about in these fields. So I was curious how many
people have ever used Siri? Probably almost every
hand would go up. But how many people
have ever used Erica? ROMAIN: So perhaps we
have Shaheryar who'd like to share his experience. Sorry for your name--
mispronouncing it. STUDENT: Yeah, so it's
essentially like Siri, but you actually-- it's a
product of Bank of America, and you can use it to
check your spending habits. You can also use it to, if
you need things with regards to check depositing, or if
you want to know something, it can do that. But currently, I believe
it's not as refined as Siri, and I still think there
is a lot room over there for improvement. GARY GENSLER: Why do you
think it is that Erica-- and JP Morgan has one as well. I can't remember
what her name is. They all do seem to be
mostly female voices, if I'm not mistaken. But why is it that
Erica and the like, a virtual assistant, as
they're called in finance, aren't as developed as the
Siris and Alexas, do you think? STUDENT: Sorry, can you
repeat the question? GARY GENSLER:
Anybody can answer. Why is it that the finance
virtual assistants like Erica are not yet as fully developed
as the Home and other ones like Siri and Alexa? STUDENT: I believe some of-- I think it's got something
to do with the fact that the number of users
for financial assistance is way lower as compared
to Alexa, Siri, or Alexa or whatever. So I believe that is something
which may explain it, the disparity between these
two kinds of assistants. GARY GENSLER: Yeah. Any other thoughts
on that, one why-- ROMAIN: Nikhil has a different
take, and that we have Laira. STUDENT: Probably along
similar lines, banks so far, the interactions probably have
been in person or over phones, and they weren't used to
processing data and processing requests. I think they have
a smaller data set to go through and understand
what problem customer questions are. And that's probably a limiting
factor versus, say, Siri, they have much more
data on everyday users. I think that's probably
the biggest differentiator. GARY GENSLER: Yeah, and was
there another comment, Romain, you said? Sure. ROMAIN: So I think
Laira had her hand up, but I think she withdrew it. So perhaps we can
hear from Brian. GARY GENSLER: OK. STUDENT: So in
addition to the data, I think there's also a
human capital element. It's possible that Apple
has better human capital capabilities than do these
financial institutions, so it's harnessing
that data as well. GARY GENSLER: Yeah. So what we've just
talked about was data, human capital, experience-- all true. Also, Erica and
the financial firms only started more
recently, and so forth. But the voice recognition
programs and then taking that data that an
Apple has or their competitors in big tech around the globe-- because it's not
just here in the US-- is really remarkable now. Even to the extent
that I don't know how many people use earbuds,
but if you look closely at the user agreement on the
airbuds that you use, it says-- and if it's an Apple,
if I'm mistaken, the Apple lawyers will chase me. But the user agreement
says that they can listen to that to
help you to make sure that there's not a drop
between your earbuds and your cell phone. They are picking up vast amounts
of data, vast amounts of data from our text messaging. If you look at
something like Google, they're picking it up from
Google Chrome, Google Maps, Gmail. Multiple places that
they can pick up, and we talked about this
conceptually, big tech versus big finance versus
startups in this triangle of competitive landscape. And why I wanted to sort of
close on Bank of America Erica and this discussion
of Erica versus Siri and Alexa is big tech using
Google, just as an example, has this remarkable network that
they're layering activities. Remember, we said data
network activities-- that's the Bank of International
Settlement way to put it, and what a perfect
example to show. Google has Gmail. It has Maps. It has Google Chrome. It has Android, the
operating system. So all of these different
ways to build their network, and they layer
activities on top of it, and then vast amounts
of data come in and the human capital that was
mentioned at the end there. And they have more experience
to move it forward. Apple, similarly. Baidu and Alibaba in China
and so forth, similarly. If I were a CEO of
a big incumbent, yes, I would be very focused
on the fintech startups, but I tell you-- I'd be looking at
big tech in a way that their advantages are
really significant, very significant advantages. Romain, I see some
hands up maybe. ROMAIN: We now have Laira
who has her hand up. STUDENT: Yeah, I'm just curious
to know what do you think? So currently, we know that tech
has kind of-- or AI has kind of emerged to the financial
and payment space in the form of
virtual assistants, but what do you anticipate the
next step would be in terms of this integration
of technology and AI into the payment space? What next after the
virtual assistants? GARY GENSLER: We're going to
have a whole class on payments specifically, but
I think that what's happening in the
payment space is we've seen specialized
payment service providers. Of course, we've seen a
lot of the competition starting with PayPal in 1998. This is not a new
space for disruption. But what we've
seen more recently is, in the retail space, whether
it's companies I've mentioned earlier, like Toast, that
got into one vertical, one slice within payments,
which was restaurants-- they can provide a better
product for that slice. And then can collect back
to AI enough robust data within that slice-- this is using Toast
as an example-- that they can provide better
software, better hardware, and also less risk loans. Basically, as Toast
started to provide lending to restaurants within that
space, built upon the payment. So it's the marriage between
the user experience providing the users-- in that case, the restaurants
in the payment space-- but providing the users
in that space something that the generalized platform-- a bank payment system
is generalized. It's multi sector. It's a general
product, and Toast was able to say, no, we
can provide something that just restaurants-- there might be something
a little unique about the restaurant
business that we can provide software and hardware. In their case, it was tablets. They were providing
tablets for the servers to walk around and
take the orders. They could integrate the menu
right into the payment app. So there was something a
little bit unique about that. But then based on that,
they get a bunch of data, and that data helps
them with, I would say, underwriting
decisions based on-- it doesn't have to
be machine learning. But it's enhanced data analytics
because of machine learning. So I hope that helps. I think on the
conversational agents and the virtual assistants,
what we're seeing in the payment space, because that
was your question, is we're moving from
card authorized payments to mobile app QR codes. Then the QR codes is not
based upon virtual assistants, but it's an interesting
question whether we'll get to some voice
authenticated payments. There are a lot of uses of
voice authentication already. Vanguard and many
other asset managers, where you can have your
brokerage accounts, you can call in and
get voice authenticated before you can do a trade. And that voice
authentication is just like other forms
of authentication, but it's not perfect
as of this moment.
