AI in fintech

is utilizing data

and specialized hardware

and algorithms--
machine learning algorithms--

to augment humans

and allow people

to utilize this technology

to do things that they never imagined

that they could do before in fintech.

What are the tasks that are being

automated and augmented?

They're these-- these nine:

market research, accounting,

assistance, investment,

credit, insurance, collections,

predictive analytics, and compliance.

So, I'll give you some examples

of each one of these,

and then we'll look at some 2017 results

in machine learning and AI.

We'll talk about new applications

of machine learning and AI

in a variety of different areas

that you might want

to invest in, outside of fintech,

and then what some

of the ethical implications are.

So, in market research,

there are companies like DataMiner

that are not originally

focused on fintech.

They vacuum up vast

amounts of information

and allow people who subscribe

to certain vectors of information--
like, finance or in weather
or in medicine--

to get specialized information,

just for them, that's pulled

out of that massive database.

And then, we have

companies like AlphaSense

that are specialized in fintech

and allow financial analysts

to ask questions they've never

asked before to make it easier

to ask those questions

and to get answers--

not just references,

like you would on Google.

In accounting,

there are companies like AppZen

that do auditing of expense reports.

Now this is a horrible task--

it tends to be very tedious,

people have to look

for needles in a haystack--

and what we're doing now

is using machine learning

to identify the needles.

Then, humans can look

at that stack of needles

and decide where

can we make an exception

and where do we have

some evidence of fraud.

And then there are companies like Zeitgold

that are producing AI-fueled

accounting systems on the cloud

and on smartphones, directly.
So that professionals
and small businesses

can run their entire operation,

right on their own smartphone.

In Assistants,

we have companies like Kasisto

that have KAI Banking,

which is a consumer-facing personal banker

that allows people to do almost anything

that they would do with a personal banker,

directly on their smartphone.

Now, Amazon's Alexa

is not specialized in finance,

but it has a 70% market share

and look at the growth of skills here!

in Q1 2016: 135 skills.

In Q1 2017, over 10,000 skills;

now many of them financial.

For example, "ask Capital One

to make a credit card payment,"

"get the rate for a credit card

from Virginia Credit Union,"

"ask Ameritrade for a market update,"

"ask Fidelity to get a quote for Amazon,"

"ask Liberty Mutual

for auto insurance estimates."

These are all vocally-enabled searches--

something that was very

difficult to do five years ago.

And here is a company called Digit

that addresses a problem

that is often associated with millenials,

but a lot of people have this problem:

they have difficulty saving.

And what this system

does is it does a forecast
of the cash flow in, say,
a student's bank account--

in their checking account--

and every so often,

will make an incremental transfer

into their savings account.

Of course, it can be reversed,

if they want to do that.

But this is a machine learning-fueled app

that allows people to save

with a little help from AI.

And here's investment--

Numerai is a hedge fund--

that is a crowd-sourced hedge fund--

with billions of individual predictions

and hundreds of thousands of models.

And they reward

their crowd-sourced partners--

not for getting the best model--

because they're not looking

for the best model--

they're looking for a model

that contributes
to their synthetic model

in their hedge fund.

And they pay their partners in bitcoin.

This company, Sentient,

came out of part

of the team that built Siri.

And this team has put together

an evolutionary platform

that builds little agents

that are hedge fund managers--

that make predictions--
and they evolve them--
they compete them against each other--

they take the top of the class

and they go again, and again, and again--

until they finally end up

with very, very sophisticated

hedge fund managers.

Credit is another big area for fintech.

Affirm, from Max Levchin,

is a company

that is all about transparency and honesty

in financial transactions.

They're trying to transform the landscape,

particularly for young people.

DELIVER HONEST FINANCIAL

PRODUCTS TO IMPROVE LIVES

This company, ZestFinance,

is focused on those people

who have a thin file--

that is, there's not a lot

of financial data about them.

And they want to enable those people

to be checked out

and for those people

who are worthy of getting a loan,

they can get a loan

even without a lot of credit history.

Insurance is another area

that is being disrupted.

You're going to hear

later on today, about Lemonade--

it's all about using machine learning

in every aspect
of the underwriting process,

of the sales process,

of interacting with customers--
this is a B Corp.

They take 20% of their premium--

they get a 20% fee on premiums--

and for people who underclaim,

that underclaimed money

goes to their favorite charity.

So they're doing some good in the world

as well as transforming the industry.

Cape Analytics is using satellite data--

and as well as other data--

to validate the claims that people

make for insurance on their properties.

So, they can find out what kind of roof

you actually have

and what condition it's in.

They can look at the square footage

of your property and check

out the square footage directly.

Collections are another big area--

collectAI is all about emulating

collection agents

including their tone of voice.

So, they want to learn, incrementally,

what works for getting

collections in the door,

and they get better

and better as time goes on.

TrueAccord takes a somewhat

different approach to collections.

They recognize that sometimes

collection problems are the result

of either misunderstandings,

or sometimes, lack of clarity,

and sometimes,
deliberate falsification
of who owes what to whom.

So they look at both sides

of the equation,

and they try to come to a true accord

of what's owed and what isn't.

And then they get very high compliance

once they get agreement about that.

Predictive analytics is another key area.

This company, Opera,

is doing amazing work with signals--

they have a Signal Hub--

and they have descriptive signals

that are pattern-recognized from the world.

And then they take those signals

and they use them in prediction.

So, they have predictive signals

that are all about

conditions of the market,

conditions of the world,

that can be acted on.

So they are looking for proactively set up

predictions on actionable information.

Kensho is a company that has

one of the world's largest

global event databases.

And they have knowledge graphs

associated with their event database.

And they are enabling

people who are interested

in events in the world,

to make ever more interesting

and precise forecasts

about what those events mean.

And in compliance, we have Trifacta--

Trifacta came out of a project
between Stanford, UC Berkeley,

and University of Washington--

it was called Data Wrangler.

It was all about marshalling

the data, and cleaning it,

and getting out the noise--

a key task in machine learning.

And Trifacta is all about enabling

data marshalling and data cleaning

for further analysis.

This company-- Digital Reasoning--

came out of the military

and intelligence analysis world--

they're in Franklin, Tennessee--

large percentage of them--

about 30% of them

have top secret clearances.

But they're now focusing

on the financial world.

And they, in particular, are able to look

at communications inside a company--

like email and other kinds

of communications--

and look for signs of fraud.

So, the question is what is AI

that is powering all this innovation?

Well, it's pattern recognition techniques

like deep learning

and other machine learning techniques,

that allow us to make inferences

from vast quantities of data.

It's software agents that may combine

a pattern recognition technique

like deep learning

with reinforcement learning,

that reinforces an agent for a high score

in an arbitrary task,
like a game or a prediction.

It's a vision of superhuman intelligence

that really hasn't happened yet,

but movies and science fiction

like to blur that line;

it's important for you to keep

your critical thinking skills

intact around that.

And it's computer science

that's accelerating

all the other exponential technologies

like synthetic biology, nanotechnology,

robotics, and yes... even AI.

People sometimes ask

about the relationship

between deep learning

and machine learning--

deep learning is a subset

of machine learning;

machine learning
is a subset of AI;

AI is a subset of computer science.

Now, even with these definitions,

people sometimes get confused.

True story, I was recently on a flight

from Seattle to San Francisco.

The guy next to me

is wearing a cowboy hat.

He says,
"Neil, what do you do in real life?"

I said, "Well, I'm in the AI business."

He said, "Really? I am, too.

How many head do you have?"

And I thought, "What is he talking-- oh!

- He means artificial insemination."

- (audience laughter)
So, there is a relationship.

You can use AI to select the genes

of your beef or dairy cattle.

Here are 100 startups that are using AI

and machine learning

to transform industries.

And the thing to get about this is:

Yes! We're using machine

learning all across fintech,

but look at these other areas--

these are the companies

that hedge funds invest in

outside of the fintech arena.

Companies that are auto companies,

robotics companies,
cybersecurity companies,

sales companies, healthcare companies,

commerce, Internet of Things,

business intelligence-- across the board,

AI and machine learning

are disrupting the landscape.

So, deep learning is a hierarchical

pattern recognition algorithm.

What's changed since 2009

is that we can now have deep stacks

of these recognition layers,

because we can preserve coherence

in between the different

layers of the stacks.

Let me show you how this works...

Yeah...

So if we start with a set

of pixels-- images--

we can recognize, in the first stage

of the hierarchy, just edges--

vertical line,
horizontal line, diagonal line.

Second stage: object parts.

Ears, nose, eyes, chins.

Third stage: objects, like faces.

But three layers

are not enough to differentiate

all of your beautiful faces;

we need more layers for that.

Microsoft has a pattern recognition

hierarchy with 152 layers;

you don't need anywhere

near that to achieve

good facial pattern recognition.

Look at the money

that's now going into AI.

In 2012: $589 million.

2016: $5 billion.

And we're seeing major

changes in hardware;

hardware that's specialized

for machine learning and AI.

This is a tensor processing unit board.

This system is blistering fast;

it just was announced by Alphabet.

It is specialized for running

machine learning algorithms

like TensorFlow, that Amin mentioned.

This system allows you to train

and test on the same board,

which was not the case

with their first version of this.

And this system has 180 teraflops--

180 x 10^12 logical operations per second.

It is blistering fast.

And it can be woven

into a larger fabric of computing--
if you see all these blue--
these two blue racks--

together, those two blue

racks are 11.5 petaflops--

10^15th logical operations per second,

or about a human equivalent

worth of computing power;

not human intelligence,

but just raw computing power.

I'm encouraging you to try AI

and machine learning at home.

You can download

the R machine learning library

right onto your laptop--

if you don't have

the chops to use it yourself,

you can invite a high

school student to help you.

R has 1,800 free extensions,

and it itself is free.

TensorFlow is free
for researchers; it's vastly powerful.

It allows you to represent

your problem as a set of algorithms

in the node of a data flow graph

and the edges that connect those nodes

are multidimensional data arrays.

So then, you can spread your problem

across a broad array of computer fabrics.

What's happening now

is an emerging synthesis

of human intelligence
and machine intelligence.

All of you have heard about the recent win

of AlphaGo in China,
and in Korea a year before that.

Just for context--

if you think about chess,

it has an average number

of moves per turn--

alternative moves-- of about 35.

And 35 after that, 35 after that;

fans out very quickly.

Go has an average number

of alternative moves per turn

of 200-- it is epically complex.

And just a few weeks ago,

Alphabet's AlphaGo player

won against Ke Jie,

the world champion in Go--

Chinese gentleman--

and he was amazingly graceful

and played games

with machine learning algorithms

after a set of matches that he lost.

These are results that would have

been unthinkable 10 years ago.

Let's look at Libratus,

which was developed by Tuomas Sandholm

and Noam Brown at CMU--

Carnegie Mellon University.

This is a system that plays

no-limit Texas hold 'em poker.

It played 120,000 hands,

just to make sure

there was no fluke results,

against four world champions in poker.

And it won 1.77 million

dollars in poker chips.

An amazing result; why should we care?

Because this system
can deal with bluffing,

and misinformation, and strategy.

You may see this as a quiz.

What's happening
in the right cell of this matrix--

this is the Raven

Progressive Matrices test.

It is a test of analogical
and visual reasoning.

The right answer is D,

because it matches the symmetry

of the column and the row.

Ken Forbus, from Northwestern,

has developed a system called CogSketch

that outperforms
the average American on this test;

it's in the 70th percentile

of test takers.

IBM beat the two world champions

in Jeopardy way back

in ancient history-- 2011.

But now, IBM is moving Watson

into virtually every area

of human endeavor,

including financial services.

The background on the AlphaGo player

was a system called the Atari game player

developed by DeepMind in the U.K.

They gave a demo in 2013

of a powerful pattern recognizer;

a convolutional neural network

that was combined

with reinforcement learning--

which just means a powerful

way of reinforcing an agent

for a high score in the game.

It starts out with zero knowledge

of a bunch of Atari games.

Watch this--

This is Demis Hassabis, CEO of DeepMind,

and this Atari game player

is playing "Breakout"

and just missed a ball--

it'll miss more; another one.

After 120 minutes of play,

it's playing at a mediocre human level.

And then, after 240 minutes of play,

it's going to tunnel

up the left-hand side of the screen

and play from the back court

where the game has no defenses.

Look at the score.

Pretty amazing.

People sometimes describe AI

as a game-changing technology--

it's more than that.

AI is disrupting the entire playing field

that we are operating on.

Not just metaphorical.

If you don't have

bench strength in AI,

you can put up prize money

and your data and your problem

and get two or three hundred

of the world's best data scientists

to bid on the solution to your problem.

If you don't want to put your data

out in public, no problem--

By the way, Kaggle just got bought.

The place where they're doing

machine learning competitions

just got bought by Google Cloud this year.

An alternative to that--
you don't have to put your data out--

would be Experfy.com

They allow you to just

put out your problem;

have very well-qualified

and curated data scientists

and AI people bid

on the solution to your problem.

And it's been very successful.

In the spirit of full disclosure,

I'm an investor here.

There are three problems that I want

to call your attention to with AI;

I'm going to have to go

over them pretty quickly.

One is lack of transparency;

AI, machine learning

tend to be like a black box,

particularly machine learning.

But there's a program at the Defense

Advanced Research Projects Agency--

called XAI at DARPA--

run by Dave Gunning--

and it's all about opening

up the black box of machine learning

so that we get good explanations.

So instead of an explanation here

like, "This is a cat,"

it tells you why it thinks it has a cat

in that image.

Another problem is that if you train

a machine learning system

like deep learning on one test-- Task A--
and then you train it on Task B,

it can sometimes undergo catastrophic

forgetting of how to do Task A.

And some folks at DeepMind, in the U.K.,

came up with a new

method of dealing with this,

called Elastic Weight Consolidation.

This is something that nature

figured out with mammalian cortex.

And what they're doing,

is preserving the knowledge

of how to solve the prior task

by providing a penalty
for forgetting the first task.

They're getting very good results

with that and it's not just two steps,

you can do end steps.

I consider that a real breakthrough

in machine learning, although it was

never labeled as such by them.

And then your mileage can vary.

In 2016, Unanimous used human intelligence

and machine intelligence to correctly

predict the winner of the Kentucky Derby.

2017, they didn't get it right.

So, again, keep your critical

thinking skills intact.

AI's been sponsored

by companies all over the world.

I added Nvidia, Amazon,

Uber and Salesforce to this list;

they're all doing amazing work

in AI research and development.

Amazon has a thousand people

on their Alexa team.
I did a study of 360
innovative applications of AI--

it's not just better,

faster, cheaper; it's different.

AI allows us to expand
the range of the possible.

And you can work in any of these areas--

I've seen great applications in design,

diagnosis, manufacturing
and management,

customer service, sales

and configuration, and quality.

Here's an example of AI in medicine

from Imperial College London.

This system predicts

the one-year survival rate of people

with pulmonary hypertension

and no robotic sleeve.

And cardiologists tend

to be 60% accurate with that;

this system is 80% accurate.

If you're interested in AI in the law,

this company called Intraspexion™

will look at emails sent to your company

like, "I'm going to sue your pants off."

And try to differentiate the cranks

from the people who really will sue you.

And if you want to use AI

in materials development,

you can insert

a problem solver-- an AI solver--

in between a high throughput

experiment and people.

And get the best of human intelligence

and machine intelligence

to do rapid experimentation.

This was the winning paper

called Phase-Mapper
at the Innovative Applications
of AI Conference in 2017.

GE has a fabric that you can

lay over a dumb factory

to make it smart, more predictive,

more responsive, more connected.

And Carnegie Learning

has systems that can model the user

and provide one-on-one tutoring

for people who want to learn new areas.

AI is going to impact Human Resources,

augmenting people and adding machines,

transforming business models

like that radically-altered model

that I showed you of a soccer field--

but that means that we are going

to have to act proactively--

I think it means that we're going

to have to consider

various basic income experiments

and free education.

The adult conversation

is that AI comes with tradeoffs.

Yes, faster, cheaper, better,

different problem-solving--

but also job disruption

and human identity change

and risk amplification.

What it means to team with AI strengths

is that you get best

in class performance--

for some tasks--

you get improvements

in speed and prediction.

No vacations or sick leave, no whining,

but limited empathy,

language understanding, and social grace.
Trust is going to be
a big issue going forward.

There have been a number

of conferences sponsored by Future of AI--

one in Puerto Rico--

where we talked about validity,

verification, security, and control,

as methods to ensure
that we have beneficial use of AI.

There was a January 2017 conference

that came up with 23 Asilomar principles

for the beneficial use of AI.

Elon Musk and Sam Altman at Y Combinator

have created OpenAI

to democratize access to AI

and provide both software and reports

to make sure that people

can have access and use it safely.

There's a partnership in AI now,

with Amazon, DeepMind, Google,

Facebook, IBM, and Microsoft,

that's all about ensuring

beneficial use of AI.

Here's some operational recommendations--

if you have small amounts of capital,

you can innovate large amounts of capital.

You can acquire and integrate.

You can use free algorithms

or sophisticated platforms.

There's lots of hardware available,

including very tiny hardware

for the companies

that you might invest in,

that are doing Internet of Things.

For data, you can marshal

the data that's available to you--

not just big data, but relevant data.

For talent, you can crowdsource your data;

if you don't have deep bench strength.

For applications,

innovate on product services

and internal operations.

And with regard to responsibility,

consider security, empathy,

ethics, policy, and liability.

I'm encouraging you all

to build an application oriented AI

toolbox for your organization;

it will pay off.

Here's what you can do with it--

you can disrupt

the financial playing field with AI.

You can augment

and automate white collar jobs.

You can build that toolbox

for application problem-solving;

you can become

an exponential organization--

Salim Ismail will be talking to you more

about that later on in the summit.

The most important thing

is to design for a world

that you actually want to live in,

if you don't know what part

of the game board you're on.

We're encouraging
you to build the future boldly,

and to do it responsibly.

- Thank you very much.

- (applause)
490
00:28:05,256 --> 00:00:00,000
♪ (upbeat outro music) ♪