Welcome to Brown Biostatistics Seminar.

And I'm Rossi Luo, faculty host for today's event.

And for those of you new to our departmental seminar,

the format is usually that the presentation

followed by a question and answer session.

And because of the size of crowd today, we

are going to also use this red box

thing to capture your questions and for videotaping and also

make sure your questions are heard.

And today I'm very pleased to introduce Professor Yann LeCun.

Professor LeCun is a director of Facebook AI Research,

also known as FAIR.

And he is also senior professor of computer science,

neuroscience, and electronic computer engineering

at New York University.

He's also the founding director of NYU Center for Data Science.

Before joining NYU, he had a research department

for industry, including AT&T and NEC.

Professor LeCun has made extraordinary research

contributions in machine learning, computer vision,

mobile robotics, computational neuroscience.

Among this, he's a pioneer in developing

convolutional neural networks.

And he is also a founding father of convolutional nets.

And these works contributed to say

the creation of new an exploding field in machine learning

called deep learning, which is now

called artificial intelligence tool for various range

of applications from image to natural text processing.

And his research on contributions

has earned him many honors and awards

including the election to the US National

Academy of Engineering.

Today he will give a seminar titled,

How Can Machines Learn as Efficiently as Animals

and Humans.

I understand some of you actually

told me you drove from Boston or many places are very far.

So without further ado, let's welcome Professor Yann LeCun

for his talk.

[APPLAUSE]

YANN LECUN: Thank you very much.

It's a pleasure to be here.

A game I play now occasionally when I give a talk here is I

count how many former colleagues from AT&T are in the room.

I count at least two.

Chris Rose here, Michael Litman.

Maybe that's it.

That's pretty good, two.

Right.

So, how can machines learn as efficiently

as animals and humans?

A have a terrible confession to make.

AI systems today suck.

[LAUGHTER]

Here it is in a slightly less vernacular form.

Recently, I gave a talk at a conference in Columbia

called the Compositional and Cognitive Neuroscience

Conference.

It was the first edition.

And there was a keynote.

And before me, Josh Tenenbaum give

a keynote where he said this.

All of these AI systems that we see now, none of them

are real AI.

And what he means by this is that none of them

actually learn stuff that are as complicated as what

humans can learn.

But also learn stuff as efficiently as

what animals seem to learn them.

So we don't have robots that are nearly as

agile as a cat for example.

You know, we have machines that can play golf better

than any humans.

But that's kind of not quite the same.

And so that tells us there are major pieces of learning

that we haven't figured out.

That animals are able to do that, we don't do--

we can't do with our machines.

And so, I'm sort of jumping ahead here

and telling you the punch line in advance, which

is that we need a new paradigm for learning,

or a new way of formulating that has old paradigms that

will allow machines to learn how the world works the way animals

and humans do that.

So the current paradigm of learning

is basically supervised learning.

So all the applications of machine learning,

AI, deep learning, all the stuff you see the actual real world

applications, most of them use supervised learning.

There's a tiny number of them that

use reinforcement learning.

Most of them use some form of supervised learning.

And you know, supervised learning, we all--

I'm sure most of you in the room know what it is.

You want to build a machine that classifies cars from airplanes.

You show an image of a car.

If a machine says car, you do nothing.

If it says airplane, you adjust the knobs on the machine

so that the output gets closer to what you want.

And then you show an example of an airplane.

And you do the same.

And then you keep showing images of airplanes and cars,

millions of them, thousands of them.

You adjust the knobs a little bit every time.

And eventually, the knobs settle on a configuration,

if you're lucky enough, that will distinguish every car

from every airplane, including the ones

that the machine has never seen before.

That's called a generalization ability.

And what deepening has brought to the table there,

unsupervised learning, is the ability

to build those machines more or less

numerically with very little sort of human input

in how the machine needs to be built,

except in very general terms.

So the limitation of this is that you

had to have lots of data that has been labeled by people.

And to get a machine to distinguish cars

from airplanes, you need to share

with thousands of examples.

And it's not the case that babies or animals

need thousands of examples of each category

to be able to recognize.

Now, I should say that even with supervised learning,

you could do something called transfer learning, where

you train a machine to recognize lots of different objects.

And then if you want to add a new object category,

you can just retrain with very few samples.

And generally it works.

And so what that says, what that tells

you is that when you train a machine,

you kind of figure out a way to represent the world that

is independent of the task somehow, even though you train

it for a particular task.

So what did deep learning bring to the table?

Deep learning brought to the table

the ability to basically train those machines

without having to hand craft too many modules of it.

The traditional way of doing pattern recognition

is you take an image, and you design a feature extractor that

turns the image into a list of numbers that can be digested

by a learning algorithm, regardless of what

your favorite learning algorithm is,

linear classifiers, [INAUDIBLE] machines, kernel machines,

trees, whatever you want, or neural nets.

But you have to preprocess it in a digestible way.

And what deep learning has allowed us to do

is basically design a learning machine

as a cascade of parametrised modules, each of which

computes a nonlinear function parametrised

by a set of coefficients, and train the whole machine end

to end to do a particular task.

And this kind of an old idea.

People even in the 60s had the idea

that this would be great to come up

with learning algorithms that would train multilayer systems

of this type.

They didn't quite have the right framework if you want,

neither the right computers for it.

And so in the 80s, something came up

called back propagation with neural nets

that allowed us to do this.

And I'm going to come to this in a minute.

So the next question you can ask of course

is what do you put in those boxes?

And the simplest thing you can imagine as a nonlinear

function, it has to be non-linear,

because otherwise there's no point in stacking boxes.

So the simplest thing you can imagine is take an image,

think of it as a vector, essentially.

Multiply it by a matrix.

The coefficient of this matrix are going to be learned.

And you can think of every row of this matrix being

used to compute a dot product with an input vector.

And that produces basically a weighted sum

of the inputs multiplied by those coefficients.

That gives you another vector.

And you pass each component of these vector

through a non-linearity like this one, for example.

Just halfway ratification.

So you have two different steps.

Linear, nonlinear.

Linear pointwise, nonlinear.

Very simple.

And you can show that by stacking two layers of this,

you can approximate any function you want, as close as you want

as long as you have sufficiently many of these guys

in the middle by tweaking the parameters of the two layers.

But in fact, most functions we're interested in

are more economically represented by many layers.

And so that's the new approach to deep learning, if you want,

that changes from the neural nets of 30 years ago,

which typically had only two or three layers.

The neural nets of today, the deep learning systems of today

have anywhere between 20, 50, or 100 layers.

OK.

So we have linear operators that are

parametrized by coefficients.

And the supervised learning, we're basically

going to train it to be some sort of objective function

that's going to measure the discrepancy

between the output the machine produces and the output

we want.

And so the subjective function is going to be differentiable.

What we're going to do is compute the gradient

of the objective function with respect

to all the parameters in the machine averaged

over a number of training samples.

Or if we use stochastic gradient decent,

averaged over a small batch of training samples,

or even a single sample.

And then take one step [INAUDIBLE]

to get your gradient using the stochastic gradient update

rule.

Basically, the parameters are going

to kind of go down to a minimum in a stochastic fashion

as you train more and more.

So now the next step you have to do

is compute the gradient of the objective function

with respect to the parameters.

And the way you do this through back propagation.

I'm not going to go through this.

The mathematical concept on which it's based

is incredibly sophisticated.

Is it's called chain rule.

[LAUGHTER]

And some people learn this in high school.

And it basically comes down to the fact

that if have-- if your ranged parametrized functions

in a graph of competition, which in this case

is a very simple one.

It's just a linear stack of modules.

But it doesn't need to be such a simple graph.

It could be any graph.

And you [? take ?] connection by propagating signals

backwards through this graph.

Basically taking the gradient of some cost

function you want to minimize with respect

to this red variable.

And so this gradient is represented

by these green variable.

And multiplying it by the Jacobian of this box,

you get the gradient respect to the input of that box.

This is chain rule.

So it's this guy here.

Gradient with respect to the input

equals gradient with respect to the output

multiplied by Jacobian.

Very easy.

And so you propagate this backwards through the graph.

And the cool thing about this is that you can do this