0:16

- It's a, I guess Computer Science's attempt

to mimic a real,

the neurons and how our brain actually functions.

So 20, 30 years ago a neural network

would have some inputs that would come in

they would be fed into different processing nodes

that would then do some transformation on them

and aggregate them or something

and then maybe go to another level of nodes

and finally some output would come out.

And I can remember training a neural network

to recognize digits, handwritten digits and stuff.

(music)

1:00

So a neural network is trying to use

a computer program that will mimic

how neurons, how our brains use neurons to process things,

brains to synapse, neurons to synapses

and building these complex networks that can be trained.

So a neural network starts out

with some inputs and some outputs

and you keep feeding these inputs in

to try to see what kinds of transformations

will get to these outputs,

and you keep doing this over and over and over again

in a way that this network should converge

so these input, the transformations

will eventually get these outputs.

The problem with neural networks was that

even though the theory was there

and they did work on small problems

like recognizing handwritten digits and things like that,

they were computationally very intensive,

and so they went out of favor.

I stopped teaching them,

well, probably 15 years ago.

Then all of a sudden we started hearing about deep learning.

I heard the term deep learning.

This is another term that

when did you first hear it?

Fours years ago, five years ago.

So I finally said,

"What the hell is deep learning?

It's really doing all this great stuff.

What is it?"

I Google it and I find this is neural networks on steroids.

What they did was they just had more

multiple layers of neural networks

and they use lots and lots and lots

of computing power to solve them.

Just before this interview

I had a young faculty member in the marketing department

whose research is partially based on deep learning.

She needs a computer that has

a graphics processing unit in it

because it takes an enormous amount of matrix

and linear algebra calculations

to actually do all of the mathematics

that you need in neural networks,

but they are now quite capable.

We now have neural networks and deep learning

that can recognize speech, can recognize people.

If you're out there and getting your face recognized

I guarantee that NSA has a lot of work

going on in neural networks.

The University, right now,

as Director of Research Computing,

I have some small set of machines

down at our South Data Center

and I went in there last week

and there were just piles and piles and piles

of cardboard boxes all from Dell with a GPU on the side.

Well, a GPU is a graphics processing unit.

There is only one application in this University

that needs 200 servers,

each with graphics processing units in it,

and each graphics processing unit

has the equivalent of 600 cores of processing,

so this is tens of thousands of processing cores.

That is for deep learning.

I guarantee.

(music)

4:12

Some of the first ones are speech recognition.

Yann LeCun who teaches the deep learning class at NYU

and is also the Head Data Scientist at Facebook

comes into class with a notebook

and it's a pretty thick notebook.

It looks a little odd because it's like this.

It's that thick because it has

a couple of graphics processing units in it

and then he will ask the class

to start to speak to this thing

and it will train while he's in class,

he will train a neural network to recognize speech.

So recognizing speech, recognizing people, images,

classifying images, almost all of the traditional tasks

that neural nets used to work on in little tiny things,

now they can do really, really large things.

It will learn, on it's own, the difference between

a cat and a dog and different kinds of objects.

It doesn't have to be taught.

It doesn't, it just learns.

That's why they call it deep learning,

and if you hear, he plays this.

If you hear how it recognizes speech and generates speech,

it sounds like a baby learning to talk.

You can just, you're like

(babbles)

5:55

You need to learn some linear algebra.

A lot of this stuff is based on matrix

and linear algebra, so you need to know how to do,

use linear algebra and do transformations.

Now, on the other hand,

there's now lots of packages out there

that will do deep learning

and they'll do all the linear algebra for you,

but you should have some idea

of what is happening underneath.

Deep learning, in particular,

needs really high powered computational power,

so it's not something that you're going to go out

and do on your notebook for, you could play with it,

but if you really want to do it seriously

you have to have some special computational resources.

(music)