LAURENCE MORONEY: Hi, everybody.

Laurence Moroney here on my TensorFlow World.

And we've just come from the keynote

that was given by Jeff Dean.

And so Jeff, welcome, and thanks for coming to talk with us.

JEFF DEAN: Thanks for having me.

LAURENCE MORONEY: So you covered lots of great contents

in the keynote, and there were so

many things that we don't have time to go over them all.

But there was one really impactful thing that I saw.

And you were talking about like in computer vision.

Now, the error rate in humans is like 5% in computer vision.

And now with machines, it's down to 3%,

and that's really, really cool.

But it's more than just a number, right?

What's the impact of this?

JEFF DEAN: Right.

I mean, it's important to understand

this is for a particular task that humans aren't necessarily

that great at.

You have to be able to distinguish 40 species of dogs

and other kinds of things in 1,000 categories.

But I do think the progress we've made from about 26%

error in 2011 down to 3% in 2016 is hugely impactful.

Because the way I like to think about it is computers

have now evolved eyes that work, right?

And so we've now got the ability for computers

to perceive the world around them

in ways that didn't exist six or seven years ago.

And all of a sudden, that opens up applications of computing

that just didn't exist before.

Because now, you can depend on being able to see

and sense of what's right.

LAURENCE MORONEY: I know one of these applications that you're

always passionate about is diabetic retinopathy

and diagnosis of that.

Could you tell us what's going on in that space?

JEFF DEAN: Yeah, I mean, I think diabetic retinopathy

is a really good example of many medical imaging fields.

Where now, all of a sudden, if you collect

a high quality [INAUDIBLE] from domain experts,

radiologists labeling x-rays, or ophthalmologists

labeling eye images, and then you train a computer vision

model on that task, whatever it might be,

you can now sort of replicate the expertise of those domain

experts in a way that makes it possible to bring and deploy

that sort of expertise much more widely.

You can get something onto a GPU card

and do 100 images a second in a rural village

all over the world.

LAURENCE MORONEY: And I think that's the important part.

It's like places where there's shortage of that expertise,

you can now have impact to change the world.

JEFF DEAN: That's right.

Yeah, yes.

So you can offer--

if you have clinicians who are already

doing this task-- you can offer them an instant second opinion,

like a second colleague they can turn to.

But you can also deploy it in places where there are just

aren't enough doctors.

LAURENCE MORONEY: I just find that amazing,

and it's one of the ways that computer vision is now

more than just a number.

It's an application that we're able to change our world

to make it--

JEFF DEAN: I mean, being able to see

has all kinds of cool implications.

LAURENCE MORONEY: Exactly.

And then you also spoke a lot about language,

and some of the new language models,

and some of the research that's been going on into there.

And you can you update us a little on that?

JEFF DEAN: Sure.

I think in the last four or so years,

we've made a lot of progress as a community

in how do we build models that can basically

understand pieces of text?

Things like a paragraph or a couple of paragraphs

long, we can actually understand them at a much deeper level

than we were able to do before.

We still don't have a good handle on

how do we read an entire book and understand that in a way

a human would get from reading a book?

But understanding a few paragraphs of text

is actually a pretty fundamentally useful thing

for all kinds of things.

They can use these to improve our search system.

Just last week, we announced the use

of a BERT model, which is a fairly sophisticated

natural language processing model

in the middle of our search ranking algorithms.

And that's been shown to improve our search results quite

a lot for lots of different kinds of queries

that were previously pretty hard.

LAURENCE MORONEY: Cool, cool.

And I'm assuming can be used, for example, for like research

at least, for translation, for bringing more languages online

for [INAUDIBLE].

JEFF DEAN: Yeah, yeah.

So there's also a lot of advances

in the field of translation using these kinds of models.

Transformer-based models for translation

are showing remarkable gains in BLEU score

which is a measure of translation quality.

LAURENCE MORONEY: Right, right.

Now, one thing that I found particularly fascinating

that you were talking about as you were wrapping up

your keynote is that a lot of time,

we have these kind of atomic models

that do all these unit tasks.

But what about this great big model,

like to be able to do multiple things

and using neural architecture search to be

able to add to that model?

And could you elaborate a little bit

on that 'cause you had a great call to action there?

JEFF DEAN: Yeah, I think today, in the machine learning field,

we mostly find a problem we care about,

we find the right data to train a model

to do that particular task.

But we usually start from nothing with that model.

We basically initialize the parameters

of the model with random floating point numbers

and then try to learn everything about that task from the data

set we've collected.

And that seems pretty unrealistic.

It's sort of akin to, like, when you

want to learn to do something new,

you forget all your education, and you

go back to being an infant.

LAURENCE MORONEY: Take a brain out

and put a different brain in.

JEFF DEAN: And now, you try to learn everything

about this task.

And that's going to require that you

have a lot more examples of what it is you're trying to do,

because you're not generalizing from all the other things

you already know how to do.

And it's also going to mean you need

a lot more computation and a lot more

effort to achieve good outcomes in those tasks.

If, instead, you had a model that

knew how to do lots and lots of things,

in the limit, all the things we're

training separate machine learning models for,

why aren't we training one large model for this

with different pieces of expertise?

I think it's really important that, if we

have a large model, that we only sort of sparsely activate it.

We call upon different pieces of it as needed.

But mostly, 99% of the model is idle for any given task.

And you call upon the right pieces of expertise

when you need them.

That, I think, is a promising direction.

There's a lot of really interesting computer systems

problems underneath there.

How do we actually scale to a model of that size?

There's a lot of interesting machine

learning research questions.

How do we have a model that evolves its structure that

learns to route to different pieces of the model that

are most appropriate?

But I'm pretty excited about it.

LAURENCE MORONEY: Yeah, me, too.

And it's like, it's one of those things that might

seem a little fantastical now.

But only two or three or four years ago,

the computer vision and natural language stuff that

we're talking about seemed fantastical then, so it's--

JEFF DEAN: Right.

And we're seeing hints of things.

Like, neural architecture search seems to work well for things.

We're seeing the fact that when you

do transfer learning from another related task,

you generally get good results with less

data for the final task you care about.

Multi-task learning at small scales of five

or six related things all tend to make things work well.

So this is just sort of the logical consequence

of extending all those ideas out.

LAURENCE MORONEY: Yeah, exactly.

So then bringing you back, for example,

to the computer vision that we spoke about early on.

It was, like, who would have thought that when we were first

researching that, that things like diabetic retinopathy

would have been possible?

And now we're at the point where with this model, this--

I don't know what to call it-- model of everything,

uber model, that kind of thing, there

were going to be implications for that

can change the world, that can make the world a better place.

JEFF DEAN: Yeah.

That's what we hope.

LAURENCE MORONEY: That's the hope,

and that's also the driving goal, I think.

And that's one of the things that I find--

and if we go back to your keynote,

towards the end of your keynote, when you spoke about fairness,

when you spoke about the engineering challenges

that we're helping to solve, that

was personally inspiring to me.

JEFF DEAN: Hmm, cool.

LAURENCE MORONEY: And I hope it's personally

inspiring to you, too.

So thanks so much, Jeff.

I really appreciate having you on and--

JEFF DEAN: Thanks very much.

Appreciate it.

LAURENCE MORONEY: Thank you.

JEFF DEAN: Thanks.

[MUSIC PLAYING]