0:00:00.800,0:00:03.924
So, I lead a team at Google[br]that works on machine intelligence;

0:00:03.948,0:00:08.598
in other words, the engineering discipline[br]of making computers and devices

0:00:08.622,0:00:11.041
able to do some of the things[br]that brains do.

0:00:11.439,0:00:14.538
And this makes us[br]interested in real brains

0:00:14.562,0:00:15.851
and neuroscience as well,

0:00:15.875,0:00:20.047
and especially interested[br]in the things that our brains do

0:00:20.071,0:00:24.113
that are still far superior[br]to the performance of computers.

0:00:25.209,0:00:28.818
Historically, one of those areas[br]has been perception,

0:00:28.842,0:00:31.881
the process by which things[br]out there in the world --

0:00:31.905,0:00:33.489
sounds and images --

0:00:33.513,0:00:35.691
can turn into concepts in the mind.

0:00:36.235,0:00:38.752
This is essential for our own brains,

0:00:38.776,0:00:41.240
and it's also pretty useful on a computer.

0:00:41.636,0:00:44.986
The machine perception algorithms,[br]for example, that our team makes,

0:00:45.010,0:00:48.884
are what enable your pictures[br]on Google Photos to become searchable,

0:00:48.908,0:00:50.305
based on what's in them.

0:00:51.594,0:00:55.087
The flip side of perception is creativity:

0:00:55.111,0:00:58.149
turning a concept into something[br]out there into the world.

0:00:58.173,0:01:01.728
So over the past year,[br]our work on machine perception

0:01:01.752,0:01:06.611
has also unexpectedly connected[br]with the world of machine creativity

0:01:06.635,0:01:07.795
and machine art.

0:01:08.556,0:01:11.840
I think Michelangelo[br]had a penetrating insight

0:01:11.864,0:01:15.520
into to this dual relationship[br]between perception and creativity.

0:01:16.023,0:01:18.029
This is a famous quote of his:

0:01:18.053,0:01:21.376
"Every block of stone[br]has a statue inside of it,

0:01:22.036,0:01:25.038
and the job of the sculptor[br]is to discover it."

0:01:26.029,0:01:29.245
So I think that what[br]Michelangelo was getting at

0:01:29.269,0:01:32.449
is that we create by perceiving,

0:01:32.473,0:01:35.496
and that perception itself[br]is an act of imagination

0:01:35.520,0:01:37.981
and is the stuff of creativity.

0:01:38.691,0:01:42.616
The organ that does all the thinking[br]and perceiving and imagining,

0:01:42.640,0:01:44.228
of course, is the brain.

0:01:45.089,0:01:47.634
And I'd like to begin[br]with a brief bit of history

0:01:47.658,0:01:49.960
about what we know about brains.

0:01:50.496,0:01:52.942
Because unlike, say,[br]the heart or the intestines,

0:01:52.966,0:01:56.110
you really can't say very much[br]about a brain by just looking at it,

0:01:56.134,0:01:57.546
at least with the naked eye.

0:01:57.983,0:02:00.399
The early anatomists who looked at brains

0:02:00.423,0:02:04.230
gave the superficial structures[br]of this thing all kinds of fanciful names,

0:02:04.254,0:02:06.687
like hippocampus, meaning "little shrimp."

0:02:06.711,0:02:09.475
But of course that sort of thing[br]doesn't tell us very much

0:02:09.499,0:02:11.817
about what's actually going on inside.

0:02:12.780,0:02:16.393
The first person who, I think, really[br]developed some kind of insight

0:02:16.417,0:02:18.347
into what was going on in the brain

0:02:18.371,0:02:22.291
was the great Spanish neuroanatomist,[br]Santiago Ramón y Cajal,

0:02:22.315,0:02:23.859
in the 19th century,

0:02:23.883,0:02:27.638
who used microscopy and special stains

0:02:27.662,0:02:31.832
that could selectively fill in[br]or render in very high contrast

0:02:31.856,0:02:33.864
the individual cells in the brain,

0:02:33.888,0:02:37.042
in order to start to understand[br]their morphologies.

0:02:37.972,0:02:40.863
And these are the kinds of drawings[br]that he made of neurons

0:02:40.887,0:02:42.096
in the 19th century.

0:02:42.120,0:02:44.004
This is from a bird brain.

0:02:44.028,0:02:47.085
And you see this incredible variety[br]of different sorts of cells,

0:02:47.109,0:02:50.544
even the cellular theory itself[br]was quite new at this point.

0:02:50.568,0:02:51.846
And these structures,

0:02:51.870,0:02:54.129
these cells that have these arborizations,

0:02:54.153,0:02:56.761
these branches that can go[br]very, very long distances --

0:02:56.785,0:02:58.401
this was very novel at the time.

0:02:58.779,0:03:01.682
They're reminiscent, of course, of wires.

0:03:01.706,0:03:05.163
That might have been obvious[br]to some people in the 19th century;

0:03:05.187,0:03:09.501
the revolutions of wiring and electricity[br]were just getting underway.

0:03:09.964,0:03:11.142
But in many ways,

0:03:11.166,0:03:14.479
these microanatomical drawings[br]of Ramón y Cajal's, like this one,

0:03:14.503,0:03:16.835
they're still in some ways unsurpassed.

0:03:16.859,0:03:18.713
We're still more than a century later,

0:03:18.737,0:03:21.562
trying to finish the job[br]that Ramón y Cajal started.

0:03:21.586,0:03:24.720
These are raw data from our collaborators

0:03:24.744,0:03:27.625
at the Max Planck Institute[br]of Neuroscience.

0:03:27.649,0:03:29.439
And what our collaborators have done

0:03:29.463,0:03:34.464
is to image little pieces of brain tissue.

0:03:34.488,0:03:37.814
The entire sample here[br]is about one cubic millimeter in size,

0:03:37.838,0:03:40.459
and I'm showing you a very,[br]very small piece of it here.

0:03:40.483,0:03:42.829
That bar on the left is about one micron.

0:03:42.853,0:03:45.262
The structures you see are mitochondria

0:03:45.286,0:03:47.330
that are the size of bacteria.

0:03:47.354,0:03:48.905
And these are consecutive slices

0:03:48.929,0:03:52.077
through this very, very[br]tiny block of tissue.

0:03:52.101,0:03:54.504
Just for comparison's sake,

0:03:54.528,0:03:58.320
the diameter of an average strand[br]of hair is about 100 microns.

0:03:58.344,0:04:00.618
So we're looking at something[br]much, much smaller

0:04:00.642,0:04:02.040
than a single strand of hair.

0:04:02.064,0:04:06.095
And from these kinds of serial[br]electron microscopy slices,

0:04:06.119,0:04:11.127
one can start to make reconstructions[br]in 3D of neurons that look like these.

0:04:11.151,0:04:14.308
So these are sort of in the same[br]style as Ramón y Cajal.

0:04:14.332,0:04:15.824
Only a few neurons lit up,

0:04:15.848,0:04:18.629
because otherwise we wouldn't[br]be able to see anything here.

0:04:18.653,0:04:19.965
It would be so crowded,

0:04:19.989,0:04:21.319
so full of structure,

0:04:21.343,0:04:24.067
of wiring all connecting[br]one neuron to another.

0:04:25.293,0:04:28.097
So Ramón y Cajal was a little bit[br]ahead of his time,

0:04:28.121,0:04:30.676
and progress on understanding the brain

0:04:30.700,0:04:32.971
proceeded slowly[br]over the next few decades.

0:04:33.455,0:04:36.308
But we knew that neurons used electricity,

0:04:36.332,0:04:39.268
and by World War II, our technology[br]was advanced enough

0:04:39.292,0:04:42.098
to start doing real electrical[br]experiments on live neurons

0:04:42.122,0:04:44.228
to better understand how they worked.

0:04:44.631,0:04:48.987
This was the very same time[br]when computers were being invented,

0:04:49.011,0:04:52.111
very much based on the idea[br]of modeling the brain --

0:04:52.135,0:04:55.220
of "intelligent machinery,"[br]as Alan Turing called it,

0:04:55.244,0:04:57.235
one of the fathers of computer science.

0:04:57.923,0:05:02.555
Warren McCulloch and Walter Pitts[br]looked at Ramón y Cajal's drawing

0:05:02.579,0:05:03.896
of visual cortex,

0:05:03.920,0:05:05.482
which I'm showing here.

0:05:05.506,0:05:09.948
This is the cortex that processes[br]imagery that comes from the eye.

0:05:10.424,0:05:13.932
And for them, this looked[br]like a circuit diagram.

0:05:14.353,0:05:18.188
So there are a lot of details[br]in McCulloch and Pitts's circuit diagram

0:05:18.212,0:05:19.564
that are not quite right.

0:05:19.588,0:05:20.823
But this basic idea

0:05:20.847,0:05:24.839
that visual cortex works like a series[br]of computational elements

0:05:24.863,0:05:27.609
that pass information[br]one to the next in a cascade,

0:05:27.633,0:05:29.235
is essentially correct.

0:05:29.259,0:05:31.609
Let's talk for a moment

0:05:31.633,0:05:35.665
about what a model for processing[br]visual information would need to do.

0:05:36.228,0:05:38.969
The basic task of perception

0:05:38.993,0:05:43.187
is to take an image like this one and say,

0:05:43.211,0:05:44.387
"That's a bird,"

0:05:44.411,0:05:47.285
which is a very simple thing[br]for us to do with our brains.

0:05:47.309,0:05:50.730
But you should all understand[br]that for a computer,

0:05:50.754,0:05:53.841
this was pretty much impossible[br]just a few years ago.

0:05:53.865,0:05:55.781
The classical computing paradigm

0:05:55.805,0:05:58.312
is not one in which[br]this task is easy to do.

0:05:59.366,0:06:01.918
So what's going on between the pixels,

0:06:01.942,0:06:05.970
between the image of the bird[br]and the word "bird,"

0:06:05.994,0:06:08.808
is essentially a set of neurons[br]connected to each other

0:06:08.832,0:06:09.987
in a neural network,

0:06:10.011,0:06:11.234
as I'm diagramming here.

0:06:11.258,0:06:14.530
This neural network could be biological,[br]inside our visual cortices,

0:06:14.554,0:06:16.716
or, nowadays, we start[br]to have the capability

0:06:16.740,0:06:19.194
to model such neural networks[br]on the computer.

0:06:19.834,0:06:22.187
And I'll show you what[br]that actually looks like.

0:06:22.211,0:06:25.627
So the pixels you can think[br]about as a first layer of neurons,

0:06:25.651,0:06:27.890
and that's, in fact,[br]how it works in the eye --

0:06:27.914,0:06:29.577
that's the neurons in the retina.

0:06:29.601,0:06:31.101
And those feed forward

0:06:31.125,0:06:34.528
into one layer after another layer,[br]after another layer of neurons,

0:06:34.552,0:06:37.585
all connected by synapses[br]of different weights.

0:06:37.609,0:06:38.944
The behavior of this network

0:06:38.968,0:06:42.252
is characterized by the strengths[br]of all of those synapses.

0:06:42.276,0:06:45.564
Those characterize the computational[br]properties of this network.

0:06:45.588,0:06:47.058
And at the end of the day,

0:06:47.082,0:06:49.529
you have a neuron[br]or a small group of neurons

0:06:49.553,0:06:51.200
that light up, saying, "bird."

0:06:51.824,0:06:54.956
Now I'm going to represent[br]those three things --

0:06:54.980,0:06:59.676
the input pixels and the synapses[br]in the neural network,

0:06:59.700,0:07:01.285
and bird, the output --

0:07:01.309,0:07:04.366
by three variables: x, w and y.

0:07:04.853,0:07:06.664
There are maybe a million or so x's --

0:07:06.688,0:07:08.641
a million pixels in that image.

0:07:08.665,0:07:11.111
There are billions or trillions of w's,

0:07:11.135,0:07:14.556
which represent the weights of all[br]these synapses in the neural network.

0:07:14.580,0:07:16.455
And there's a very small number of y's,

0:07:16.479,0:07:18.337
of outputs that that network has.

0:07:18.361,0:07:20.110
"Bird" is only four letters, right?

0:07:21.088,0:07:24.514
So let's pretend that this[br]is just a simple formula,

0:07:24.538,0:07:26.701
x "x" w = y.

0:07:26.725,0:07:28.761
I'm putting the times in scare quotes

0:07:28.785,0:07:31.065
because what's really[br]going on there, of course,

0:07:31.089,0:07:34.135
is a very complicated series[br]of mathematical operations.

0:07:35.172,0:07:36.393
That's one equation.

0:07:36.417,0:07:38.089
There are three variables.

0:07:38.113,0:07:40.839
And we all know[br]that if you have one equation,

0:07:40.863,0:07:44.505
you can solve one variable[br]by knowing the other two things.

0:07:45.158,0:07:48.538
So the problem of inference,

0:07:48.562,0:07:51.435
that is, figuring out[br]that the picture of a bird is a bird,

0:07:51.459,0:07:52.733
is this one:

0:07:52.757,0:07:56.216
it's where y is the unknown[br]and w and x are known.

0:07:56.240,0:07:58.699
You know the neural network,[br]you know the pixels.

0:07:58.723,0:08:02.050
As you can see, that's actually[br]a relatively straightforward problem.

0:08:02.074,0:08:04.260
You multiply two times three[br]and you're done.

0:08:04.862,0:08:06.985
I'll show you an artificial neural network

0:08:07.009,0:08:09.305
that we've built recently,[br]doing exactly that.

0:08:09.634,0:08:12.494
This is running in real time[br]on a mobile phone,

0:08:12.518,0:08:15.831
and that's, of course,[br]amazing in its own right,

0:08:15.855,0:08:19.323
that mobile phones can do so many[br]billions and trillions of operations

0:08:19.347,0:08:20.595
per second.

0:08:20.619,0:08:22.234
What you're looking at is a phone

0:08:22.258,0:08:25.805
looking at one after another[br]picture of a bird,

0:08:25.829,0:08:28.544
and actually not only saying,[br]"Yes, it's a bird,"

0:08:28.568,0:08:31.979
but identifying the species of bird[br]with a network of this sort.

0:08:32.890,0:08:34.716
So in that picture,

0:08:34.740,0:08:38.542
the x and the w are known,[br]and the y is the unknown.

0:08:38.566,0:08:41.074
I'm glossing over the very[br]difficult part, of course,

0:08:41.098,0:08:44.959
which is how on earth[br]do we figure out the w,

0:08:44.983,0:08:47.170
the brain that can do such a thing?

0:08:47.194,0:08:49.028
How would we ever learn such a model?

0:08:49.418,0:08:52.651
So this process of learning,[br]of solving for w,

0:08:52.675,0:08:55.322
if we were doing this[br]with the simple equation

0:08:55.346,0:08:57.346
in which we think about these as numbers,

0:08:57.370,0:09:00.057
we know exactly how to do that: 6 = 2 x w,

0:09:00.081,0:09:03.393
well, we divide by two and we're done.

0:09:04.001,0:09:06.221
The problem is with this operator.

0:09:06.823,0:09:07.974
So, division --

0:09:07.998,0:09:11.119
we've used division because[br]it's the inverse to multiplication,

0:09:11.143,0:09:12.583
but as I've just said,

0:09:12.607,0:09:15.056
the multiplication is a bit of a lie here.

0:09:15.080,0:09:18.406
This is a very, very complicated,[br]very non-linear operation;

0:09:18.430,0:09:20.134
it has no inverse.

0:09:20.158,0:09:23.308
So we have to figure out a way[br]to solve the equation

0:09:23.332,0:09:25.356
without a division operator.

0:09:25.380,0:09:27.723
And the way to do that[br]is fairly straightforward.

0:09:27.747,0:09:30.418
You just say, let's play[br]a little algebra trick,

0:09:30.442,0:09:33.348
and move the six over[br]to the right-hand side of the equation.

0:09:33.372,0:09:35.198
Now, we're still using multiplication.

0:09:35.675,0:09:39.255
And that zero -- let's think[br]about it as an error.

0:09:39.279,0:09:41.794
In other words, if we've solved[br]for w the right way,

0:09:41.818,0:09:43.474
then the error will be zero.

0:09:43.498,0:09:45.436
And if we haven't gotten it quite right,

0:09:45.460,0:09:47.209
the error will be greater than zero.

0:09:47.233,0:09:50.599
So now we can just take guesses[br]to minimize the error,

0:09:50.623,0:09:53.310
and that's the sort of thing[br]computers are very good at.

0:09:53.334,0:09:54.927
So you've taken an initial guess:

0:09:54.951,0:09:56.107
what if w = 0?

0:09:56.131,0:09:57.371
Well, then the error is 6.

0:09:57.395,0:09:58.841
What if w = 1? The error is 4.

0:09:58.865,0:10:01.232
And then the computer can[br]sort of play Marco Polo,

0:10:01.256,0:10:03.623
and drive down the error close to zero.

0:10:03.647,0:10:07.021
As it does that, it's getting[br]successive approximations to w.

0:10:07.045,0:10:10.701
Typically, it never quite gets there,[br]but after about a dozen steps,

0:10:10.725,0:10:15.349
we're up to w = 2.999,[br]which is close enough.

0:10:16.302,0:10:18.116
And this is the learning process.

0:10:18.140,0:10:20.870
So remember that what's been going on here

0:10:20.894,0:10:25.272
is that we've been taking[br]a lot of known x's and known y's

0:10:25.296,0:10:28.750
and solving for the w in the middle[br]through an iterative process.

0:10:28.774,0:10:32.330
It's exactly the same way[br]that we do our own learning.

0:10:32.354,0:10:34.584
We have many, many images as babies

0:10:34.608,0:10:37.241
and we get told, "This is a bird;[br]this is not a bird."

0:10:37.714,0:10:39.812
And over time, through iteration,

0:10:39.836,0:10:42.764
we solve for w, we solve[br]for those neural connections.

0:10:43.460,0:10:47.546
So now, we've held[br]x and w fixed to solve for y;

0:10:47.570,0:10:49.417
that's everyday, fast perception.

0:10:49.441,0:10:51.204
We figure out how we can solve for w,

0:10:51.228,0:10:53.131
that's learning, which is a lot harder,

0:10:53.155,0:10:55.140
because we need to do error minimization,

0:10:55.164,0:10:56.851
using a lot of training examples.

0:10:56.875,0:11:00.062
And about a year ago,[br]Alex Mordvintsev, on our team,

0:11:00.086,0:11:03.636
decided to experiment[br]with what happens if we try solving for x,

0:11:03.660,0:11:05.697
given a known w and a known y.

0:11:06.124,0:11:07.275
In other words,

0:11:07.299,0:11:08.651
you know that it's a bird,

0:11:08.675,0:11:11.978
and you already have your neural network[br]that you've trained on birds,

0:11:12.002,0:11:14.346
but what is the picture of a bird?

0:11:15.034,0:11:20.058
It turns out that by using exactly[br]the same error-minimization procedure,

0:11:20.082,0:11:23.512
one can do that with the network[br]trained to recognize birds,

0:11:23.536,0:11:26.924
and the result turns out to be ...

0:11:30.400,0:11:31.705
a picture of birds.

0:11:32.814,0:11:36.551
So this is a picture of birds[br]generated entirely by a neural network

0:11:36.575,0:11:38.401
that was trained to recognize birds,

0:11:38.425,0:11:41.963
just by solving for x[br]rather than solving for y,

0:11:41.987,0:11:43.275
and doing that iteratively.

0:11:43.732,0:11:45.579
Here's another fun example.

0:11:45.603,0:11:49.040
This was a work made[br]by Mike Tyka in our group,

0:11:49.064,0:11:51.372
which he calls "Animal Parade."

0:11:51.396,0:11:54.272
It reminds me a little bit[br]of William Kentridge's artworks,

0:11:54.296,0:11:56.785
in which he makes sketches, rubs them out,

0:11:56.809,0:11:58.269
makes sketches, rubs them out,

0:11:58.293,0:11:59.691
and creates a movie this way.

0:11:59.715,0:12:00.866
In this case,

0:12:00.890,0:12:04.167
what Mike is doing is varying y[br]over the space of different animals,

0:12:04.191,0:12:06.573
in a network designed[br]to recognize and distinguish

0:12:06.597,0:12:08.407
different animals from each other.

0:12:08.431,0:12:12.182
And you get this strange, Escher-like[br]morph from one animal to another.

0:12:14.221,0:12:18.835
Here he and Alex together[br]have tried reducing

0:12:18.859,0:12:21.618
the y's to a space of only two dimensions,

0:12:21.642,0:12:25.080
thereby making a map[br]out of the space of all things

0:12:25.104,0:12:26.823
recognized by this network.

0:12:26.847,0:12:28.870
Doing this kind of synthesis

0:12:28.894,0:12:31.276
or generation of imagery[br]over that entire surface,

0:12:31.300,0:12:34.146
varying y over the surface,[br]you make a kind of map --

0:12:34.170,0:12:37.311
a visual map of all the things[br]the network knows how to recognize.

0:12:37.335,0:12:40.200
The animals are all here;[br]"armadillo" is right in that spot.

0:12:40.919,0:12:43.398
You can do this with other kinds[br]of networks as well.

0:12:43.422,0:12:46.296
This is a network designed[br]to recognize faces,

0:12:46.320,0:12:48.320
to distinguish one face from another.

0:12:48.344,0:12:51.593
And here, we're putting[br]in a y that says, "me,"

0:12:51.617,0:12:53.192
my own face parameters.

0:12:53.216,0:12:54.922
And when this thing solves for x,

0:12:54.946,0:12:57.564
it generates this rather crazy,

0:12:57.588,0:13:02.016
kind of cubist, surreal,[br]psychedelic picture of me

0:13:02.040,0:13:03.846
from multiple points of view at once.

0:13:03.870,0:13:06.604
The reason it looks like[br]multiple points of view at once

0:13:06.628,0:13:10.315
is because that network is designed[br]to get rid of the ambiguity

0:13:10.339,0:13:12.815
of a face being in one pose[br]or another pose,

0:13:12.839,0:13:16.215
being looked at with one kind of lighting,[br]another kind of lighting.

0:13:16.239,0:13:18.324
So when you do[br]this sort of reconstruction,

0:13:18.348,0:13:20.652
if you don't use some sort of guide image

0:13:20.676,0:13:21.887
or guide statistics,

0:13:21.911,0:13:25.676
then you'll get a sort of confusion[br]of different points of view,

0:13:25.700,0:13:27.068
because it's ambiguous.

0:13:27.786,0:13:32.009
This is what happens if Alex uses[br]his own face as a guide image

0:13:32.033,0:13:35.354
during that optimization process[br]to reconstruct my own face.

0:13:36.284,0:13:38.612
So you can see it's not perfect.

0:13:38.636,0:13:40.510
There's still quite a lot of work to do

0:13:40.534,0:13:42.987
on how we optimize[br]that optimization process.

0:13:43.011,0:13:45.838
But you start to get something[br]more like a coherent face,

0:13:45.862,0:13:47.876
rendered using my own face as a guide.

0:13:48.892,0:13:51.393
You don't have to start[br]with a blank canvas

0:13:51.417,0:13:52.573
or with white noise.

0:13:52.597,0:13:53.901
When you're solving for x,

0:13:53.925,0:13:57.814
you can begin with an x,[br]that is itself already some other image.

0:13:57.838,0:14:00.394
That's what this little demonstration is.

0:14:00.418,0:14:04.540
This is a network[br]that is designed to categorize

0:14:04.564,0:14:07.683
all sorts of different objects --[br]man-made structures, animals ...

0:14:07.707,0:14:10.300
Here we're starting[br]with just a picture of clouds,

0:14:10.324,0:14:11.995
and as we optimize,

0:14:12.019,0:14:16.505
basically, this network is figuring out[br]what it sees in the clouds.

0:14:16.931,0:14:19.251
And the more time[br]you spend looking at this,

0:14:19.275,0:14:22.028
the more things you also[br]will see in the clouds.

0:14:23.004,0:14:26.379
You could also use the face network[br]to hallucinate into this,

0:14:26.403,0:14:28.215
and you get some pretty crazy stuff.

0:14:28.239,0:14:29.389
(Laughter)

0:14:30.401,0:14:33.145
Or, Mike has done some other experiments

0:14:33.169,0:14:37.074
in which he takes that cloud image,

0:14:37.098,0:14:40.605
hallucinates, zooms, hallucinates,[br]zooms hallucinates, zooms.

0:14:40.629,0:14:41.780
And in this way,

0:14:41.804,0:14:45.479
you can get a sort of fugue state[br]of the network, I suppose,

0:14:45.503,0:14:49.183
or a sort of free association,

0:14:49.207,0:14:51.434
in which the network[br]is eating its own tail.

0:14:51.458,0:14:54.879
So every image is now the basis for,

0:14:54.903,0:14:56.324
"What do I think I see next?

0:14:56.348,0:14:59.151
What do I think I see next?[br]What do I think I see next?"

0:14:59.487,0:15:02.423
I showed this for the first time in public

0:15:02.447,0:15:07.884
to a group at a lecture in Seattle[br]called "Higher Education" --

0:15:07.908,0:15:10.345
this was right after[br]marijuana was legalized.

0:15:10.369,0:15:12.784
(Laughter)

0:15:14.627,0:15:16.731
So I'd like to finish up quickly

0:15:16.755,0:15:21.010
by just noting that this technology[br]is not constrained.

0:15:21.034,0:15:24.699
I've shown you purely visual examples[br]because they're really fun to look at.

0:15:24.723,0:15:27.174
It's not a purely visual technology.

0:15:27.198,0:15:29.191
Our artist collaborator, Ross Goodwin,

0:15:29.215,0:15:32.886
has done experiments involving[br]a camera that takes a picture,

0:15:32.910,0:15:37.144
and then a computer in his backpack[br]writes a poem using neural networks,

0:15:37.168,0:15:39.112
based on the contents of the image.

0:15:39.136,0:15:42.083
And that poetry neural network[br]has been trained

0:15:42.107,0:15:44.341
on a large corpus of 20th-century poetry.

0:15:44.365,0:15:45.864
And the poetry is, you know,

0:15:45.888,0:15:47.802
I think, kind of not bad, actually.

0:15:47.826,0:15:49.210
(Laughter)

0:15:49.234,0:15:50.393
In closing,

0:15:50.417,0:15:52.549
I think that per Michelangelo,

0:15:52.573,0:15:53.807
I think he was right;

0:15:53.831,0:15:57.267
perception and creativity[br]are very intimately connected.

0:15:57.611,0:16:00.245
What we've just seen are neural networks

0:16:00.269,0:16:02.572
that are entirely trained to discriminate,

0:16:02.596,0:16:04.838
or to recognize different[br]things in the world,

0:16:04.862,0:16:08.023
able to be run in reverse, to generate.

0:16:08.047,0:16:09.830
One of the things that suggests to me

0:16:09.854,0:16:12.252
is not only that[br]Michelangelo really did see

0:16:12.276,0:16:14.728
the sculpture in the blocks of stone,

0:16:14.752,0:16:18.390
but that any creature,[br]any being, any alien

0:16:18.414,0:16:22.071
that is able to do[br]perceptual acts of that sort

0:16:22.095,0:16:23.470
is also able to create

0:16:23.494,0:16:26.718
because it's exactly the same[br]machinery that's used in both cases.

0:16:26.742,0:16:31.274
Also, I think that perception[br]and creativity are by no means

0:16:31.298,0:16:32.508
uniquely human.

0:16:32.532,0:16:36.240
We start to have computer models[br]that can do exactly these sorts of things.

0:16:36.264,0:16:39.592
And that ought to be unsurprising;[br]the brain is computational.

0:16:39.616,0:16:41.273
And finally,

0:16:41.297,0:16:45.965
computing began as an exercise[br]in designing intelligent machinery.

0:16:45.989,0:16:48.451
It was very much modeled after the idea

0:16:48.475,0:16:51.488
of how could we make machines intelligent.

0:16:51.512,0:16:53.674
And we finally are starting to fulfill now

0:16:53.698,0:16:56.104
some of the promises[br]of those early pioneers,

0:16:56.128,0:16:57.841
of Turing and von Neumann

0:16:57.865,0:17:00.130
and McCulloch and Pitts.

0:17:00.154,0:17:04.252
And I think that computing[br]is not just about accounting

0:17:04.276,0:17:06.423
or playing Candy Crush or something.

0:17:06.447,0:17:09.025
From the beginning,[br]we modeled them after our minds.

0:17:09.049,0:17:12.318
And they give us both the ability[br]to understand our own minds better

0:17:12.342,0:17:13.871
and to extend them.

0:17:14.627,0:17:15.794
Thank you very much.

0:17:15.818,0:17:21.757
(Applause)