TED@BCG Paris
Blaise Agüera y Arcas: How computers are learning to be creative
布莱斯 · 阿尔卡斯: 计算机如何学习具有创造力
Filmed:
Readability: 4.4
1,934,067 views
我们站在了艺术和创造力的前沿——而做到这一点的并不是人类。谷歌首席科学家，布莱斯 · 阿尔卡斯，致力于研究机器的感知和分布式学习的深度神经网络。在这个视屏中，他将会展示可以识别图片的神经网络如何反向运行。结果显示：壮观，梦幻的拼贴画（还有诗歌！）很难被归类。布莱斯 · 阿尔卡斯说：“感知和创造力联系的特别紧密。任何生物，任何可以感知行动的事物，都可以进行创造。”
Blaise Agüera y Arcas - Software architect
Blaise Agüera y Arcas works on machine learning at Google. Previously a Distinguished Engineer at Microsoft, he has worked on augmented reality, mapping, wearable computing and natural user interfaces. Full bio
Blaise Agüera y Arcas works on machine learning at Google. Previously a Distinguished Engineer at Microsoft, he has worked on augmented reality, mapping, wearable computing and natural user interfaces. Full bio
Double-click the English transcript below to play the video.
00:12
So, I lead a team at Google
that works on machine intelligence;
that works on machine intelligence;
0
800
3124
我在谷歌领导着一个
机器智能的项目组，
机器智能的项目组，
00:15
in other words, the engineering discipline
of making computers and devices
of making computers and devices
1
3948
4650
换句话说，利用工程学原理制造出
00:20
able to do some of the things
that brains do.
that brains do.
2
8622
2419
能够像人脑一样
完成某些任务的电脑和设备。
完成某些任务的电脑和设备。
00:23
And this makes us
interested in real brains
interested in real brains
3
11439
3099
这也使我们对人类的
大脑以及神经科学
大脑以及神经科学
00:26
and neuroscience as well,
4
14562
1289
产生了兴趣，
00:27
and especially interested
in the things that our brains do
in the things that our brains do
5
15875
4172
尤其在那些大脑的表现
00:32
that are still far superior
to the performance of computers.
to the performance of computers.
6
20071
4042
比电脑强太多的领域。
00:37
Historically, one of those areas
has been perception,
has been perception,
7
25209
3609
长期以来，我们研究的
其中一个领域便是感知，
其中一个领域便是感知，
00:40
the process by which things
out there in the world --
out there in the world --
8
28842
3039
一种将外界事物——
比如图像或声音—
00:43
sounds and images --
9
31905
1584
00:45
can turn into concepts in the mind.
10
33513
2178
转化为大脑内概念的过程。
00:48
This is essential for our own brains,
11
36235
2517
这对我们的大脑很重要，
00:50
and it's also pretty useful on a computer.
12
38776
2464
对计算机的作用也非同小可。
00:53
The machine perception algorithms,
for example, that our team makes,
for example, that our team makes,
13
41636
3350
例如，我们团队开发的机器感知算法
会根据图片的内容
让你在谷歌相册的图片
让你在谷歌相册的图片
00:57
are what enable your pictures
on Google Photos to become searchable,
on Google Photos to become searchable,
14
45010
3874
01:00
based on what's in them.
15
48908
1397
出现在搜索结果中。
01:03
The flip side of perception is creativity:
16
51594
3493
感知的另一方面是创意：
01:07
turning a concept into something
out there into the world.
out there into the world.
17
55111
3038
将概念变成现实。
01:10
So over the past year,
our work on machine perception
our work on machine perception
18
58173
3555
因此，这些年我们
在机器感知能力方面的工作
在机器感知能力方面的工作
01:13
has also unexpectedly connected
with the world of machine creativity
with the world of machine creativity
19
61752
4859
也意外地跟机器创意以及机器艺术
01:18
and machine art.
20
66635
1160
联系在了一起。
01:20
I think Michelangelo
had a penetrating insight
had a penetrating insight
21
68556
3284
我觉得米开朗基罗对感知和创意
01:23
into to this dual relationship
between perception and creativity.
between perception and creativity.
22
71864
3656
之间的双重关系有着深刻的见解。
01:28
This is a famous quote of his:
23
76023
2006
他有一句名言：
01:30
"Every block of stone
has a statue inside of it,
has a statue inside of it,
24
78053
3323
“每一块石头里都藏着一尊雕像，
01:34
and the job of the sculptor
is to discover it."
is to discover it."
25
82036
3002
而雕塑家的工作就是去发现它。”
01:38
So I think that what
Michelangelo was getting at
Michelangelo was getting at
26
86029
3216
我想米开朗基罗意思是
01:41
is that we create by perceiving,
27
89269
3180
我们通过感知来创造，
01:44
and that perception itself
is an act of imagination
is an act of imagination
28
92473
3023
而感知本身是想象力的表现，
01:47
and is the stuff of creativity.
29
95520
2461
以及创意的来源。
01:50
The organ that does all the thinking
and perceiving and imagining,
and perceiving and imagining,
30
98691
3925
而进行思考、感知和想象的器官，
01:54
of course, is the brain.
31
102640
1588
毫无疑问，就是大脑。
01:57
And I'd like to begin
with a brief bit of history
with a brief bit of history
32
105089
2545
我想先简单地谈一谈
01:59
about what we know about brains.
33
107658
2302
我们对大脑的了解。
02:02
Because unlike, say,
the heart or the intestines,
the heart or the intestines,
34
110496
2446
因为不像心脏或其它内脏，
02:04
you really can't say very much
about a brain by just looking at it,
about a brain by just looking at it,
35
112966
3144
你无法仅仅通过观察
就能看出点什么来，
就能看出点什么来，
02:08
at least with the naked eye.
36
116134
1412
至少仅凭肉眼看不出来。
02:09
The early anatomists who looked at brains
37
117983
2416
早期的解剖学家看着大脑，
02:12
gave the superficial structures
of this thing all kinds of fanciful names,
of this thing all kinds of fanciful names,
38
120423
3807
给它的表面结构
取了各种充满想象力的名字。
取了各种充满想象力的名字。
02:16
like hippocampus, meaning "little shrimp."
39
124254
2433
比如说海马体，意思是“小虾子”。
02:18
But of course that sort of thing
doesn't tell us very much
doesn't tell us very much
40
126711
2764
但这些并不能告诉我们
大脑里面究竟是怎样工作的。
02:21
about what's actually going on inside.
41
129499
2318
02:24
The first person who, I think, really
developed some kind of insight
developed some kind of insight
42
132780
3613
我认为第一个真正对大脑的工作方式
02:28
into what was going on in the brain
43
136417
1930
有所洞悉的人，
02:30
was the great Spanish neuroanatomist,
Santiago Ramón y Cajal,
Santiago Ramón y Cajal,
44
138371
3920
是19世纪西班牙
伟大的神经解剖学家
伟大的神经解剖学家
02:34
in the 19th century,
45
142315
1544
圣地亚哥 · 拉蒙 · 卡哈尔
（Santiago Ramón y Cajal），
（Santiago Ramón y Cajal），
他使用了显微镜以及某种特殊染色剂，
02:35
who used microscopy and special stains
46
143883
3755
02:39
that could selectively fill in
or render in very high contrast
or render in very high contrast
47
147662
4170
有选择性地将大脑中的
单个细胞填充或者渲染上
单个细胞填充或者渲染上
02:43
the individual cells in the brain,
48
151856
2008
高对比度的颜色，
以便了解它们的形态。
02:45
in order to start to understand
their morphologies.
their morphologies.
49
153888
3154
02:49
And these are the kinds of drawings
that he made of neurons
that he made of neurons
50
157972
2891
这些就是他在19世纪
完成的的神经元手绘图。
02:52
in the 19th century.
51
160887
1209
这是一只鸟的大脑。
02:54
This is from a bird brain.
52
162120
1884
02:56
And you see this incredible variety
of different sorts of cells,
of different sorts of cells,
53
164028
3057
能看到这些形态各异的细胞，
甚至在当时对细胞学说
本身还是新鲜事物。
本身还是新鲜事物。
02:59
even the cellular theory itself
was quite new at this point.
was quite new at this point.
54
167109
3435
而这些结构，
03:02
And these structures,
55
170568
1278
像树枝一样分岔，
03:03
these cells that have these arborizations,
56
171870
2259
03:06
these branches that can go
very, very long distances --
very, very long distances --
57
174153
2608
能够延伸到很长的距离——
这些在当时都是闻所未闻。
03:08
this was very novel at the time.
58
176785
1616
03:10
They're reminiscent, of course, of wires.
59
178779
2903
他们让人联想到的，当然是电线。
03:13
That might have been obvious
to some people in the 19th century;
to some people in the 19th century;
60
181706
3457
这对于很多19世纪的人
来说是显而易见的，
来说是显而易见的，
03:17
the revolutions of wiring and electricity
were just getting underway.
were just getting underway.
61
185187
4314
因为那时电线和电力革命刚刚兴起。
03:21
But in many ways,
62
189964
1178
但是在许多方面
03:23
these microanatomical drawings
of Ramón y Cajal's, like this one,
of Ramón y Cajal's, like this one,
63
191166
3313
拉蒙 · 卡哈尔的神经解剖学
绘画，比如这一张，
绘画，比如这一张，
从某些方面来说是很卓越的。
03:26
they're still in some ways unsurpassed.
64
194503
2332
03:28
We're still more than a century later,
65
196859
1854
一个多世纪后的我们，仍然在继续
03:30
trying to finish the job
that Ramón y Cajal started.
that Ramón y Cajal started.
66
198737
2825
尝试完成拉蒙 · 卡哈尔开启的事业。
03:33
These are raw data from our collaborators
67
201586
3134
提供这些原始数据的，是我们来自
03:36
at the Max Planck Institute
of Neuroscience.
of Neuroscience.
68
204744
2881
马克斯 · 普朗克
神经科学研究所的合作者。
神经科学研究所的合作者。
03:39
And what our collaborators have done
69
207649
1790
他们的工作
是对那些小块的脑组织进行成像。
03:41
is to image little pieces of brain tissue.
70
209463
5001
03:46
The entire sample here
is about one cubic millimeter in size,
is about one cubic millimeter in size,
71
214488
3326
这一整个样品的大小
是1立方毫米左右，
是1立方毫米左右，
03:49
and I'm showing you a very,
very small piece of it here.
very small piece of it here.
72
217838
2621
而我展示的只是它上面
很小很小的一块区域。
很小很小的一块区域。
03:52
That bar on the left is about one micron.
73
220483
2346
左边那段比例尺的长度是1微米。
03:54
The structures you see are mitochondria
74
222853
2409
你看到的这个结构
是一个细菌大小的线粒体。
03:57
that are the size of bacteria.
75
225286
2044
03:59
And these are consecutive slices
76
227354
1551
这些是利用这个非常微小的组织
04:00
through this very, very
tiny block of tissue.
tiny block of tissue.
77
228929
3148
所制作成的连续的切片。
04:04
Just for comparison's sake,
78
232101
2403
我们来做个对比。
04:06
the diameter of an average strand
of hair is about 100 microns.
of hair is about 100 microns.
79
234528
3792
通常一根头发的直径是
100微米左右。
100微米左右。
04:10
So we're looking at something
much, much smaller
much, much smaller
80
238344
2274
所以我们看到的东西
比一根头发丝还要细很多。
04:12
than a single strand of hair.
81
240642
1398
通过这些连续的电子显微镜切片，
04:14
And from these kinds of serial
electron microscopy slices,
electron microscopy slices,
82
242064
4031
人们可以重构出类似这样的
神经元三维图像。
神经元三维图像。
04:18
one can start to make reconstructions
in 3D of neurons that look like these.
in 3D of neurons that look like these.
83
246119
5008
04:23
So these are sort of in the same
style as Ramón y Cajal.
style as Ramón y Cajal.
84
251151
3157
某种程度上，这跟拉蒙 · 卡哈尔
所用的方式是一样的。
所用的方式是一样的。
04:26
Only a few neurons lit up,
85
254332
1492
我们只对少量的神经元进行了突出显示，
04:27
because otherwise we wouldn't
be able to see anything here.
be able to see anything here.
86
255848
2781
否则我们不可能看到任何东西，
因为那样一来画面会很拥挤，
04:30
It would be so crowded,
87
258653
1312
04:31
so full of structure,
88
259989
1330
充满了组织结构，
04:33
of wiring all connecting
one neuron to another.
one neuron to another.
89
261343
2724
充满了各个神经元间
纵横交错的通路。
纵横交错的通路。
04:37
So Ramón y Cajal was a little bit
ahead of his time,
ahead of his time,
90
265293
2804
显然，拉蒙 · 卡哈尔
有一点超前于他的时代，
有一点超前于他的时代，
04:40
and progress on understanding the brain
91
268121
2555
接下来的几十年间
04:42
proceeded slowly
over the next few decades.
over the next few decades.
92
270700
2271
人们对大脑的理解进展非常缓慢。
04:45
But we knew that neurons used electricity,
93
273455
2853
但是我们已经知道，
神经元通过电流传导信息，
神经元通过电流传导信息，
而到二战时，我们的技术
已取得了长足的进步，
已取得了长足的进步，
04:48
and by World War II, our technology
was advanced enough
was advanced enough
94
276332
2936
04:51
to start doing real electrical
experiments on live neurons
experiments on live neurons
95
279292
2806
可以开始在活的
神经元细胞上做电流实验，
神经元细胞上做电流实验，
04:54
to better understand how they worked.
96
282122
2106
以便更好地理解它们的工作原理。
04:56
This was the very same time
when computers were being invented,
when computers were being invented,
97
284631
4356
而电脑也正是在
这个时候被发明了出来，
这个时候被发明了出来，
05:01
very much based on the idea
of modeling the brain --
of modeling the brain --
98
289011
3100
它的发明是基于对大脑的模拟——
05:04
of "intelligent machinery,"
as Alan Turing called it,
as Alan Turing called it,
99
292135
3085
也就是阿兰 · 图灵
所称的“智能机器”理念，
所称的“智能机器”理念，
05:07
one of the fathers of computer science.
100
295244
1991
图灵是计算机科学的开创者之一。
05:09
Warren McCulloch and Walter Pitts
looked at Ramón y Cajal's drawing
looked at Ramón y Cajal's drawing
101
297923
4632
沃伦 · 麦卡洛克（Warren McCulloch）和
沃尔特 · 皮兹（Walter Pitts）看到了
沃尔特 · 皮兹（Walter Pitts）看到了
拉蒙 · 卡哈尔所画的
大脑视觉皮层，
大脑视觉皮层，
05:14
of visual cortex,
102
302579
1317
05:15
which I'm showing here.
103
303920
1562
就是我给你们看的这个。
05:17
This is the cortex that processes
imagery that comes from the eye.
imagery that comes from the eye.
104
305506
4442
这是负责处理我们视觉信息的大脑皮层。
05:22
And for them, this looked
like a circuit diagram.
like a circuit diagram.
105
310424
3508
对他们来说，这看起来像一个电路图。
05:26
So there are a lot of details
in McCulloch and Pitts's circuit diagram
in McCulloch and Pitts's circuit diagram
106
314353
3835
在麦卡洛克和皮兹的电路图上，
05:30
that are not quite right.
107
318212
1352
有许多细节并不是那么正确。
但基本概念是对的，
05:31
But this basic idea
108
319588
1235
05:32
that visual cortex works like a series
of computational elements
of computational elements
109
320847
3992
他们认为视觉皮层工作起来
就像一系列计算机元件
就像一系列计算机元件
05:36
that pass information
one to the next in a cascade,
one to the next in a cascade,
110
324863
2746
在同一个层级中传递信息，
05:39
is essentially correct.
111
327633
1602
这一点是对的。
05:41
Let's talk for a moment
112
329259
2350
我们再聊一聊
05:43
about what a model for processing
visual information would need to do.
visual information would need to do.
113
331633
4032
视觉信息处理模型需要做些什么。
05:48
The basic task of perception
114
336228
2741
感知的基本任务就是
05:50
is to take an image like this one and say,
115
338993
4194
抓取这样的图像并且告诉我们
“这是一只鸟”，
05:55
"That's a bird,"
116
343211
1176
这对我们的大脑来说非常简单。
05:56
which is a very simple thing
for us to do with our brains.
for us to do with our brains.
117
344411
2874
05:59
But you should all understand
that for a computer,
that for a computer,
118
347309
3421
但对一台电脑来说，
06:02
this was pretty much impossible
just a few years ago.
just a few years ago.
119
350754
3087
在几年前，这还是完全不可能的事。
06:05
The classical computing paradigm
120
353865
1916
传统的计算模式
06:07
is not one in which
this task is easy to do.
this task is easy to do.
121
355805
2507
很难完成这个任务。
06:11
So what's going on between the pixels,
122
359366
2552
像素、鸟的图像以及“鸟”这个词，
06:13
between the image of the bird
and the word "bird,"
and the word "bird,"
123
361942
4028
这三者之间所产生的联系，
本质上是在一个神经网络中各神经元
06:17
is essentially a set of neurons
connected to each other
connected to each other
124
365994
2814
相互连接的结果，
06:20
in a neural network,
125
368832
1155
正如这张图所示。
06:22
as I'm diagramming here.
126
370011
1223
06:23
This neural network could be biological,
inside our visual cortices,
inside our visual cortices,
127
371258
3272
这种神经网络可能是生物学上的，
存在于我们大脑视觉皮层里，
存在于我们大脑视觉皮层里，
06:26
or, nowadays, we start
to have the capability
to have the capability
128
374554
2162
或者，现如今我们开始有能力
06:28
to model such neural networks
on the computer.
on the computer.
129
376740
2454
在电脑上模拟这种神经网络。
06:31
And I'll show you what
that actually looks like.
that actually looks like.
130
379834
2353
我们来看一下它的工作原理。
06:34
So the pixels you can think
about as a first layer of neurons,
about as a first layer of neurons,
131
382211
3416
可以将像素想像成第一层的神经元，
06:37
and that's, in fact,
how it works in the eye --
how it works in the eye --
132
385651
2239
这实际上就是在
眼睛内部的工作原理——
眼睛内部的工作原理——
是视网膜上的神经元。
06:39
that's the neurons in the retina.
133
387914
1663
06:41
And those feed forward
134
389601
1500
然后这些前馈信息
06:43
into one layer after another layer,
after another layer of neurons,
after another layer of neurons,
135
391125
3403
通过一层层神经元往下传递，
这些神经元通过突触彼此连接。
06:46
all connected by synapses
of different weights.
of different weights.
136
394552
3033
06:49
The behavior of this network
137
397609
1335
这个神经网络的行为
06:50
is characterized by the strengths
of all of those synapses.
of all of those synapses.
138
398968
3284
是通过所有这些突触的强度来表达的，
06:54
Those characterize the computational
properties of this network.
properties of this network.
139
402276
3288
也塑造了这个网络的计算性能。
06:57
And at the end of the day,
140
405588
1470
最终，
一个或者一小群神经元
06:59
you have a neuron
or a small group of neurons
or a small group of neurons
141
407082
2447
07:01
that light up, saying, "bird."
142
409553
1647
会亮起来，说，“鸟”。
07:03
Now I'm going to represent
those three things --
those three things --
143
411824
3132
接下来我会将这三部分——
07:06
the input pixels and the synapses
in the neural network,
in the neural network,
144
414980
4696
输入的像素，神经网络中的突触，
07:11
and bird, the output --
145
419700
1585
以及“鸟”，这个输出结果——
07:13
by three variables: x, w and y.
146
421309
3057
用三个变量来表示：x、w和y。
07:16
There are maybe a million or so x's --
147
424853
1811
在那张图片上可能会有一百万个x——
07:18
a million pixels in that image.
148
426688
1953
代表一百万个像素点。
然后有几十亿或几万亿的w，
07:20
There are billions or trillions of w's,
149
428665
2446
07:23
which represent the weights of all
these synapses in the neural network.
these synapses in the neural network.
150
431135
3421
代表着神经网络中所有突触的权重。
只有很少数量的y，
07:26
And there's a very small number of y's,
151
434580
1875
07:28
of outputs that that network has.
152
436479
1858
代表整个网络的输出结果。
07:30
"Bird" is only four letters, right?
153
438361
1749
“Bird（鸟）"这个单词
只有四个字母，对吧？
只有四个字母，对吧？
07:33
So let's pretend that this
is just a simple formula,
is just a simple formula,
154
441088
3426
我们假定这只是一个很简单的公式
07:36
x "x" w = y.
155
444538
2163
x 乘以 w 等于 y。
我把乘号打上了引号，
07:38
I'm putting the times in scare quotes
156
446725
2036
因为实际的过程要复杂得多。
07:40
because what's really
going on there, of course,
going on there, of course,
157
448785
2280
07:43
is a very complicated series
of mathematical operations.
of mathematical operations.
158
451089
3046
牵涉到一系列非常复杂的数学运算。
07:47
That's one equation.
159
455172
1221
这是一个方程式，
07:48
There are three variables.
160
456417
1672
有三个变量。
07:50
And we all know
that if you have one equation,
that if you have one equation,
161
458113
2726
而我们知道在一个方程式中
07:52
you can solve one variable
by knowing the other two things.
by knowing the other two things.
162
460863
3642
通过两个已知数
你就能算出另一个未知数。
你就能算出另一个未知数。
07:57
So the problem of inference,
163
465158
3380
所以这道推论题，
08:00
that is, figuring out
that the picture of a bird is a bird,
that the picture of a bird is a bird,
164
468562
2873
即判断出图中是一只鸟，
08:03
is this one:
165
471459
1274
可以这样来描述：
08:04
it's where y is the unknown
and w and x are known.
and w and x are known.
166
472757
3459
y是未知数，w跟x都是已知数。
08:08
You know the neural network,
you know the pixels.
you know the pixels.
167
476240
2459
也就是神经网络和像素是已知的。
08:10
As you can see, that's actually
a relatively straightforward problem.
a relatively straightforward problem.
168
478723
3327
实际上这是一个相当简单的问题。
你只需要用2乘以3，就完事儿了。
08:14
You multiply two times three
and you're done.
and you're done.
169
482074
2186
08:16
I'll show you an artificial neural network
170
484862
2123
我会给你们展示我们最近
完成的人工神经网络，
完成的人工神经网络，
08:19
that we've built recently,
doing exactly that.
doing exactly that.
171
487009
2296
它的工作原理正是如此。
08:21
This is running in real time
on a mobile phone,
on a mobile phone,
172
489634
2860
这是在一台在手机上
实时运行的神经网络，
实时运行的神经网络，
08:24
and that's, of course,
amazing in its own right,
amazing in its own right,
173
492518
3313
当然，令人惊叹的是它自身的运算能力，
08:27
that mobile phones can do so many
billions and trillions of operations
billions and trillions of operations
174
495855
3468
每秒钟可以进行
几十亿甚至几万亿次的
几十亿甚至几万亿次的
运算。
08:31
per second.
175
499347
1248
08:32
What you're looking at is a phone
176
500619
1615
你所看到的是一台手机的
08:34
looking at one after another
picture of a bird,
picture of a bird,
177
502258
3547
相机对准了一张张含有鸟的图片，
08:37
and actually not only saying,
"Yes, it's a bird,"
"Yes, it's a bird,"
178
505829
2715
并且它不只能判断出，
“是的，这是一只鸟”，
“是的，这是一只鸟”，
而且还能用这种网络
来判断这些鸟的种类。
来判断这些鸟的种类。
08:40
but identifying the species of bird
with a network of this sort.
with a network of this sort.
179
508568
3411
08:44
So in that picture,
180
512890
1826
因此在这张图片中，
08:46
the x and the w are known,
and the y is the unknown.
and the y is the unknown.
181
514740
3802
x和w是已知的，y是未知的。
08:50
I'm glossing over the very
difficult part, of course,
difficult part, of course,
182
518566
2508
当然，我省略了非常复杂的那一部分，
08:53
which is how on earth
do we figure out the w,
do we figure out the w,
183
521098
3861
也就是我们如何判断出w？
为什么大脑能做出这样的判断？
08:56
the brain that can do such a thing?
184
524983
2187
08:59
How would we ever learn such a model?
185
527194
1834
我们是如何学会这种模式的？
09:01
So this process of learning,
of solving for w,
of solving for w,
186
529418
3233
在学习以及解出w的过程中，
如果我们使用简单的等式
09:04
if we were doing this
with the simple equation
with the simple equation
187
532675
2647
09:07
in which we think about these as numbers,
188
535346
2000
将这些都想象成数字，
那这道题就简单了： 6 = 2 x W，
09:09
we know exactly how to do that: 6 = 2 x w,
189
537370
2687
那么，用6除以2就可以得出答案。
09:12
well, we divide by two and we're done.
190
540081
3312
09:16
The problem is with this operator.
191
544001
2220
现在的问题就是这个运算符号。
09:18
So, division --
192
546823
1151
除法——
我们用除法是因为它是乘法的逆运算。
09:19
we've used division because
it's the inverse to multiplication,
it's the inverse to multiplication,
193
547998
3121
但就像我刚才说的，
09:23
but as I've just said,
194
551143
1440
乘法表述在这里其实不太准确。
09:24
the multiplication is a bit of a lie here.
195
552607
2449
09:27
This is a very, very complicated,
very non-linear operation;
very non-linear operation;
196
555080
3326
这是一个非常非常
复杂的非线性运算，
复杂的非线性运算，
它没有逆运算。
09:30
it has no inverse.
197
558430
1704
09:32
So we have to figure out a way
to solve the equation
to solve the equation
198
560158
3150
所以我们要找出一个不使用除号
09:35
without a division operator.
199
563332
2024
就能解出这个方程式的方法。
09:37
And the way to do that
is fairly straightforward.
is fairly straightforward.
200
565380
2343
其实非常简单。
只需要使用一点代数上的小技巧，
09:39
You just say, let's play
a little algebra trick,
a little algebra trick,
201
567747
2671
09:42
and move the six over
to the right-hand side of the equation.
to the right-hand side of the equation.
202
570442
2906
将6移到等式的右边。
09:45
Now, we're still using multiplication.
203
573372
1826
现在我们仍然使用乘法。
09:47
And that zero -- let's think
about it as an error.
about it as an error.
204
575675
3580
而这个0——我们就当它是一个误差。
09:51
In other words, if we've solved
for w the right way,
for w the right way,
205
579279
2515
换句话说，如果我们
能用正确的方法解出w，
能用正确的方法解出w，
09:53
then the error will be zero.
206
581818
1656
那么这个误差就为0。
09:55
And if we haven't gotten it quite right,
207
583498
1938
如果我们没有找到正确的答案，
那么这个误差就会大于0。
09:57
the error will be greater than zero.
208
585460
1749
所以现在我们可以通过
假设去缩小这个误差，
假设去缩小这个误差，
09:59
So now we can just take guesses
to minimize the error,
to minimize the error,
209
587233
3366
10:02
and that's the sort of thing
computers are very good at.
computers are very good at.
210
590623
2687
而这正是电脑所擅长的。
比如你最开始假设：
10:05
So you've taken an initial guess:
211
593334
1593
如果w = 0呢？
10:06
what if w = 0?
212
594951
1156
那么误差就为6。
如果w = 1呢？误差就变成了4。
10:08
Well, then the error is 6.
213
596131
1240
10:09
What if w = 1? The error is 4.
214
597395
1446
然后电脑就像玩游戏一样不断测试，
10:10
And then the computer can
sort of play Marco Polo,
sort of play Marco Polo,
215
598865
2367
10:13
and drive down the error close to zero.
216
601256
2367
将误差降低到接近于0。
10:15
As it does that, it's getting
successive approximations to w.
successive approximations to w.
217
603647
3374
这样就逐步逼近了w的值。
通常来说，它不可能获得完全精确的值，
但是经过很多步运算以后，
但是经过很多步运算以后，
10:19
Typically, it never quite gets there,
but after about a dozen steps,
but after about a dozen steps,
218
607045
3656
10:22
we're up to w = 2.999,
which is close enough.
which is close enough.
219
610725
4624
我们得到了 w = 2.999，
已经足够精确了。
已经足够精确了。
10:28
And this is the learning process.
220
616302
1814
以上就是这个学习过程。
10:30
So remember that what's been going on here
221
618140
2730
大家回想一下刚刚我们所做的，
10:32
is that we've been taking
a lot of known x's and known y's
a lot of known x's and known y's
222
620894
4378
我们用了很多已知的x和y的值，
10:37
and solving for the w in the middle
through an iterative process.
through an iterative process.
223
625296
3454
通过迭代法去解出中间的w，
10:40
It's exactly the same way
that we do our own learning.
that we do our own learning.
224
628774
3556
这也正是我们自己
在学习时所使用的方法。
在学习时所使用的方法。
在我们很小的时候，
会看到很多很多图像，
会看到很多很多图像，
10:44
We have many, many images as babies
225
632354
2230
10:46
and we get told, "This is a bird;
this is not a bird."
this is not a bird."
226
634608
2633
然后有人告诉我们：
“这个是鸟，这个不是鸟。”
“这个是鸟，这个不是鸟。”
10:49
And over time, through iteration,
227
637714
2098
经过一段时间的重复，
我们解出了w，建立起了
神经元之间的连接。
神经元之间的连接。
10:51
we solve for w, we solve
for those neural connections.
for those neural connections.
228
639836
2928
10:55
So now, we've held
x and w fixed to solve for y;
x and w fixed to solve for y;
229
643460
4086
那么现在，我们有了确定的
x和w。再要去解出Y
x和w。再要去解出Y
10:59
that's everyday, fast perception.
230
647570
1847
就会非常快了。
我们找到解出w的方法，
11:01
We figure out how we can solve for w,
231
649441
1763
这是一种学习，要困难得多，
11:03
that's learning, which is a lot harder,
232
651228
1903
11:05
because we need to do error minimization,
233
653155
1985
因为我们要用很多的训练样本，
去将误差最小化。
11:07
using a lot of training examples.
234
655164
1687
一年前，我们团队的
亚历克斯 · 莫尔德温采夫
亚历克斯 · 莫尔德温采夫
11:08
And about a year ago,
Alex Mordvintsev, on our team,
Alex Mordvintsev, on our team,
235
656875
3187
决定做一个实验，
看如果给定已知的w和y，
看如果给定已知的w和y，
11:12
decided to experiment
with what happens if we try solving for x,
with what happens if we try solving for x,
236
660086
3550
去解出x，会发生什么。
11:15
given a known w and a known y.
237
663660
2037
11:18
In other words,
238
666124
1151
换句话说，
你已经知道那是一只鸟
11:19
you know that it's a bird,
239
667299
1352
11:20
and you already have your neural network
that you've trained on birds,
that you've trained on birds,
240
668675
3303
并且也有一个接受过
鸟类识别训练的神经网络，
鸟类识别训练的神经网络，
那么一只鸟的图像是怎样的呢？
11:24
but what is the picture of a bird?
241
672002
2344
11:27
It turns out that by using exactly
the same error-minimization procedure,
the same error-minimization procedure,
242
675034
5024
我们发现，通过运用相同的
将误差最小化的步骤，
将误差最小化的步骤，
11:32
one can do that with the network
trained to recognize birds,
trained to recognize birds,
243
680082
3430
加上一个受过鸟类识别
训练的神经网络，
训练的神经网络，
11:35
and the result turns out to be ...
244
683536
3388
我们就可以得到
11:42
a picture of birds.
245
690400
1305
一张含有鸟的图片。
11:44
So this is a picture of birds
generated entirely by a neural network
generated entirely by a neural network
246
692814
3737
这是一张由一个进行过
鸟类识别训练的
鸟类识别训练的
神经网络所生成的鸟的图片，
11:48
that was trained to recognize birds,
247
696575
1826
11:50
just by solving for x
rather than solving for y,
rather than solving for y,
248
698425
3538
仅仅是通过解出x，而不是y，
并且重复不断的运行。
11:53
and doing that iteratively.
249
701987
1288
11:55
Here's another fun example.
250
703732
1847
这是另外一个有趣的例子
11:57
This was a work made
by Mike Tyka in our group,
by Mike Tyka in our group,
251
705603
3437
是我们团队的迈克 · 泰卡制作的 ，
他称之为“动物大游行”。
12:01
which he calls "Animal Parade."
252
709064
2308
12:03
It reminds me a little bit
of William Kentridge's artworks,
of William Kentridge's artworks,
253
711396
2876
这让我想起了威廉 ·肯特里奇的作品，
12:06
in which he makes sketches, rubs them out,
254
714296
2489
他先画一些素描，然后擦掉，
12:08
makes sketches, rubs them out,
255
716809
1460
再画一些素描，再擦掉，
用这种方法创作了一部影片。
12:10
and creates a movie this way.
256
718293
1398
在我们这个案例中，
12:11
In this case,
257
719715
1151
迈克在一个旨在识别和辨认
12:12
what Mike is doing is varying y
over the space of different animals,
over the space of different animals,
258
720890
3277
不同种类动物的神经网络中
12:16
in a network designed
to recognize and distinguish
to recognize and distinguish
259
724191
2382
将y变换成各种不同的动物。
12:18
different animals from each other.
260
726597
1810
这样你就得到了这个奇特的
动物图像的埃舍尔式变换效果。
动物图像的埃舍尔式变换效果。
12:20
And you get this strange, Escher-like
morph from one animal to another.
morph from one animal to another.
261
728431
3751
12:26
Here he and Alex together
have tried reducing
have tried reducing
262
734221
4614
他和亚历克斯还一起尝试了
将这些y降低到一个二维空间内，
12:30
the y's to a space of only two dimensions,
263
738859
2759
从而将被该神经网络识别出来的
12:33
thereby making a map
out of the space of all things
out of the space of all things
264
741642
3438
12:37
recognized by this network.
265
745104
1719
所有对象放到一张图上来。
通过这样的合成
12:38
Doing this kind of synthesis
266
746847
2023
12:40
or generation of imagery
over that entire surface,
over that entire surface,
267
748894
2382
或者在整个表面上生成图像，
在表面上不断的变换y，
你就创造出了一种图像——
你就创造出了一种图像——
12:43
varying y over the surface,
you make a kind of map --
you make a kind of map --
268
751300
2846
一个包含该神经网络能够
分辨出来的所有对象的视觉图像。
分辨出来的所有对象的视觉图像。
12:46
a visual map of all the things
the network knows how to recognize.
the network knows how to recognize.
269
754170
3141
所有的动物都在这儿，
犰狳在那个点上。
犰狳在那个点上。
12:49
The animals are all here;
"armadillo" is right in that spot.
"armadillo" is right in that spot.
270
757335
2865
12:52
You can do this with other kinds
of networks as well.
of networks as well.
271
760919
2479
你也可以用其它的神经网络
实现类似的目的。
实现类似的目的。
这是一个为识别和分辨出不同面孔
12:55
This is a network designed
to recognize faces,
to recognize faces,
272
763422
2874
而设计的神经网络。
12:58
to distinguish one face from another.
273
766320
2000
这里，我们输入一个y值，代表“我”，
13:00
And here, we're putting
in a y that says, "me,"
in a y that says, "me,"
274
768344
3249
我自己的面部参数。
13:03
my own face parameters.
275
771617
1575
13:05
And when this thing solves for x,
276
773216
1706
当它在解出x的时候，
13:06
it generates this rather crazy,
277
774946
2618
就生成了这张集不同视角
于一体，相当不可思议的，
于一体，相当不可思议的，
立体的、超现实的、迷幻版本的
13:09
kind of cubist, surreal,
psychedelic picture of me
psychedelic picture of me
278
777588
4428
我的面部图像。
13:14
from multiple points of view at once.
279
782040
1806
它之所以看起来像是集不同视角于一体，
13:15
The reason it looks like
multiple points of view at once
multiple points of view at once
280
783870
2734
是因为这个神经网络被设计成将一张脸
13:18
is because that network is designed
to get rid of the ambiguity
to get rid of the ambiguity
281
786628
3687
在不同姿势、不同光线之间产生的
13:22
of a face being in one pose
or another pose,
or another pose,
282
790339
2476
13:24
being looked at with one kind of lighting,
another kind of lighting.
another kind of lighting.
283
792839
3376
模棱两可的地方抹掉了。
因此当你开始这项复原工作时，
13:28
So when you do
this sort of reconstruction,
this sort of reconstruction,
284
796239
2085
如果不利用某种影像引导，
13:30
if you don't use some sort of guide image
285
798348
2304
13:32
or guide statistics,
286
800676
1211
或者统计引导，
13:33
then you'll get a sort of confusion
of different points of view,
of different points of view,
287
801911
3765
那么你就会得到一种
令人困惑的多视角的图像，
令人困惑的多视角的图像，
因为它是模棱两可的。
13:37
because it's ambiguous.
288
805700
1368
13:39
This is what happens if Alex uses
his own face as a guide image
his own face as a guide image
289
807786
4223
这就是亚历克斯在复原
我的面部的优化流程中，
我的面部的优化流程中，
用他自己的脸作为
影像引导时所得到的图像。
影像引导时所得到的图像。
13:44
during that optimization process
to reconstruct my own face.
to reconstruct my own face.
290
812033
3321
13:48
So you can see it's not perfect.
291
816284
2328
你可以看到它还不是十分完美。
我们在完善这个优化流程方面
13:50
There's still quite a lot of work to do
292
818636
1874
还有许多的工作要做。
13:52
on how we optimize
that optimization process.
that optimization process.
293
820534
2453
但是通过将我自己的脸
作为渲染过程中的引导，
作为渲染过程中的引导，
13:55
But you start to get something
more like a coherent face,
more like a coherent face,
294
823011
2827
13:57
rendered using my own face as a guide.
295
825862
2014
你已经可以得到一个
更清晰的面孔了。
更清晰的面孔了。
14:00
You don't have to start
with a blank canvas
with a blank canvas
296
828892
2501
你不需要完全从一块空白的画布
或白噪音开始。
14:03
or with white noise.
297
831417
1156
当你在解出x时，
14:04
When you're solving for x,
298
832597
1304
你可以从一个本身已经是
别的图像的x开始。
别的图像的x开始。
14:05
you can begin with an x,
that is itself already some other image.
that is itself already some other image.
299
833925
3889
正如这个小小的展示那样。
14:09
That's what this little demonstration is.
300
837838
2556
这是一个设计为用来将所有物品——
14:12
This is a network
that is designed to categorize
that is designed to categorize
301
840418
4122
人造结构、动物等进行分类的神经网络。
14:16
all sorts of different objects --
man-made structures, animals ...
man-made structures, animals ...
302
844564
3119
我们从一张云图开始，
14:19
Here we're starting
with just a picture of clouds,
with just a picture of clouds,
303
847707
2593
在优化过程中，
14:22
and as we optimize,
304
850324
1671
这个神经网络正在不停地计算
它在云中看到了什么。
它在云中看到了什么。
14:24
basically, this network is figuring out
what it sees in the clouds.
what it sees in the clouds.
305
852019
4486
14:28
And the more time
you spend looking at this,
you spend looking at this,
306
856931
2320
你花越多的时间盯着这张图，
14:31
the more things you also
will see in the clouds.
will see in the clouds.
307
859275
2753
你就会在云中看到越多的东西。
14:35
You could also use the face network
to hallucinate into this,
to hallucinate into this,
308
863004
3375
你也可以使用面部识别
神经网络去产生迷幻效果，
神经网络去产生迷幻效果，
然后就可以得到这种不可思议的东西。
14:38
and you get some pretty crazy stuff.
309
866403
1812
（观众笑声）
14:40
(Laughter)
310
868239
1150
14:42
Or, Mike has done some other experiments
311
870401
2744
或者可以像迈克做的另外一个实验那样，
他还是利用那张云图，
14:45
in which he takes that cloud image,
312
873169
3905
使它幻化、再放大，
幻化再放大，幻化再放大.
幻化再放大，幻化再放大.
14:49
hallucinates, zooms, hallucinates,
zooms hallucinates, zooms.
zooms hallucinates, zooms.
313
877098
3507
这样一来，
14:52
And in this way,
314
880629
1151
我想你就可以得到
这个网络的神游状态，
这个网络的神游状态，
14:53
you can get a sort of fugue state
of the network, I suppose,
of the network, I suppose,
315
881804
3675
或者某种自由联想，
14:57
or a sort of free association,
316
885503
3680
仿佛这个网络正在吞噬自己的尾巴。
15:01
in which the network
is eating its own tail.
is eating its own tail.
317
889207
2227
15:03
So every image is now the basis for,
318
891458
3421
因此每一张图都是
下一张图的基础，决定了
下一张图的基础，决定了
“我觉得接下来会看到什么？
15:06
"What do I think I see next?
319
894903
1421
接下来又会看到什么？
接下来还会看到什么？”
接下来还会看到什么？”
15:08
What do I think I see next?
What do I think I see next?"
What do I think I see next?"
320
896348
2803
15:11
I showed this for the first time in public
321
899487
2936
我第一次公开展示这些是在西雅图，
15:14
to a group at a lecture in Seattle
called "Higher Education" --
called "Higher Education" --
322
902447
5437
为一个团队做的一次名为
“高等教育”的讲座上——
“高等教育”的讲座上——
15:19
this was right after
marijuana was legalized.
marijuana was legalized.
323
907908
2437
刚好就在大麻合法化之后。
15:22
(Laughter)
324
910369
2415
（观众笑声）
15:26
So I'd like to finish up quickly
325
914627
2104
在结束我的演讲前，
15:28
by just noting that this technology
is not constrained.
is not constrained.
326
916755
4255
我想再提醒各位，
这种技术是不受限的。
这种技术是不受限的。
15:33
I've shown you purely visual examples
because they're really fun to look at.
because they're really fun to look at.
327
921034
3665
我给你们看了一些纯粹的视觉实例，
因为它们看起来真的很有趣。
因为它们看起来真的很有趣。
它不是一种纯粹的视觉技术。
15:36
It's not a purely visual technology.
328
924723
2451
15:39
Our artist collaborator, Ross Goodwin,
329
927198
1993
我们的合作者，艺术家罗斯 · 古德温
做了一个实验，他用相机拍了一张照片，
15:41
has done experiments involving
a camera that takes a picture,
a camera that takes a picture,
330
929215
3671
15:44
and then a computer in his backpack
writes a poem using neural networks,
writes a poem using neural networks,
331
932910
4234
然后他背包里的电脑
基于这张照片的内容，
基于这张照片的内容，
15:49
based on the contents of the image.
332
937168
1944
用神经网络作了一首诗。
15:51
And that poetry neural network
has been trained
has been trained
333
939136
2947
这个作诗的神经网络已经接受过
15:54
on a large corpus of 20th-century poetry.
334
942107
2234
大量的20世纪诗歌的训练。
15:56
And the poetry is, you know,
335
944365
1499
其实我觉得
那首诗还不赖。
15:57
I think, kind of not bad, actually.
336
945888
1914
（观众笑声）
15:59
(Laughter)
337
947826
1384
下面，
16:01
In closing,
338
949234
1159
16:02
I think that per Michelangelo,
339
950417
2132
再回到米开朗基罗那句名言，
我想他是对的，
16:04
I think he was right;
340
952573
1234
16:05
perception and creativity
are very intimately connected.
are very intimately connected.
341
953831
3436
感知和创意是密不可分的。
16:09
What we've just seen are neural networks
342
957611
2634
我们刚刚所看到的是一些
16:12
that are entirely trained to discriminate,
343
960269
2303
完全被训练成去区分，
16:14
or to recognize different
things in the world,
things in the world,
344
962596
2242
或辨别世上的不同物品，
能够逆向运行、成生图像的神经网络。
16:16
able to be run in reverse, to generate.
345
964862
3161
16:20
One of the things that suggests to me
346
968047
1783
我从中受到的启发之一就是，
不仅米开朗基罗真的看到了
16:21
is not only that
Michelangelo really did see
Michelangelo really did see
347
969854
2398
石头中的雕像，
16:24
the sculpture in the blocks of stone,
348
972276
2452
而且任何的生物、任何人、任何外星人，
16:26
but that any creature,
any being, any alien
any being, any alien
349
974752
3638
只要能够有这样的感知，
16:30
that is able to do
perceptual acts of that sort
perceptual acts of that sort
350
978414
3657
也就能够创造，
16:34
is also able to create
351
982095
1375
16:35
because it's exactly the same
machinery that's used in both cases.
machinery that's used in both cases.
352
983494
3224
因为它们都运用了截然相同的机制。
16:38
Also, I think that perception
and creativity are by no means
and creativity are by no means
353
986742
4532
另外，我想感知和创意决不是
人类所特有的。
16:43
uniquely human.
354
991298
1210
16:44
We start to have computer models
that can do exactly these sorts of things.
that can do exactly these sorts of things.
355
992532
3708
我们开始有了可以
完成这些事的电脑模型。
完成这些事的电脑模型。
这应当不足为奇，因为大脑会运算。
16:48
And that ought to be unsurprising;
the brain is computational.
the brain is computational.
356
996264
3328
16:51
And finally,
357
999616
1657
最后，
电脑运算最开始是作为
设计智能机器的一种练习。
设计智能机器的一种练习。
16:53
computing began as an exercise
in designing intelligent machinery.
in designing intelligent machinery.
358
1001297
4668
16:57
It was very much modeled after the idea
359
1005989
2462
它在很大程度上仿照了我们如何
17:00
of how could we make machines intelligent.
360
1008475
3013
让机器变得智能这一理念。
17:03
And we finally are starting to fulfill now
361
1011512
2162
而我们也终于开始能够实现
图灵、冯 · 诺依曼、
17:05
some of the promises
of those early pioneers,
of those early pioneers,
362
1013698
2406
17:08
of Turing and von Neumann
363
1016128
1713
麦卡洛克和皮兹
17:09
and McCulloch and Pitts.
364
1017865
2265
这些先驱的一些期望了。
17:12
And I think that computing
is not just about accounting
is not just about accounting
365
1020154
4098
我觉得电脑不仅仅是拿来计算，
17:16
or playing Candy Crush or something.
366
1024276
2147
或者玩游戏的。
17:18
From the beginning,
we modeled them after our minds.
we modeled them after our minds.
367
1026447
2578
从一开始，我们就是
仿照大脑来制造它们的。
仿照大脑来制造它们的。
而它们也赋予了我们能够
更好的理解我们的大脑，
更好的理解我们的大脑，
17:21
And they give us both the ability
to understand our own minds better
to understand our own minds better
368
1029049
3269
并且拓展其潜力的能力。
17:24
and to extend them.
369
1032342
1529
17:26
Thank you very much.
370
1034627
1167
非常感谢。
（观众掌声）
17:27
(Applause)
371
1035818
5939
ABOUT THE SPEAKER
Blaise Agüera y Arcas - Software architectBlaise Agüera y Arcas works on machine learning at Google. Previously a Distinguished Engineer at Microsoft, he has worked on augmented reality, mapping, wearable computing and natural user interfaces.
Why you should listen
Blaise Agüera y Arcas is principal scientist at Google, where he leads a team working on machine intelligence for mobile devices. His group works extensively with deep neural nets for machine perception and distributed learning, and it also investigates so-called "connectomics" research, assessing maps of connections within the brain.
Agüera y Arcas' background is as multidimensional as the visions he helps create. In the 1990s, he authored patents on both video compression and 3D visualization techniques, and in 2001, he made an influential computational discovery that cast doubt on Gutenberg's role as the father of movable type.
He also created Seadragon (acquired by Microsoft in 2006), the visualization technology that gives Photosynth its amazingly smooth digital rendering and zoom capabilities. Photosynth itself is a vastly powerful piece of software capable of taking a wide variety of images, analyzing them for similarities, and grafting them together into an interactive three-dimensional space. This seamless patchwork of images can be viewed via multiple angles and magnifications, allowing us to look around corners or “fly” in for a (much) closer look. Simply put, it could utterly transform the way we experience digital images.
He joined Microsoft when Seadragon was acquired by Live Labs in 2006. Shortly after the acquisition of Seadragon, Agüera y Arcas directed his team in a collaboration with Microsoft Research and the University of Washington, leading to the first public previews of Photosynth several months later. His TED Talk on Seadragon and Photosynth in 2007 is rated one of TED's "most jaw-dropping." He returned to TED in 2010 to demo Bing’s augmented reality maps.
Fun fact: According to the author, Agüera y Arcas is the inspiration for the character Elgin in the 2012 best-selling novel Where'd You Go, Bernadette?
Blaise Agüera y Arcas | Speaker | TED.com