請雙擊下方英文字幕播放視頻。
譯者: 易帆 余
00:12
So, I lead a team at Google
that works on machine intelligence;
0
12800
3124
我在 Google 帶領
一個團隊做機械智慧;
00:15
in other words, the engineering discipline
of making computers and devices
1
15948
4650
換句話說,就是制定一些訓練方法,
00:20
able to do some of the things
that brains do.
2
20622
2419
讓電腦和裝置能做些大腦做的事。
00:23
And this makes us
interested in real brains
3
23439
3099
而這也讓我們對真實的大腦
00:26
and neuroscience as well,
4
26562
1289
以及神經科學產生了興趣,
00:27
and especially interested
in the things that our brains do
5
27875
4172
特別是一些我們大腦能做
但電腦仍無法呈現出來的事。
00:32
that are still far superior
to the performance of computers.
6
32071
4042
00:37
Historically, one of those areas
has been perception,
7
37209
3609
長期以來,機械智慧的
其中一個領域談的就是機械感知,
00:40
the process by which things
out there in the world --
8
40842
3039
它是一種轉化的過程——
00:43
sounds and images --
9
43905
1584
像是把聲音和影像——
00:45
can turn into concepts in the mind.
10
45513
2178
轉化成心智上的概念。
00:48
This is essential for our own brains,
11
48235
2517
這是我們大腦必備的能力,
00:50
and it's also pretty useful on a computer.
12
50776
2464
這個能力對電腦來說也很有用。
00:53
The machine perception algorithms,
for example, that our team makes,
13
53636
3350
所謂的機械感知演算法,
像是我們團隊做的,
能讓你 Google 相簿裡的照片
00:57
are what enable your pictures
on Google Photos to become searchable,
14
57010
3874
01:00
based on what's in them.
15
60908
1397
根據照片裡的東西
把它們變成可以被搜尋的資料。
01:03
The flip side of perception is creativity:
16
63594
3493
感知的另一面是創意:
把概念轉化成另一種東西。
01:07
turning a concept into something
out there into the world.
17
67111
3038
所以過去幾年,
我們團隊在機器感知上的努力,
01:10
So over the past year,
our work on machine perception
18
70173
3555
01:13
has also unexpectedly connected
with the world of machine creativity
19
73752
4859
已經可以把創意與
機器藝術結合在一起。
01:18
and machine art.
20
78635
1160
01:20
I think Michelangelo
had a penetrating insight
21
80556
3284
我覺得米開朗基羅對「感知」
與「創意」這兩者之間的關係
01:23
into to this dual relationship
between perception and creativity.
22
83864
3656
有一種很透析的看法。
他有一句名言:
01:28
This is a famous quote of his:
23
88023
2006
「每一塊石頭裡都藏著一座雕像,
01:30
"Every block of stone
has a statue inside of it,
24
90053
3323
等待雕刻家將它雕塑出來。」
01:34
and the job of the sculptor
is to discover it."
25
94036
3002
所以我覺得米開朗基羅
當時的體悟是:
01:38
So I think that what
Michelangelo was getting at
26
98029
3216
01:41
is that we create by perceiving,
27
101269
3180
我們的「創意」來自「感知」,
01:44
and that perception itself
is an act of imagination
28
104473
3023
而感知本身就是一個想像行為
01:47
and is the stuff of creativity.
29
107520
2461
及創意的來源。
01:50
The organ that does all the thinking
and perceiving and imagining,
30
110691
3925
人體中有一個器官
能做出思考、感受和想像,
01:54
of course, is the brain.
31
114640
1588
當然,那就是我們的大腦。
我想先簡單地來談一談
01:57
And I'd like to begin
with a brief bit of history
32
117089
2545
01:59
about what we know about brains.
33
119658
2302
我們對大腦認知的歷史。
02:02
Because unlike, say,
the heart or the intestines,
34
122496
2446
因為大腦不像我們的心臟或腸道,
02:04
you really can't say very much
about a brain by just looking at it,
35
124966
3144
你不能光用看的來瞭解大腦,
光靠肉眼根本看不出個所以然來。
02:08
at least with the naked eye.
36
128134
1412
02:09
The early anatomists who looked at brains
37
129983
2416
早期研究大腦的解剖學家,
02:12
gave the superficial structures
of this thing all kinds of fanciful names,
38
132423
3807
在大腦表皮結構上
取了許多稀奇古怪的名字,
02:16
like hippocampus, meaning "little shrimp."
39
136254
2433
例如海馬體,意思是「小蝦子」。
02:18
But of course that sort of thing
doesn't tell us very much
40
138711
2764
當然,這樣的命名方式
並沒有讓我們對
大腦的認識有太多的幫助。
02:21
about what's actually going on inside.
41
141499
2318
02:24
The first person who, I think, really
developed some kind of insight
42
144780
3613
我認為,第一個有真正深入了解
02:28
into what was going on in the brain
43
148417
1930
大腦如何運作的,
02:30
was the great Spanish neuroanatomist,
Santiago Ramón y Cajal,
44
150371
3920
是偉大的西班牙神經解剖學家
桑地牙哥·拉蒙卡哈,
02:34
in the 19th century,
45
154315
1544
他在十九世紀,
02:35
who used microscopy and special stains
46
155883
3755
就已經開始用顯微鏡和特殊染劑
02:39
that could selectively fill in
or render in very high contrast
47
159662
4170
把大腦裡的特定細胞篩選出來染色,
02:43
the individual cells in the brain,
48
163856
2008
或以強烈的對比色來觀察細胞,
02:45
in order to start to understand
their morphologies.
49
165888
3154
這樣做,是為了瞭解
它們的形態結構。
02:49
And these are the kinds of drawings
that he made of neurons
50
169972
2891
這些是他在十九世紀時
02:52
in the 19th century.
51
172887
1209
畫的神經細胞圖,
這一張是鳥的大腦。
02:54
This is from a bird brain.
52
174120
1884
但當時已經可以看到
各式各樣不同的細胞圖片,
02:56
And you see this incredible variety
of different sorts of cells,
53
176028
3057
即使細胞的原理
在當時是個相當新穎的概念。
02:59
even the cellular theory itself
was quite new at this point.
54
179109
3435
03:02
And these structures,
55
182568
1278
這些結構,
03:03
these cells that have these arborizations,
56
183870
2259
這些樹枝狀的細胞結構,
可以延伸到相當相當長──
03:06
these branches that can go
very, very long distances --
57
186153
2608
03:08
this was very novel at the time.
58
188785
1616
在當時來講,
這樣的發現算是相當神奇了。
03:10
They're reminiscent, of course, of wires.
59
190779
2903
當然,它們也會讓人聯想到電線,
03:13
That might have been obvious
to some people in the 19th century;
60
193706
3457
這對 19 世紀的人來說,
這樣的比喻可能比較恰當,
03:17
the revolutions of wiring and electricity
were just getting underway.
61
197187
4314
因為當時電線和電力的變革
正如火如荼的進行。
03:21
But in many ways,
62
201964
1178
但就很多方面來說,
像拉蒙卡哈這樣的顯微鏡解剖圖
03:23
these microanatomical drawings
of Ramón y Cajal's, like this one,
63
203166
3313
03:26
they're still in some ways unsurpassed.
64
206503
2332
現在看來還是很厲害。
03:28
We're still more than a century later,
65
208859
1854
但我們卻在一個世紀後,
03:30
trying to finish the job
that Ramón y Cajal started.
66
210737
2825
才想試著去完成
當年拉蒙卡哈的研究。
03:33
These are raw data from our collaborators
67
213586
3134
這些原始資料,來自我們
03:36
at the Max Planck Institute
of Neuroscience.
68
216744
2881
馬克斯·普朗克
神經科學機構的合作夥伴。
03:39
And what our collaborators have done
69
219649
1790
而我們的合作夥伴的工作就是
03:41
is to image little pieces of brain tissue.
70
221463
5001
把大腦組織切成
一小片一小片的圖像。
03:46
The entire sample here
is about one cubic millimeter in size,
71
226488
3326
整個樣本的大小
大約只有 1 立方毫米,
03:49
and I'm showing you a very,
very small piece of it here.
72
229838
2621
我展示給各位看的只有小小的一片。
03:52
That bar on the left is about one micron.
73
232483
2346
你可以看到,
左邊的長度標誌僅有一微米。
03:54
The structures you see are mitochondria
74
234853
2409
各位現在看到的結構是粒線體,
03:57
that are the size of bacteria.
75
237286
2044
大小跟細菌一樣。
03:59
And these are consecutive slices
76
239354
1551
這些連續切片圖,
04:00
through this very, very
tiny block of tissue.
77
240929
3148
是由一塊很小的組織中
一片片切出來的。
舉個例子做比較,
04:04
Just for comparison's sake,
78
244101
2403
04:06
the diameter of an average strand
of hair is about 100 microns.
79
246528
3792
一根頭髮的直徑
大約有 100 微米。
04:10
So we're looking at something
much, much smaller
80
250344
2274
我們在研究的
是比一根頭髮還更細更小的東西。
04:12
than a single strand of hair.
81
252642
1398
而這一系列的電子顯微鏡切片圖像,
04:14
And from these kinds of serial
electron microscopy slices,
82
254064
4031
可以組成像這樣的
神經元 3D 立體成像。
04:18
one can start to make reconstructions
in 3D of neurons that look like these.
83
258119
5008
這些和拉蒙卡哈
當年的研究相去不遠。
04:23
So these are sort of in the same
style as Ramón y Cajal.
84
263151
3157
04:26
Only a few neurons lit up,
85
266332
1492
但只有幾個神經元可以打光,
04:27
because otherwise we wouldn't
be able to see anything here.
86
267848
2781
否則我們會看不到東西。
04:30
It would be so crowded,
87
270653
1312
因為空間太壅擠、
04:31
so full of structure,
88
271989
1330
結構太複雜了,
04:33
of wiring all connecting
one neuron to another.
89
273343
2724
神經元蜿蜒地一個接著一個。
04:37
So Ramón y Cajal was a little bit
ahead of his time,
90
277293
2804
所以,拉蒙卡哈在當時
也算是走在時代的尖端,
但在那之後的幾十年,
04:40
and progress on understanding the brain
91
280121
2555
04:42
proceeded slowly
over the next few decades.
92
282700
2271
人類對大腦的認識卻相當緩慢。
04:45
But we knew that neurons used electricity,
93
285455
2853
但我們已經知道
神經元是利用電子傳遞訊號,
04:48
and by World War II, our technology
was advanced enough
94
288332
2936
到第二次世界大戰前,
我們的科技已經進步到
04:51
to start doing real electrical
experiments on live neurons
95
291292
2806
可以在活體神經元上做電子實驗,
用來更好地理解它們是如何運作的。
04:54
to better understand how they worked.
96
294122
2106
04:56
This was the very same time
when computers were being invented,
97
296631
4356
這也是電腦被發明出來的時間,
當初有一個模擬人腦的基礎想法——
05:01
very much based on the idea
of modeling the brain --
98
301011
3100
是由艾倫·圖靈所提出,
他稱之為「智能機械」,
05:04
of "intelligent machinery,"
as Alan Turing called it,
99
304135
3085
05:07
one of the fathers of computer science.
100
307244
1991
他是計算機科學之父之一。
05:09
Warren McCulloch and Walter Pitts
looked at Ramón y Cajal's drawing
101
309923
4632
當時沃倫麥卡洛克和華特彼特斯
(人工神經科學家)
05:14
of visual cortex,
102
314579
1317
看到的視覺皮質圖,
05:15
which I'm showing here.
103
315920
1562
就是上面這張拉蒙卡哈的圖片。
05:17
This is the cortex that processes
imagery that comes from the eye.
104
317506
4442
這個皮質層是負責把
眼睛傳來的訊號轉換成圖像。
05:22
And for them, this looked
like a circuit diagram.
105
322424
3508
他們當時發現,
它看起來像是一張電路圖。
05:26
So there are a lot of details
in McCulloch and Pitts's circuit diagram
106
326353
3835
雖然麥卡洛克和彼特斯
在電路圖上有很多細節不太正確,
05:30
that are not quite right.
107
330212
1352
05:31
But this basic idea
108
331588
1235
但這樣的基礎概念,
05:32
that visual cortex works like a series
of computational elements
109
332847
3992
視覺皮層的工作原理
05:36
that pass information
one to the next in a cascade,
110
336863
2746
像一系列的計算子
在串聯的電路圖上傳遞著資訊,
05:39
is essentially correct.
111
339633
1602
這樣的概念卻是相當正確的。
05:41
Let's talk for a moment
112
341259
2350
我們稍微聊一下,
05:43
about what a model for processing
visual information would need to do.
113
343633
4032
產生視覺資訊的模型,
需要做哪些事情。
05:48
The basic task of perception
114
348228
2741
覺察力的基本任務就是
05:50
is to take an image like this one and say,
115
350993
4194
比如說,看到這一張圖片,
就要會判斷出,「這是一隻鳥」,
05:55
"That's a bird,"
116
355211
1176
05:56
which is a very simple thing
for us to do with our brains.
117
356411
2874
這對我們大腦來說是很簡單的任務。
05:59
But you should all understand
that for a computer,
118
359309
3421
但各位要知道,這對電腦來說
06:02
this was pretty much impossible
just a few years ago.
119
362754
3087
在幾年前根本是不可能的事。
06:05
The classical computing paradigm
120
365865
1916
傳統的計算模式
06:07
is not one in which
this task is easy to do.
121
367805
2507
根本不太容易跑出來這樣的任務。
06:11
So what's going on between the pixels,
122
371366
2552
所以,像素、
06:13
between the image of the bird
and the word "bird,"
123
373942
4028
鳥圖與文字之間,
06:17
is essentially a set of neurons
connected to each other
124
377994
2814
一定要有一組彼此連結的神經元
06:20
in a neural network,
125
380832
1155
在神經網路內相互作用著,
就像我這張示意圖。
06:22
as I'm diagramming here.
126
382011
1223
06:23
This neural network could be biological,
inside our visual cortices,
127
383258
3272
這張神經網路圖
就像我們的視覺皮質運作原理。
06:26
or, nowadays, we start
to have the capability
128
386554
2162
如今,我們已經有能力
06:28
to model such neural networks
on the computer.
129
388740
2454
用電腦來模擬這樣的神經網路。
06:31
And I'll show you what
that actually looks like.
130
391834
2353
接下來我向各位展示一下,
實際的操作大概是怎樣。
06:34
So the pixels you can think
about as a first layer of neurons,
131
394211
3416
圖片的像素你可以把它想像成是
第一層的神經元,
06:37
and that's, in fact,
how it works in the eye --
132
397651
2239
實際上,就是眼睛裡面
像素的呈現方式,
06:39
that's the neurons in the retina.
133
399914
1663
像素是透過
視網膜上的神經元做傳遞。
06:41
And those feed forward
134
401601
1500
而這些前饋資訊
會一層一層地傳遞到下一層神經元,
06:43
into one layer after another layer,
after another layer of neurons,
135
403125
3403
06:46
all connected by synapses
of different weights.
136
406552
3033
全部由不同的「突觸權重」所連結。
06:49
The behavior of this network
137
409609
1335
神經網路的行為
06:50
is characterized by the strengths
of all of those synapses.
138
410968
3284
全都由這些突觸的強度所控制。
06:54
Those characterize the computational
properties of this network.
139
414276
3288
它們決定了神經網路的計算模式。
06:57
And at the end of the day,
140
417588
1470
最後,
會有一個或一小群的
神經元發出訊號,
06:59
you have a neuron
or a small group of neurons
141
419082
2447
07:01
that light up, saying, "bird."
142
421553
1647
辨識出該圖片就是,「鳥」。
07:03
Now I'm going to represent
those three things --
143
423824
3132
我現在要來解釋一下這三個元素——
07:06
the input pixels and the synapses
in the neural network,
144
426980
4696
輸入的「像素」、
神經網路裡的「突觸」、
07:11
and bird, the output --
145
431700
1585
還有「鳥」這個輸出的字元——
它們是如何運作的。
07:13
by three variables: x, w and y.
146
433309
3057
它們是由三種變數所組成,
x、w 和 y。
07:16
There are maybe a million or so x's --
147
436853
1811
圖片中可能有一百多萬個 x ——
07:18
a million pixels in that image.
148
438688
1953
100 多萬個像素。
07:20
There are billions or trillions of w's,
149
440665
2446
而 w 可能有數十億或好幾兆個,
它們代表著神經網路中
各個突觸的權重。
07:23
which represent the weights of all
these synapses in the neural network.
150
443135
3421
07:26
And there's a very small number of y's,
151
446580
1875
而這個網路能輸出的 y
07:28
of outputs that that network has.
152
448479
1858
只有少數幾個。
07:30
"Bird" is only four letters, right?
153
450361
1749
「bird」只有四個字母,對吧?
我們假設它的原理是
一個簡單的公式,
07:33
So let's pretend that this
is just a simple formula,
154
453088
3426
07:36
x "x" w = y.
155
456538
2163
x 「乘以」 w = y
07:38
I'm putting the times in scare quotes
156
458725
2036
我把乘法符號用引號標示起來
07:40
because what's really
going on there, of course,
157
460785
2280
因為它其實是一個
非常複雜的數學運算概念。
07:43
is a very complicated series
of mathematical operations.
158
463089
3046
這個方程式
07:47
That's one equation.
159
467172
1221
07:48
There are three variables.
160
468417
1672
有三個變數,
我們都知道,如果你想要
解開這個方程式,
07:50
And we all know
that if you have one equation,
161
470113
2726
07:52
you can solve one variable
by knowing the other two things.
162
472863
3642
可以從兩個已知數
交叉算出未知的數。
所以要推斷出
07:57
So the problem of inference,
163
477158
3380
08:00
that is, figuring out
that the picture of a bird is a bird,
164
480562
2873
圖片中的影像是一隻鳥,
08:03
is this one:
165
483459
1274
可以用這種方式得知:
08:04
it's where y is the unknown
and w and x are known.
166
484757
3459
y 是未知數,而 w 和 x 是已知數。
08:08
You know the neural network,
you know the pixels.
167
488240
2459
已知神經網路和圖片像素,
08:10
As you can see, that's actually
a relatively straightforward problem.
168
490723
3327
其實可以很直接的就得到答案,
2x3=6,就做完了。
08:14
You multiply two times three
and you're done.
169
494074
2186
08:16
I'll show you an artificial neural network
170
496862
2123
我向各位展示一個
我們最近做的人工神經網路,
08:19
that we've built recently,
doing exactly that.
171
499009
2296
08:21
This is running in real time
on a mobile phone,
172
501634
2860
它可以在手機上做及時的操作,
08:24
and that's, of course,
amazing in its own right,
173
504518
3313
當然,手機的運算能力相當驚人,
08:27
that mobile phones can do so many
billions and trillions of operations
174
507855
3468
手機每秒
可以做出數十億至上兆次的運算。
08:31
per second.
175
511347
1248
08:32
What you're looking at is a phone
176
512619
1615
你現在看到的是一隻手機
08:34
looking at one after another
picture of a bird,
177
514258
3547
正對著一張張的鳥圖拍照,
08:37
and actually not only saying,
"Yes, it's a bird,"
178
517829
2715
手機不但可以正確的說出,
「是的,這是一隻鳥。」
08:40
but identifying the species of bird
with a network of this sort.
179
520568
3411
還能透過神經網路分類
分辨出這是哪一種鳥。
08:44
So in that picture,
180
524890
1826
所以,在這些圖片上,
08:46
the x and the w are known,
and the y is the unknown.
181
526740
3802
x 和 w 是已知,而 y 是未知。
08:50
I'm glossing over the very
difficult part, of course,
182
530566
2508
我現在來解釋一下這個
最困難的 「w」,
我們到底是如何算出來的?
08:53
which is how on earth
do we figure out the w,
183
533098
3861
08:56
the brain that can do such a thing?
184
536983
2187
為什麼大腦可以做出這樣的判斷?
08:59
How would we ever learn such a model?
185
539194
1834
我們到底是如何學到
這樣的認知模式的?
09:01
So this process of learning,
of solving for w,
186
541418
3233
這個學習的過程,
是一個求解 w 的過程,
09:04
if we were doing this
with the simple equation
187
544675
2647
如果我們要解這個一次方程式,
09:07
in which we think about these as numbers,
188
547346
2000
當它們都是數字時,
09:09
we know exactly how to do that: 6 = 2 x w,
189
549370
2687
我們都知道如何解 6=2 x w,
我們只要把 6 除以 2
就可以得到答案。
09:12
well, we divide by two and we're done.
190
552081
3312
問題在於這個運算符號,
09:16
The problem is with this operator.
191
556001
2220
09:18
So, division --
192
558823
1151
除法這個符號——
09:19
we've used division because
it's the inverse to multiplication,
193
559998
3121
我們會用除法的方式求解,
是因為它跟乘法相反,
但就如同我剛剛提到的,
09:23
but as I've just said,
194
563143
1440
09:24
the multiplication is a bit of a lie here.
195
564607
2449
乘法在這裡有點像是個幌子。
這是非常非常複雜的概念,
它們是「非線性運算」的概念;
09:27
This is a very, very complicated,
very non-linear operation;
196
567080
3326
09:30
it has no inverse.
197
570430
1704
無法直接用除的求解。
所以,我們要另外
找個方法來解方程式,
09:32
So we have to figure out a way
to solve the equation
198
572158
3150
09:35
without a division operator.
199
575332
2024
而不能直接用除的。
09:37
And the way to do that
is fairly straightforward.
200
577380
2343
方法相當簡單,
09:39
You just say, let's play
a little algebra trick,
201
579747
2671
可以說,我們只用了點
代數的小技巧,
09:42
and move the six over
to the right-hand side of the equation.
202
582442
2906
將 6 移動到等號的右邊。
09:45
Now, we're still using multiplication.
203
585372
1826
如此我們就可以繼續用乘法來運算。
09:47
And that zero -- let's think
about it as an error.
204
587675
3580
而等號左邊的零——
我們把它想像成是誤差。
09:51
In other words, if we've solved
for w the right way,
205
591279
2515
換言之,如果要解出 w,
09:53
then the error will be zero.
206
593818
1656
誤差就要變成 0。
09:55
And if we haven't gotten it quite right,
207
595498
1938
如果我們沒找到答案
09:57
the error will be greater than zero.
208
597460
1749
誤差會永遠大於 0。
09:59
So now we can just take guesses
to minimize the error,
209
599233
3366
所以,我們現在
只能用猜的來縮小誤差,
10:02
and that's the sort of thing
computers are very good at.
210
602623
2687
而這就是電腦非常擅長的地方。
10:05
So you've taken an initial guess:
211
605334
1593
所以,你會從頭開始猜:
10:06
what if w = 0?
212
606951
1156
假設 w=0
那誤差會等於6
10:08
Well, then the error is 6.
213
608131
1240
10:09
What if w = 1? The error is 4.
214
609395
1446
但假如 w=1 呢?誤差等於 4。
10:10
And then the computer can
sort of play Marco Polo,
215
610865
2367
接下來電腦有點像是在玩
馬可波羅探索遊戲,
10:13
and drive down the error close to zero.
216
613256
2367
探索到誤差接近零為止。
10:15
As it does that, it's getting
successive approximations to w.
217
615647
3374
當它一直探索到零,
那麼 w 就解出來了。
原則上,它會不停探索直到接近零,
但大約經過多次步驟後,
10:19
Typically, it never quite gets there,
but after about a dozen steps,
218
619045
3656
10:22
we're up to w = 2.999,
which is close enough.
219
622725
4624
我們就能得出 w=2.999,
相當接近了。
10:28
And this is the learning process.
220
628302
1814
這就是電腦學習的過程。
回想一下剛剛發生了什麼事情,
10:30
So remember that what's been going on here
221
630140
2730
10:32
is that we've been taking
a lot of known x's and known y's
222
632894
4378
我們有很多已知的 x 和 y,
10:37
and solving for the w in the middle
through an iterative process.
223
637296
3454
透過重複迭代的過程解出了 w。
10:40
It's exactly the same way
that we do our own learning.
224
640774
3556
而這就是我們人類學習的過程,
10:44
We have many, many images as babies
225
644354
2230
我們從小看了很多圖片
10:46
and we get told, "This is a bird;
this is not a bird."
226
646608
2633
被告知「這是鳥」,「這不是鳥」;
10:49
And over time, through iteration,
227
649714
2098
經過了一段時間,不停地重複,
10:51
we solve for w, we solve
for those neural connections.
228
651836
2928
我們解出了 w,
產生了神經元的連結關係。
10:55
So now, we've held
x and w fixed to solve for y;
229
655460
4086
所以現在,我們的 x 和 w
是固定數,可以解出 y;
10:59
that's everyday, fast perception.
230
659570
1847
這就是我們人類每天
經常性的快速直覺判斷。
11:01
We figure out how we can solve for w,
231
661441
1763
我們搞懂了如何解出 w,
11:03
that's learning, which is a lot harder,
232
663228
1903
而學習本身是一條相當艱辛的路程,
因為為了讓誤差最小化,
11:05
because we need to do error minimization,
233
665155
1985
我們必須使用很多的訓練樣本。
11:07
using a lot of training examples.
234
667164
1687
11:08
And about a year ago,
Alex Mordvintsev, on our team,
235
668875
3187
約一年前,我們團隊的
艾力克斯摩文斯夫
決定做個實驗,
11:12
decided to experiment
with what happens if we try solving for x,
236
672086
3550
看看如果我們試著給出了 w 和 y,
解出來的 x 會變什麼樣。
11:15
given a known w and a known y.
237
675660
2037
換句話說,
11:18
In other words,
238
678124
1151
11:19
you know that it's a bird,
239
679299
1352
電腦知道它是一隻鳥,
11:20
and you already have your neural network
that you've trained on birds,
240
680675
3303
電腦有你給它訓練出來
辨識鳥圖片的神經網路,
但對電腦而言,鳥是怎樣的圖像?
11:24
but what is the picture of a bird?
241
684002
2344
原來,使用一模一樣的
「誤差最小化」程序
11:27
It turns out that by using exactly
the same error-minimization procedure,
242
687034
5024
以及訓練出來
用來辨識鳥的神經網路,
11:32
one can do that with the network
trained to recognize birds,
243
692082
3430
11:35
and the result turns out to be ...
244
695536
3388
你就能辨識出……
11:42
a picture of birds.
245
702400
1305
這是一張鳥圖,
11:44
So this is a picture of birds
generated entirely by a neural network
246
704814
3737
所以,這是一張完全由
訓練辨認鳥的神經網路
自行創造出來的鳥圖,
11:48
that was trained to recognize birds,
247
708575
1826
11:50
just by solving for x
rather than solving for y,
248
710425
3538
只要透過不斷地重複解出 x,
11:53
and doing that iteratively.
249
713987
1288
而不是解 y 就可以了。
11:55
Here's another fun example.
250
715732
1847
這裡有另一個有趣的範例。
11:57
This was a work made
by Mike Tyka in our group,
251
717603
3437
我們團隊裡的
另外一位組員麥克泰卡,
他稱這些畫為《動物大遊行》。
12:01
which he calls "Animal Parade."
252
721064
2308
12:03
It reminds me a little bit
of William Kentridge's artworks,
253
723396
2876
這讓我有點回想起了
威廉肯特基的作品,
12:06
in which he makes sketches, rubs them out,
254
726296
2489
他畫好素描後,擦掉它,
12:08
makes sketches, rubs them out,
255
728809
1460
然後反覆地畫、反覆地擦
12:10
and creates a movie this way.
256
730293
1398
透過這樣的方式,
創造出了一部影片。
12:11
In this case,
257
731715
1151
在這個展示裡,
12:12
what Mike is doing is varying y
over the space of different animals,
258
732890
3277
麥可做的就是把不同動物的 y ,
12:16
in a network designed
to recognize and distinguish
259
736191
2382
透過設計好的神經網路,
12:18
different animals from each other.
260
738597
1810
彼此辨認並分別出不一樣的動物。
12:20
And you get this strange, Escher-like
morph from one animal to another.
261
740431
3751
如此,你就能得到一張像艾雪一樣的
不同動物的變體圖像。
12:26
Here he and Alex together
have tried reducing
262
746221
4614
這一張是他和艾力克斯一起完成的,
12:30
the y's to a space of only two dimensions,
263
750859
2759
他們試著減少 y 的數量,
將這些圖案丟到一個 2D 平面上,
12:33
thereby making a map
out of the space of all things
264
753642
3438
透過這個網路的辨識,
創造出了這一張有各種動物的地圖。
12:37
recognized by this network.
265
757104
1719
12:38
Doing this kind of synthesis
266
758847
2023
要做出這樣的綜合體,
12:40
or generation of imagery
over that entire surface,
267
760894
2382
或透過整張圖面產出圖像,
12:43
varying y over the surface,
you make a kind of map --
268
763300
2846
你只要在圖面上給出各式各樣的 y ,
你就能做出一張地圖來——
一張由神經網路辨識出的視覺地圖。
12:46
a visual map of all the things
the network knows how to recognize.
269
766170
3141
12:49
The animals are all here;
"armadillo" is right in that spot.
270
769335
2865
所有動物都會在這上面,
犰狳就在圖上這個點。
12:52
You can do this with other kinds
of networks as well.
271
772919
2479
你也可以透過不同的神經網路,
做出類似這樣的作品,
12:55
This is a network designed
to recognize faces,
272
775422
2874
這一張由辨識臉的神經網路
12:58
to distinguish one face from another.
273
778320
2000
所做出來的作品,
13:00
And here, we're putting
in a y that says, "me,"
274
780344
3249
這一張是用「我」當作 y ,
所做出來的圖畫,
13:03
my own face parameters.
275
783617
1575
用我的臉當參數。
13:05
And when this thing solves for x,
276
785216
1706
當電腦解出 x 後,
13:06
it generates this rather crazy,
277
786946
2618
它就畫出了這一張相當瘋狂、
13:09
kind of cubist, surreal,
psychedelic picture of me
278
789588
4428
有點像立體派藝術、
超現實、迷幻效果的我,
同一張圖卻有不同的視角。
13:14
from multiple points of view at once.
279
794040
1806
13:15
The reason it looks like
multiple points of view at once
280
795870
2734
而會有這種「同一張圖
不同視角」的感覺,
13:18
is because that network is designed
to get rid of the ambiguity
281
798628
3687
是因為這個神經網路的設計,
13:22
of a face being in one pose
or another pose,
282
802339
2476
可以將不同姿勢臉之間的
模糊地帶移除掉,
13:24
being looked at with one kind of lighting,
another kind of lighting.
283
804839
3376
透過觀察不同的光源就可以做到。
13:28
So when you do
this sort of reconstruction,
284
808239
2085
所以,當你重新製作圖像時,
13:30
if you don't use some sort of guide image
285
810348
2304
如果你沒有使用指導圖,
13:32
or guide statistics,
286
812676
1211
或特定的統計資料,
13:33
then you'll get a sort of confusion
of different points of view,
287
813911
3765
那你就能得到來自
不同角度的混合體圖像,
13:37
because it's ambiguous.
288
817700
1368
因為它是模糊的。
13:39
This is what happens if Alex uses
his own face as a guide image
289
819786
4223
所以如果艾力克斯
用他自己的臉當作指導圖
在優化過程中重新建造我的臉,
就會產生這樣的圖像。
13:44
during that optimization process
to reconstruct my own face.
290
824033
3321
13:48
So you can see it's not perfect.
291
828284
2328
各位可以看到,
這作品還不是很完美,
13:50
There's still quite a lot of work to do
292
830636
1874
在圖像優化的過程方面,
13:52
on how we optimize
that optimization process.
293
832534
2453
還有很多工作要做。
但如果用我的臉當指導圖,
13:55
But you start to get something
more like a coherent face,
294
835011
2827
13:57
rendered using my own face as a guide.
295
837862
2014
就能漸漸地顯現出比較
條理分明的臉。
14:00
You don't have to start
with a blank canvas
296
840892
2501
你不需要從一張空白的畫布
14:03
or with white noise.
297
843417
1156
或用白雜訊畫起。
14:04
When you're solving for x,
298
844597
1304
當你解出 x 後,
14:05
you can begin with an x,
that is itself already some other image.
299
845925
3889
你就可以從 x 開始畫起,
因為它本身就有一些圖像。
14:09
That's what this little demonstration is.
300
849838
2556
這個小小的展示
說明了它的運作原理。
14:12
This is a network
that is designed to categorize
301
852418
4122
這個網路是設計用來
分辨各種不同的物體,
14:16
all sorts of different objects --
man-made structures, animals ...
302
856564
3119
像是人造結構、動物……等。
14:19
Here we're starting
with just a picture of clouds,
303
859707
2593
這一張畫我們是從
雲朵的圖像開始畫起的,
14:22
and as we optimize,
304
862324
1671
當我們把它優化後,
基本上,這個神經網路
正在搞懂它在雲朵中看見了什麼。
14:24
basically, this network is figuring out
what it sees in the clouds.
305
864019
4486
14:28
And the more time
you spend looking at this,
306
868931
2320
當你看得越久,
14:31
the more things you also
will see in the clouds.
307
871275
2753
你就能在雲層中看得越多。
你也可以運用人臉網路
讓它產生幻覺,
14:35
You could also use the face network
to hallucinate into this,
308
875004
3375
14:38
and you get some pretty crazy stuff.
309
878403
1812
然後就會跑出相當瘋狂的畫作。
14:40
(Laughter)
310
880239
1150
(笑聲)
14:42
Or, Mike has done some other experiments
311
882401
2744
或者,麥可已經有作出
一些其它的實驗,
他用那張雲朵的圖像,
14:45
in which he takes that cloud image,
312
885169
3905
使電腦產生幻覺、然後放大、
產生幻覺、再放大。
14:49
hallucinates, zooms, hallucinates,
zooms hallucinates, zooms.
313
889098
3507
14:52
And in this way,
314
892629
1151
用這樣的方式,
14:53
you can get a sort of fugue state
of the network, I suppose,
315
893804
3675
我在想,你就能得到一種
像是在神遊狀態的網路,
14:57
or a sort of free association,
316
897503
3680
或者像是一種無拘束的聯想,
15:01
in which the network
is eating its own tail.
317
901207
2227
彷彿神經網路正在吃著自己的尾巴。
15:03
So every image is now the basis for,
318
903458
3421
所以每一張圖像基本上像是正在想:
15:06
"What do I think I see next?
319
906903
1421
「我接下來會看到什麼?
15:08
What do I think I see next?
What do I think I see next?"
320
908348
2803
接下來會看到什麼?
接下來會看到什麼?」
15:11
I showed this for the first time in public
321
911487
2936
我第一次在一個
公眾場合上展示這個影片,
15:14
to a group at a lecture in Seattle
called "Higher Education" --
322
914447
5437
是在西雅圖的「高等教育」
機構做演說時展示的,
15:19
this was right after
marijuana was legalized.
323
919908
2437
當時剛好是大麻剛合法化的時候。
15:22
(Laughter)
324
922369
2415
(笑聲)
15:26
So I'd like to finish up quickly
325
926627
2104
所以,我快速總結一下,
15:28
by just noting that this technology
is not constrained.
326
928755
4255
這項技術並不會受到約束。
我剛剛展示的是純粹的視覺範例,
因為觀察它的變化,真的很好玩。
15:33
I've shown you purely visual examples
because they're really fun to look at.
327
933034
3665
15:36
It's not a purely visual technology.
328
936723
2451
它不單只有視覺科技。
15:39
Our artist collaborator, Ross Goodwin,
329
939198
1993
我們的藝術合作者,羅斯谷穎
已經做了一些實驗,
15:41
has done experiments involving
a camera that takes a picture,
330
941215
3671
他用相機拍了一張照片,
15:44
and then a computer in his backpack
writes a poem using neural networks,
331
944910
4234
然後他背包裡的電腦
會根據圖片上的內容,
透過神經網路,創作出一首詩。
15:49
based on the contents of the image.
332
949168
1944
這個會作詩的神經網路
15:51
And that poetry neural network
has been trained
333
951136
2947
是透過大量 20 世紀的詩集
所訓練出來的,
15:54
on a large corpus of 20th-century poetry.
334
954107
2234
15:56
And the poetry is, you know,
335
956365
1499
而做出來的詩,
15:57
I think, kind of not bad, actually.
336
957888
1914
實際上,我覺得還得不錯。
15:59
(Laughter)
337
959826
1384
(笑聲)
16:01
In closing,
338
961234
1159
整體而言,
16:02
I think that per Michelangelo,
339
962417
2132
我在想,米開朗基羅,
16:04
I think he was right;
340
964573
1234
他是對的;
16:05
perception and creativity
are very intimately connected.
341
965831
3436
感知和創意的關係是相當緊密的。
16:09
What we've just seen are neural networks
342
969611
2634
我們剛剛看的神經網路,
16:12
that are entirely trained to discriminate,
343
972269
2303
它們是被訓練出來分辯
16:14
or to recognize different
things in the world,
344
974596
2242
或辨認世界上不同的東西,
16:16
able to be run in reverse, to generate.
345
976862
3161
也可以反過來,自行創作出東西來。
而我從中所得到的
16:20
One of the things that suggests to me
346
980047
1783
16:21
is not only that
Michelangelo really did see
347
981854
2398
不僅有米開朗基羅的啟發:
16:24
the sculpture in the blocks of stone,
348
984276
2452
「看見石頭裡的雕像」,
16:26
but that any creature,
any being, any alien
349
986752
3638
還有任何能做出感知活動的
生物、生命、外來物種
16:30
that is able to do
perceptual acts of that sort
350
990414
3657
都能透過這樣的方式
被呈現並創造出來,
16:34
is also able to create
351
994095
1375
16:35
because it's exactly the same
machinery that's used in both cases.
352
995494
3224
因為這兩者與剛才舉的例子
都有著相同的機制。
16:38
Also, I think that perception
and creativity are by no means
353
998742
4532
我也認為,感知及創意
不是只有我們人類獨有。
16:43
uniquely human.
354
1003298
1210
16:44
We start to have computer models
that can do exactly these sorts of things.
355
1004532
3708
我們已經有電腦模式
可以做出相當類似的事。
16:48
And that ought to be unsurprising;
the brain is computational.
356
1008264
3328
所以不需要感到驚訝;
因為大腦是會運算的。
16:51
And finally,
357
1011616
1657
最後,我要說的是,
16:53
computing began as an exercise
in designing intelligent machinery.
358
1013297
4668
設計智能機器已經開始成為
電腦界的活動。
16:57
It was very much modeled after the idea
359
1017989
2462
在如何讓機器更智能的領域方面,
17:00
of how could we make machines intelligent.
360
1020475
3013
已經有很多的模式產生。
17:03
And we finally are starting to fulfill now
361
1023512
2162
我們終於開始
17:05
some of the promises
of those early pioneers,
362
1025698
2406
完成一些早期前輩們
像是圖靈、馮諾伊曼、
17:08
of Turing and von Neumann
363
1028128
1713
17:09
and McCulloch and Pitts.
364
1029865
2265
馬庫洛奇和皮斯的期望。
而我也認為電腦不是只有拿來計算
17:12
And I think that computing
is not just about accounting
365
1032154
4098
17:16
or playing Candy Crush or something.
366
1036276
2147
或玩玩 Candy Crush 而已,
17:18
From the beginning,
we modeled them after our minds.
367
1038447
2578
回到初衷,我們想要的
是讓電腦能仿效人腦。
它不僅讓我們更了解了人類的心智,
17:21
And they give us both the ability
to understand our own minds better
368
1041049
3269
並讓我們獲得延伸發展心智的能力。
17:24
and to extend them.
369
1044342
1529
17:26
Thank you very much.
370
1046627
1167
非常感謝大家。
17:27
(Applause)
371
1047818
5939
(掌聲)
New videos
Original video on YouTube.com
關於本網站
本網站將向您介紹對學習英語有用的 YouTube 視頻。 您將看到來自世界各地的一流教師教授的英語課程。 雙擊每個視頻頁面上顯示的英文字幕,從那裡播放視頻。 字幕與視頻播放同步滾動。 如果您有任何意見或要求,請使用此聯繫表與我們聯繫。