This computer is learning to read your mind | DIY Neuroscience, a TED series

126,401 views ・ 2018-09-15

TED


请双击下面的英文字幕来播放视频。

00:00
Translator: Joseph Geni Reviewer: Krystian Aparta
0
0
7000
翻译人员: yiwen zhang 校对人员: Wanting Zhong
00:12
Greg Gage: Mind-reading. You've seen this in sci-fi movies:
1
12203
2859
格雷戈 · 盖奇(Greg Gage):读心术。 你在科幻电影中曾经看到过:
00:15
machines that can read our thoughts.
2
15086
1857
那是可以读出我们想法的机器。
00:16
However, there are devices today
3
16967
1798
然而,如今有很多机器
00:18
that can read the electrical activity from our brains.
4
18789
2524
可以读出我们大脑中的电波。
00:21
We call this the EEG.
5
21337
1272
我们把它叫做 “EEG”。
00:23
Is there information contained in these brainwaves?
6
23695
2829
这些脑电波中含有信息吗?
00:26
And if so, could we train a computer to read our thoughts?
7
26548
2813
如果含有信息,我们可以训练 计算机读懂我们的思想吗?
00:29
My buddy Nathan has been working to hack the EEG
8
29385
2904
我的好友内森一直 致力于研究如何破译 EEG
00:32
to build a mind-reading machine.
9
32313
1676
以建造一台可以读心的机器。
00:34
[DIY Neuroscience]
10
34013
2457
【DIY 神经科学】
00:36
So this is how the EEG works.
11
36939
1561
这就是 EEG 的工作原理。
00:38
Inside your head is a brain,
12
38524
1844
在你的脑袋里有一个大脑,
00:40
and that brain is made out of billions of neurons.
13
40392
2558
大脑是由数十亿个神经元构成的,
00:42
Each of those neurons sends an electrical message to each other.
14
42974
3007
每个神经元都在 互相传送电子信息,
00:46
These small messages can combine to make an electrical wave
15
46005
2814
这些微小的信息可以结合在一起
00:48
that we can detect on a monitor.
16
48843
1570
形成我们能在显示器上 探测到的电波。
00:50
Now traditionally, the EEG can tell us large-scale things,
17
50437
2724
传统意义上而言, EEG 能告诉我们大维度的事情,
00:53
for example if you're asleep or if you're alert.
18
53185
2350
例如你是睡着还是清醒。
00:55
But can it tell us anything else?
19
55559
1594
但它可以告诉我们其它事情吗?
00:57
Can it actually read our thoughts?
20
57177
1702
它是否能够读出我们心中所想?
00:58
We're going to test this,
21
58903
1219
我们要去测试这一点,
01:00
and we're not going to start with some complex thoughts.
22
60146
2623
而我们不打算从一些 复杂的想法开始。
01:02
We're going to do something very simple.
23
62793
1977
我们打算做一件非常简单的事情。
01:04
Can we interpret what someone is seeing using only their brainwaves?
24
64794
3253
我们可以仅仅依据脑电波 判读出一个人看到了什么吗?
01:08
Nathan's going to begin by placing electrodes on Christy's head.
25
68071
3000
内森先要在克里斯蒂的头上安装电极。
01:11
Nathan: My life is tangled.
26
71095
1523
内:我的人生一团糟。
01:12
(Laughter)
27
72642
1150
(笑声)
01:14
GG: And then he's going to show her a bunch of pictures
28
74152
2584
格:之后他会给她看一些图片,
01:16
from four different categories.
29
76760
1521
这些图片出自四种不同类别:
01:18
Nathan: Face, house, scenery and weird pictures.
30
78305
2654
内森:面孔,房子, 风景和古怪的图片。
01:20
GG: As we show Christy hundreds of these images,
31
80983
2498
格:当我们向克里斯蒂 展示数百张这种图片时,
01:23
we are also capturing the electrical waves onto Nathan's computer.
32
83505
3543
我们也在内森的电脑上 捕捉她的脑电波。
01:27
We want to see if we can detect any visual information about the photos
33
87072
3386
我们想知道我们是否 能通过这些脑电波,
探测到任何与这些 图片相关的视觉信息。
01:30
contained in the brainwaves,
34
90482
1352
01:31
so when we're done, we're going to see if the EEG
35
91858
2331
在实验结束后, 我们将会看到 EEG 是否
01:34
can tell us what kind of picture Christy is looking at,
36
94213
2598
可以告诉我们克里斯蒂 在看哪种图片。
01:36
and if it does, each category should trigger a different brain signal.
37
96835
3584
如果可以,不同类别的图片 应该会触发不同的大脑信号。
01:40
OK, so we collected all the raw EEG data,
38
100443
2628
好的,我们收集完了 所有的原始 EEG 数据,
01:43
and this is what we got.
39
103095
1150
这就是我收集到的样子。
01:45
It all looks pretty messy, so let's arrange them by picture.
40
105389
2938
它看上去很混乱,于是我们 根据图片类别将它们排序。
01:48
Now, still a bit too noisy to see any differences,
41
108826
2656
现在,还是有点太嘈杂, 无法看出任何区别,
01:51
but if we average the EEG across all image types
42
111506
3040
但是如果我们根据图片 出现的时间将信号对齐,
01:54
by aligning them to when the image first appeared,
43
114570
2436
并对每种图片类别的 EEG 取平均值,
01:57
we can remove this noise,
44
117030
1617
我们就能移除其中的噪声。
01:58
and pretty soon, we can see some dominant patterns
45
118671
2334
很快,我们就可以从各个类别中
02:01
emerge for each category.
46
121029
1564
看到一些明显的规律。
02:02
Now the signals all still look pretty similar.
47
122617
2156
现在这些信号看起来还是很相似,
02:04
Let's take a closer look.
48
124797
1215
让我们再仔细看看。
02:06
About a hundred milliseconds after the image comes on,
49
126036
2525
大约在一张图片 出现后的一百毫秒后,
02:08
we see a positive bump in all four cases,
50
128585
2628
我们在四个类别中 都看到了正向波动,
02:11
and we call this the P100, and what we think that is
51
131237
2789
我们把它叫作 P100 我们认为这是
02:14
is what happens in your brain when you recognize an object.
52
134050
3075
当你识别物体时 大脑中发生的活动。
02:17
But damn, look at that signal for the face.
53
137149
2086
但是见鬼,看看“面孔“ 图片对应的信号,
02:19
It looks different than the others.
54
139259
1711
它看起来与众不同,
02:20
There's a negative dip about 170 milliseconds
55
140994
2890
在图片出现后的约 170 毫秒时,
02:23
after the image comes on.
56
143908
1540
出现了负向波动。
02:25
What could be going on here?
57
145472
1750
这里可能发生了什么?
02:27
Research shows that our brain has a lot of neurons that are dedicated
58
147246
3240
研究显示,我们大脑有大量神经元
专门负责识别人类的面孔,
02:30
to recognizing human faces,
59
150510
1459
02:31
so this N170 spike could be all those neurons
60
151993
2844
所以这个 N170 负波可能是
02:34
firing at once in the same location,
61
154861
1985
所有这些神经元 在同一地方同时激活,
02:36
and we can detect that in the EEG.
62
156870
1634
而我们可以在 EEG 中探测到。
02:39
So there are two takeaways here.
63
159083
1820
于是从中得出两个结论,
02:40
One, our eyes can't really detect the differences in patterns
64
160927
3085
第一,在没有经过平均化降噪时,
我们的眼睛并不能真的 识别脑波规律的不同;
02:44
without averaging out the noise,
65
164036
1571
02:45
and two, even after removing the noise,
66
165631
2237
第二,即使移除噪声后,
02:47
our eyes can only pick up the signals associated with faces.
67
167892
3001
我们的眼睛也只能 识别出和面孔有关的信号。
02:50
So this is where we turn to machine learning.
68
170917
2268
于是我们在此转而借助机器学习。
02:53
Now, our eyes are not very good at picking up patterns in noisy data,
69
173209
3976
我们的眼睛并不擅长 在嘈杂的数据中发现规律,
02:57
but machine learning algorithms are designed to do just that,
70
177209
2946
但是机器学习算法的设计 初衷就是解决这类问题。
03:00
so could we take a lot of pictures and a lot of data
71
180179
3201
所以我们可以将许多图片和数据
03:03
and feed it in and train a computer
72
183404
1790
输入到电脑中进行训练,
03:05
to be able to interpret what Christy is looking at in real time?
73
185218
3381
从而实时判断克里斯蒂正在看什么。
03:09
We're trying to code the information that's coming out of her EEG
74
189088
4117
我们尝试将她的 EEG 信息
进行实时编码,
03:13
in real time
75
193229
1175
03:14
and predict what it is that her eyes are looking at.
76
194428
2461
并预测她眼睛在看的东西。
03:16
And if it works, what we should see
77
196913
1727
如果这样有效,我们应该能看到
03:18
is every time that she gets a picture of scenery,
78
198664
2381
每次她看到风景的图片时,
03:21
it should say scenery, scenery, scenery, scenery.
79
201069
2286
机器应该显示风景, 风景,风景,风景。
03:23
A face -- face, face, face, face,
80
203379
1957
如果她看到面孔,机器则显示 面孔,面孔,面孔,面孔,
03:25
but it's not quite working that way, is what we're discovering.
81
205360
3531
但是我们发现, 实际上并非如此。
03:33
(Laughter)
82
213385
3548
(笑声)
03:36
OK.
83
216957
1151
好的。
03:38
Director: So what's going on here? GG: We need a new career, I think.
84
218132
3382
导演:所以发生了什么? 格:我觉得我们应该转行。
03:41
(Laughter)
85
221538
1070
(笑声)
03:42
OK, so that was a massive failure.
86
222632
2444
好吧,所以刚刚那是个重大失败。
03:45
But we're still curious: How far could we push this technology?
87
225100
3212
但是我们依然好奇: 我们能这项技术发展到多深?
03:48
And we looked back at what we did.
88
228336
1640
于是我们回顾了我们的做法。
03:50
We noticed that the data was coming into our computer very quickly,
89
230000
3143
我们发现电脑在飞快地获取数据,
03:53
without any timing of when the images came on,
90
233167
2241
但没有对图片出现的时间进行计时,
03:55
and that's the equivalent of reading a very long sentence
91
235432
2876
这等同于读一个
在单词间没有空格的长句。
03:58
without spaces between the words.
92
238332
1605
03:59
It would be hard to read,
93
239961
1438
这样的句子很难读懂,
04:01
but once we add the spaces, individual words appear
94
241423
3713
不过一旦我们添加了空格, 我们就能看到独立的单词,
04:05
and it becomes a lot more understandable.
95
245160
2044
句子也就变得容易理解多了,
04:07
But what if we cheat a little bit?
96
247228
1847
但如果我们做一点弊呢?
04:09
By using a sensor, we can tell the computer when the image first appears.
97
249099
3537
通过使用传感器, 我们能告诉电脑每张图片出现的时机。
04:12
That way, the brainwave stops being a continuous stream of information,
98
252660
3602
这样,脑波就不再是 一个没有间断的信息流,
04:16
and instead becomes individual packets of meaning.
99
256286
2711
而是变成了一个个 有意义的信息小包裹。
04:19
Also, we're going to cheat a little bit more,
100
259021
2368
另外,我们还要再做一点弊,
04:21
by limiting the categories to two.
101
261413
1812
把图片限制到两个类别。
04:23
Let's see if we can do some real-time mind-reading.
102
263249
2383
让我们看看我们是否 能够进行实时读心。
04:25
In this new experiment,
103
265656
1235
在这个新实验中,
04:26
we're going to constrict it a little bit more
104
266915
2097
我们将限制实验条件:
04:29
so that we know the onset of the image
105
269036
2252
我们会知道图片出现的时间,
04:31
and we're going to limit the categories to "face" or "scenery."
106
271312
3382
并将类别限制为 "面孔” 或 “风景” 。
04:35
Nathan: Face. Correct.
107
275097
1511
内:面孔。正确。
04:37
Scenery. Correct.
108
277780
1351
风景。正确。
04:40
GG: So right now, every time the image comes on,
109
280251
2373
格:所以现在,每当图片出现时,
04:42
we're taking a picture of the onset of the image
110
282648
2266
我们对图片出现的时刻进行记录,
04:44
and decoding the EEG.
111
284938
1695
并对 EEG 解码。
04:46
It's getting correct.
112
286657
1256
它变得越来越正确。
04:47
Nathan: Yes. Face. Correct.
113
287937
1579
内:是的。面孔。正确。
04:49
GG: So there is information in the EEG signal, which is cool.
114
289540
2859
格:所以 EEG 的信号中 包含信息,这很棒。
04:52
We just had to align it to the onset of the image.
115
292423
2537
我们仅仅需要把它 和图片出现的时刻对齐。
04:55
Nathan: Scenery. Correct.
116
295307
1311
内:风景。正确。
04:59
Face. Yeah.
117
299344
1150
面孔。没错。
05:00
GG: This means there is some information there,
118
300518
2288
格:这意味着它包含了一些信息,
05:02
so if we know at what time the picture came on,
119
302830
2913
如果我们知道图片出现的时间,
05:05
we can tell what type of picture it was,
120
305767
1999
我们就有可能根据 这些由图片诱发的电位
05:07
possibly, at least on average, by looking at these evoked potentials.
121
307790
5096
判断它是哪个类别的图片, 至少一般可以做到。
05:12
Nathan: Exactly.
122
312910
1325
内:说得没错。
05:14
GG: If you had told me at the beginning of this project this was possible,
123
314259
3521
格:如果你一开始跟我说, 这个项目有可能实现,
05:17
I would have said no way.
124
317804
1251
我会说 “怎么可能” 。
05:19
I literally did not think we could do this.
125
319079
2000
我真的觉得我们不可能做到。
05:21
Did our mind-reading experiment really work?
126
321103
2066
我们的读心术实验 真的成功了吗?
05:23
Yes, but we had to do a lot of cheating.
127
323193
1975
成功了,但是我们必须做很多弊。
05:25
It turns out you can find some interesting things in the EEG,
128
325192
2905
结果就是,你能通过 EEG 发现一些有趣的事,
05:28
for example if you're looking at someone's face,
129
328121
2290
比如,你是否在看某人的脸,
05:30
but it does have a lot of limitations.
130
330435
2157
但它确实有很多限制。
05:32
Perhaps advances in machine learning will make huge strides,
131
332616
2946
也许机器学习领域的进步 会带来重大突破。
05:35
and one day we will be able to decode what's going on in our thoughts.
132
335586
3390
有朝一日,我们能够解码心中所想。
05:39
But for now, the next time a company says that they can harness your brainwaves
133
339000
4077
可是现在来说, 当一个公司说它能利用你的脑波
05:43
to be able to control devices,
134
343101
1750
来控制一些设备,
05:44
it is your right, it is your duty to be skeptical.
135
344875
3310
你有权利和义务对此保持怀疑。
【格雷戈 · 盖奇内森 · YH · 权】
关于本网站

这个网站将向你介绍对学习英语有用的YouTube视频。你将看到来自世界各地的一流教师教授的英语课程。双击每个视频页面上显示的英文字幕,即可从那里播放视频。字幕会随着视频的播放而同步滚动。如果你有任何意见或要求,请使用此联系表与我们联系。

https://forms.gle/WvT1wiN1qDtmnspy7