How to Get Inside the "Brain" of AI | Alona Fyshe | TED

60,071 views ・ 2023-04-03

TED


Please double-click on the English subtitles below to play the video.

00:04
People are funny.
0
4417
1127
00:05
We're constantly trying to understand and interpret
1
5585
2544
00:08
the world around us.
2
8171
1460
00:10
I live in a house with two black cats, and let me tell you,
3
10090
2794
00:12
every time I see a black, bunched up sweater out of the corner of my eye,
4
12926
3462
00:16
I think it's a cat.
5
16388
1209
00:18
It's not just the things we see.
6
18098
1585
00:19
Sometimes we attribute more intelligence than might actually be there.
7
19724
3587
00:23
Maybe you've seen the dogs on TikTok.
8
23770
1794
00:25
They have these little buttons that say things like "walk" or "treat."
9
25605
3379
00:29
They can push them to communicate some things with their owners,
10
29025
3128
00:32
and their owners think they use them
11
32195
1752
00:33
to communicate some pretty impressive things.
12
33947
2169
00:36
But do the dogs know what they're saying?
13
36491
2002
00:39
Or perhaps you've heard the story of Clever Hans the horse,
14
39494
3462
00:42
and he could do math.
15
42956
1376
00:44
And not just like, simple math problems, really complicated ones,
16
44374
3086
00:47
like, if the eighth day of the month falls on a Tuesday,
17
47460
3003
00:50
what's the date of the following Friday?
18
50463
1919
00:52
It's like, pretty impressive for a horse.
19
52841
2043
00:55
Unfortunately, Hans wasn't doing math,
20
55927
2586
00:58
but what he was doing was equally impressive.
21
58555
2586
01:01
Hans had learned to watch the people in the room
22
61182
2670
01:03
to tell when he should tap his hoof.
23
63893
2002
01:05
So he communicated his answers by tapping his hoof.
24
65937
2461
01:08
It turns out that if you know the answer
25
68982
1918
01:10
to "if the eighth day of the month falls on a Tuesday,
26
70942
2544
01:13
what's the date of the following Friday,"
27
73528
1960
01:15
you will subconsciously change your posture
28
75488
2044
01:17
once the horse has given the correct 18 taps.
29
77532
2169
01:20
So Hans couldn't do math,
30
80493
1293
01:21
but he had learned to watch the people in the room who could do math,
31
81786
3254
01:25
which, I mean, still pretty impressive for a horse.
32
85081
2419
01:28
But this is an old picture,
33
88418
1418
01:29
and we would not fall for Clever Hans today.
34
89836
2252
01:32
Or would we?
35
92672
1168
01:34
Well, I work in AI,
36
94549
2169
01:36
and let me tell you, things are wild.
37
96718
2002
01:38
There have been multiple examples of people being completely convinced
38
98720
3754
01:42
that AI understands them.
39
102515
1502
01:44
In 2022,
40
104559
2711
01:47
a Google engineer thought that Google’s AI was sentient.
41
107270
3337
01:50
And you may have had a really human-like conversation
42
110649
2752
01:53
with something like ChatGPT.
43
113401
1877
01:55
But models we're training today are so much better
44
115779
2377
01:58
than the models we had even five years ago.
45
118198
2002
02:00
It really is remarkable.
46
120200
1710
02:02
So at this super crazy moment in time,
47
122744
2502
02:05
let’s ask the super crazy question:
48
125288
2127
02:07
Does AI understand us,
49
127832
1293
02:09
or are we having our own Clever Hans moment?
50
129167
2878
02:13
Some philosophers think that computers will never understand language.
51
133254
3462
02:16
To illustrate this, they developed something they call
52
136716
2628
02:19
the Chinese room argument.
53
139344
1460
02:21
In the Chinese room, there is a person, hypothetical person,
54
141429
3629
02:25
who does not understand Chinese,
55
145100
1835
02:26
but he has along with him a set of instructions
56
146976
2211
02:29
that tell him how to respond in Chinese to any Chinese sentence.
57
149229
4212
02:33
Here's how the Chinese room works.
58
153983
1669
02:35
A piece of paper comes in through a slot in the door,
59
155652
2502
02:38
has something written in Chinese on it.
60
158196
2461
02:40
The person uses their instructions to figure out how to respond.
61
160699
3003
02:43
They write the response down on a piece of paper
62
163702
2293
02:46
and then send it back out through the door.
63
166037
2252
02:48
To somebody who speaks Chinese,
64
168331
1502
02:49
standing outside this room,
65
169833
1293
02:51
it might seem like the person inside the room speaks Chinese.
66
171167
3170
02:54
But we know they do not,
67
174337
2753
02:57
because no knowledge of Chinese is required to follow the instructions.
68
177090
4546
03:01
Performance on this task does not show that you know Chinese.
69
181636
2961
03:05
So what does that tell us about AI?
70
185807
2002
03:08
Well, when you and I stand outside of the room,
71
188226
2753
03:11
when we speak to one of these AIs like ChatGPT,
72
191020
4547
03:15
we are the person standing outside the room.
73
195567
2085
03:17
We're feeding in English sentences,
74
197652
1710
03:19
we're getting English sentences back.
75
199362
2127
03:21
It really looks like the models understand us.
76
201531
2419
03:23
It really looks like they know English.
77
203992
2502
03:27
But under the hood,
78
207203
1168
03:28
these models are just following a set of instructions, albeit complex.
79
208371
3754
03:32
How do we know if AI understands us?
80
212917
2795
03:36
To answer that question, let's go back to the Chinese room again.
81
216880
3086
03:39
Let's say we have two Chinese rooms.
82
219966
1794
03:41
In one Chinese room is somebody who actually speaks Chinese,
83
221801
3879
03:46
and in the other room is our impostor.
84
226014
1877
03:48
When the person who actually speaks Chinese gets a piece of paper
85
228224
3087
03:51
that says something in Chinese in it, they can read it, no problem.
86
231311
3170
03:54
But when our imposter gets it again,
87
234522
1752
03:56
he has to use his set of instructions to figure out how to respond.
88
236274
3170
03:59
From the outside, it might be impossible to distinguish these two rooms,
89
239819
3671
04:03
but we know inside something really different is happening.
90
243531
3587
04:07
To illustrate that,
91
247660
1168
04:08
let's say inside the minds of our two people,
92
248870
2836
04:11
inside of our two rooms,
93
251748
1585
04:13
is a little scratch pad.
94
253374
1752
04:15
And everything they have to remember in order to do this task
95
255168
2878
04:18
has to be written on that little scratch pad.
96
258046
2169
04:20
If we could see what was written on that scratch pad,
97
260757
2586
04:23
we would be able to tell how different their approach to the task is.
98
263384
3421
04:27
So though the input and the output of these two rooms
99
267514
2502
04:30
might be exactly the same,
100
270016
1293
04:31
the process of getting from input to output -- completely different.
101
271309
3337
04:35
So again, what does that tell us about AI?
102
275897
2294
04:38
Again, if AI, even if it generates completely plausible dialogue,
103
278691
3838
04:42
answers questions just like we would expect,
104
282529
2085
04:44
it may still be an imposter of sorts.
105
284614
2377
04:47
If we want to know if AI understands language like we do,
106
287033
3003
04:50
we need to know what it's doing.
107
290078
1835
04:51
We need to get inside to see what it's doing.
108
291955
2335
04:54
Is it an imposter or not?
109
294332
1585
04:55
We need to see its scratch pad,
110
295959
1835
04:57
and we need to be able to compare it
111
297794
1752
04:59
to the scratch pad of somebody who actually understands language.
112
299587
3212
05:02
But like scratch pads in brains,
113
302841
1751
05:04
that's not something we can actually see, right?
114
304634
2336
05:07
Well, it turns out that we can kind of see scratch pads in brains.
115
307720
3337
05:11
Using something like fMRI or EEG,
116
311099
2127
05:13
we can take what are like little snapshots of the brain while it’s reading.
117
313268
3712
05:17
So have people read words or stories and then take pictures of their brain.
118
317021
4171
05:21
And those brain images are like fuzzy,
119
321192
2252
05:23
out-of-focus pictures of the scratch pad of the brain.
120
323486
3253
05:26
They tell us a little bit about how the brain is processing
121
326739
3045
05:29
and representing information while you read.
122
329784
2461
05:33
So here are three brain images taken while a person read the word "apartment,"
123
333079
4087
05:37
"house" and "celery."
124
337208
1669
05:39
You can see just with your naked eye
125
339252
2002
05:41
that the brain image for "apartment" and "house"
126
341296
2252
05:43
are more similar to each other
127
343548
1585
05:45
than they are to the brain image for "celery."
128
345133
2252
05:47
And you know, of course that apartments and houses are more similar
129
347385
3170
05:50
than they are to celery, just the words.
130
350555
2210
05:52
So said another way,
131
352807
2544
05:55
the brain uses its scratchpad when reading the words "apartment" and "house"
132
355393
4296
05:59
in a way that's more similar than when you read the word "celery."
133
359731
3128
06:03
The scratch pad tells us a little bit
134
363693
1793
06:05
about how the brain represents the language.
135
365486
2086
06:07
It's not a perfect picture of what the brain's doing,
136
367572
2502
06:10
but it's good enough.
137
370074
1335
06:11
OK, so we have scratch pads for the brain.
138
371409
2169
06:13
Now we need a scratch pad for AI.
139
373620
2127
06:16
So inside a lot of AIs is a neural network.
140
376706
2878
06:19
And inside of a neural network is a bunch of these little neurons.
141
379626
3253
06:22
So here the neurons are like these little gray circles.
142
382921
2919
06:25
And we would like to know
143
385840
1210
06:27
what is the scratch pad of a neural network?
144
387091
2086
06:29
Well, when we feed in a word into a neural network,
145
389177
4254
06:33
each of the little neurons computes a number.
146
393473
2794
06:36
Those little numbers I'm representing here with colors.
147
396893
2627
06:39
So every neuron computes this little number,
148
399562
2795
06:42
and those numbers tell us something
149
402398
1710
06:44
about how the neural network is processing language.
150
404108
2711
06:47
Taken together,
151
407862
1168
06:49
all of those little circles paint us a picture
152
409030
2753
06:51
of how the neural network is representing language,
153
411824
2419
06:54
and they give us the scratch pad of the neural network.
154
414243
2753
06:57
OK, great.
155
417580
1168
06:58
Now we have two scratch pads, one from the brain and one from AI.
156
418790
3086
07:01
And we want to know: Is AI doing something like what the brain is doing?
157
421876
3629
07:05
How can we test that?
158
425838
1377
07:07
Here's what researchers have come up with.
159
427757
2002
07:09
We're going to train a new model.
160
429801
1877
07:11
That new model is going to look at neural network scratch pad
161
431678
2877
07:14
for a particular word
162
434555
1168
07:15
and try to predict the brain scratch pad for the same word.
163
435723
3087
07:18
We can do it, by the way, around two.
164
438851
1919
07:20
So let's train a new model.
165
440812
2043
07:22
It’s going to look at the neural network scratch pad for a particular word
166
442897
3504
07:26
and try to predict the brain scratchpad.
167
446401
1918
07:28
If the brain and AI are doing nothing alike,
168
448319
2586
07:30
have nothing in common,
169
450947
1168
07:32
we won't be able to do this prediction task.
170
452115
2085
07:34
It won't be possible to predict one from the other.
171
454200
2461
07:36
So we've reached a fork in the road
172
456995
1710
07:38
and you can probably tell I'm about to tell you one of two things.
173
458705
3295
07:42
I’m going to tell you AI is amazing,
174
462458
2044
07:44
or I'm going to tell you AI is an imposter.
175
464544
2669
07:48
Researchers like me love to remind you
176
468047
2211
07:50
that AI is nothing like the brain.
177
470299
1627
07:51
And that is true.
178
471968
1501
07:53
But could it also be the AI and the brain share something in common?
179
473845
3253
07:58
So we’ve done this scratch pad prediction task,
180
478516
2211
08:00
and it turns out, 75 percent of the time
181
480727
2669
08:03
the predicted neural network scratchpad for a particular word
182
483438
3169
08:06
is more similar to the true neural network scratchpad for that word
183
486649
3754
08:10
than it is to the neural network scratch pad
184
490403
2085
08:12
for some other randomly chosen word --
185
492488
1835
08:14
75 percent is much better than chance.
186
494323
2962
08:17
What about for more complicated things,
187
497285
1918
08:19
not just words, but sentences, even stories?
188
499203
2086
08:21
Again, this scratch pad prediction task works.
189
501330
2420
08:23
We’re able to predict the neural network scratch pad from the brain and vice versa.
190
503791
4213
08:28
Amazing.
191
508838
1210
08:30
So does that mean
192
510548
1168
08:31
that neural networks and AI understand language just like we do?
193
511758
3378
08:35
Well, truthfully, no.
194
515511
1585
08:37
Though these scratch pad prediction tasks show above-chance accuracy,
195
517513
4880
08:42
the underlying correlations are still pretty weak.
196
522393
2378
08:45
And though neural networks are inspired by the brain,
197
525188
2502
08:47
they don't have the same kind of structure and complexity
198
527690
2711
08:50
that we see in the brain.
199
530401
1210
08:52
Neural networks also don't exist in the world.
200
532028
2461
08:54
A neural network has never opened a door
201
534530
2086
08:56
or seen a sunset, heard a baby cry.
202
536657
2962
09:00
Can a neural network that doesn't actually exist in the world,
203
540161
2961
09:03
hasn't really experienced the world,
204
543122
1794
09:04
really understand language about the world?
205
544916
2294
09:08
Still, these scratch pad prediction experiments have held up --
206
548169
2961
09:11
multiple brain imaging experiments,
207
551130
1710
09:12
multiple neural networks.
208
552840
1418
09:14
We've also found that as the neural networks get more accurate,
209
554300
2961
09:17
they also start to use their scratch pad in a way
210
557261
2711
09:20
that becomes more brain-like.
211
560014
2002
09:22
And it's not just language.
212
562058
1293
09:23
We've seen similar results in navigation and vision.
213
563392
2503
09:26
So AI is not doing exactly what the brain is doing,
214
566270
3879
09:30
but it's not completely random either.
215
570191
2377
09:34
So from where I sit,
216
574112
1835
09:35
if we want to know if AI really understands language like we do,
217
575988
3337
09:39
we need to get inside of the Chinese room.
218
579325
2336
09:41
We need to know what the AI is doing,
219
581702
1836
09:43
and we need to be able to compare that to what people are doing
220
583579
2962
09:46
when they understand language.
221
586541
1751
09:48
AI is moving so fast.
222
588334
1877
09:50
Today, I'm asking you, does AI understand language
223
590545
2377
09:52
that might seem like a silly question in ten years.
224
592964
2460
09:55
Or ten months.
225
595466
1168
09:56
(Laughter)
226
596634
1919
09:58
But one thing will remain true.
227
598594
1544
10:00
We are meaning-making humans,
228
600179
1460
10:01
and we are going to continue to look for meaning
229
601681
2878
10:04
and interpret the world around us.
230
604559
1918
10:06
And we will need to remember
231
606477
2127
10:08
that if we only look at the input and output of AI,
232
608646
2794
10:11
it's very easy to be fooled.
233
611440
2128
10:13
We need to get inside of the metaphorical room of AI
234
613609
3712
10:17
in order to see what's happening.
235
617363
1627
10:19
It's what's inside the counts.
236
619740
2461
10:22
Thank you.
237
622660
1168
10:23
(Applause)
238
623828
4004
About this website

This site will introduce you to YouTube videos that are useful for learning English. You will see English lessons taught by top-notch teachers from around the world. Double-click on the English subtitles displayed on each video page to play the video from there. The subtitles scroll in sync with the video playback. If you have any comments or requests, please contact us using this contact form.

https://forms.gle/WvT1wiN1qDtmnspy7