How we can teach computers to make sense of our emotions | Raphael Arar

64,579 views ・ 2018-04-24

TED


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

00:00
Translator: Ivana Korom Reviewer: Joanna Pietrulewicz
0
0
7000
00:13
I consider myself one part artist and one part designer.
1
13760
4680
00:18
And I work at an artificial intelligence research lab.
2
18480
3160
00:22
We're trying to create technology
3
22720
1696
00:24
that you'll want to interact with in the far future.
4
24440
3296
00:27
Not just six months from now, but try years and decades from now.
5
27760
4640
00:33
And we're taking a moonshot
6
33120
1616
00:34
that we'll want to be interacting with computers
7
34760
2456
00:37
in deeply emotional ways.
8
37240
2120
00:40
So in order to do that,
9
40280
1456
00:41
the technology has to be just as much human as it is artificial.
10
41760
4480
00:46
It has to get you.
11
46920
2256
00:49
You know, like that inside joke that'll have you and your best friend
12
49200
3336
00:52
on the floor, cracking up.
13
52560
1936
00:54
Or that look of disappointment that you can just smell from miles away.
14
54520
4560
01:00
I view art as the gateway to help us bridge this gap between human and machine:
15
60560
6040
01:07
to figure out what it means to get each other
16
67280
3136
01:10
so that we can train AI to get us.
17
70440
2760
01:13
See, to me, art is a way to put tangible experiences
18
73920
3816
01:17
to intangible ideas, feelings and emotions.
19
77760
3240
01:21
And I think it's one of the most human things about us.
20
81800
2600
01:25
See, we're a complicated and complex bunch.
21
85480
2936
01:28
We have what feels like an infinite range of emotions,
22
88440
3136
01:31
and to top it off, we're all different.
23
91600
2496
01:34
We have different family backgrounds,
24
94120
2296
01:36
different experiences and different psychologies.
25
96440
3080
01:40
And this is what makes life really interesting.
26
100240
2720
01:43
But this is also what makes working on intelligent technology
27
103440
3496
01:46
extremely difficult.
28
106960
1600
01:49
And right now, AI research, well,
29
109640
3456
01:53
it's a bit lopsided on the tech side.
30
113120
2016
01:55
And that makes a lot of sense.
31
115160
2136
01:57
See, for every qualitative thing about us --
32
117320
2456
01:59
you know, those parts of us that are emotional, dynamic and subjective --
33
119800
4456
02:04
we have to convert it to a quantitative metric:
34
124280
3136
02:07
something that can be represented with facts, figures and computer code.
35
127440
4360
02:13
The issue is, there are many qualitative things
36
133000
3376
02:16
that we just can't put our finger on.
37
136400
1960
02:20
So, think about hearing your favorite song for the first time.
38
140400
3200
02:25
What were you doing?
39
145200
1200
02:28
How did you feel?
40
148000
1200
02:30
Did you get goosebumps?
41
150720
1360
02:33
Or did you get fired up?
42
153240
1640
02:36
Hard to describe, right?
43
156400
1200
02:38
See, parts of us feel so simple,
44
158800
2096
02:40
but under the surface, there's really a ton of complexity.
45
160920
3656
02:44
And translating that complexity to machines
46
164600
2936
02:47
is what makes them modern-day moonshots.
47
167560
2856
02:50
And I'm not convinced that we can answer these deeper questions
48
170440
4176
02:54
with just ones and zeros alone.
49
174640
1480
02:57
So, in the lab, I've been creating art
50
177120
1936
02:59
as a way to help me design better experiences
51
179080
2456
03:01
for bleeding-edge technology.
52
181560
2096
03:03
And it's been serving as a catalyst
53
183680
1736
03:05
to beef up the more human ways that computers can relate to us.
54
185440
3840
03:10
Through art, we're tacking some of the hardest questions,
55
190000
2696
03:12
like what does it really mean to feel?
56
192720
2360
03:16
Or how do we engage and know how to be present with each other?
57
196120
4080
03:20
And how does intuition affect the way that we interact?
58
200800
4000
03:26
So, take for example human emotion.
59
206440
2056
03:28
Right now, computers can make sense of our most basic ones,
60
208520
3256
03:31
like joy, sadness, anger, fear and disgust,
61
211800
3696
03:35
by converting those characteristics to math.
62
215520
3000
03:39
But what about the more complex emotions?
63
219400
2536
03:41
You know, those emotions
64
221960
1216
03:43
that we have a hard time describing to each other?
65
223200
2376
03:45
Like nostalgia.
66
225600
1440
03:47
So, to explore this, I created a piece of art, an experience,
67
227640
3936
03:51
that asked people to share a memory,
68
231600
2096
03:53
and I teamed up with some data scientists
69
233720
2136
03:55
to figure out how to take an emotion that's so highly subjective
70
235880
3576
03:59
and convert it into something mathematically precise.
71
239480
3200
04:03
So, we created what we call a nostalgia score
72
243840
2136
04:06
and it's the heart of this installation.
73
246000
2216
04:08
To do that, the installation asks you to share a story,
74
248240
3056
04:11
the computer then analyzes it for its simpler emotions,
75
251320
3256
04:14
it checks for your tendency to use past-tense wording
76
254600
2656
04:17
and also looks for words that we tend to associate with nostalgia,
77
257280
3336
04:20
like "home," "childhood" and "the past."
78
260640
3040
04:24
It then creates a nostalgia score
79
264760
2056
04:26
to indicate how nostalgic your story is.
80
266840
2736
04:29
And that score is the driving force behind these light-based sculptures
81
269600
4136
04:33
that serve as physical embodiments of your contribution.
82
273760
3896
04:37
And the higher the score, the rosier the hue.
83
277680
3216
04:40
You know, like looking at the world through rose-colored glasses.
84
280920
3936
04:44
So, when you see your score
85
284880
2616
04:47
and the physical representation of it,
86
287520
2656
04:50
sometimes you'd agree and sometimes you wouldn't.
87
290200
2936
04:53
It's as if it really understood how that experience made you feel.
88
293160
3480
04:57
But other times it gets tripped up
89
297400
2216
04:59
and has you thinking it doesn't understand you at all.
90
299640
2560
05:02
But the piece really serves to show
91
302680
1896
05:04
that if we have a hard time explaining the emotions that we have to each other,
92
304600
4056
05:08
how can we teach a computer to make sense of them?
93
308680
2360
05:12
So, even the more objective parts about being human are hard to describe.
94
312360
3576
05:15
Like, conversation.
95
315960
1240
05:17
Have you ever really tried to break down the steps?
96
317880
2736
05:20
So think about sitting with your friend at a coffee shop
97
320640
2656
05:23
and just having small talk.
98
323320
1320
05:25
How do you know when to take a turn?
99
325160
1720
05:27
How do you know when to shift topics?
100
327440
1840
05:29
And how do you even know what topics to discuss?
101
329960
2720
05:33
See, most of us don't really think about it,
102
333560
2096
05:35
because it's almost second nature.
103
335680
1656
05:37
And when we get to know someone, we learn more about what makes them tick,
104
337360
3496
05:40
and then we learn what topics we can discuss.
105
340880
2376
05:43
But when it comes to teaching AI systems how to interact with people,
106
343280
3656
05:46
we have to teach them step by step what to do.
107
346960
2856
05:49
And right now, it feels clunky.
108
349840
3496
05:53
If you've ever tried to talk with Alexa, Siri or Google Assistant,
109
353360
4136
05:57
you can tell that it or they can still sound cold.
110
357520
4200
06:02
And have you ever gotten annoyed
111
362440
1656
06:04
when they didn't understand what you were saying
112
364120
2256
06:06
and you had to rephrase what you wanted 20 times just to play a song?
113
366400
3840
06:11
Alright, to the credit of the designers, realistic communication is really hard.
114
371440
4896
06:16
And there's a whole branch of sociology,
115
376360
2136
06:18
called conversation analysis,
116
378520
1936
06:20
that tries to make blueprints for different types of conversation.
117
380480
3136
06:23
Types like customer service or counseling, teaching and others.
118
383640
4080
06:28
I've been collaborating with a conversation analyst at the lab
119
388880
2936
06:31
to try to help our AI systems hold more human-sounding conversations.
120
391840
4696
06:36
This way, when you have an interaction with a chatbot on your phone
121
396560
3176
06:39
or a voice-based system in the car,
122
399760
1856
06:41
it sounds a little more human and less cold and disjointed.
123
401640
3320
06:46
So I created a piece of art
124
406360
1336
06:47
that tries to highlight the robotic, clunky interaction
125
407720
2816
06:50
to help us understand, as designers,
126
410560
1976
06:52
why it doesn't sound human yet and, well, what we can do about it.
127
412560
4576
06:57
The piece is called Bot to Bot
128
417160
1456
06:58
and it puts one conversational system against another
129
418640
2936
07:01
and then exposes it to the general public.
130
421600
2536
07:04
And what ends up happening is that you get something
131
424160
2496
07:06
that tries to mimic human conversation,
132
426680
1896
07:08
but falls short.
133
428600
1896
07:10
Sometimes it works and sometimes it gets into these, well,
134
430520
2736
07:13
loops of misunderstanding.
135
433280
1536
07:14
So even though the machine-to-machine conversation can make sense,
136
434840
3096
07:17
grammatically and colloquially,
137
437960
2016
07:20
it can still end up feeling cold and robotic.
138
440000
3216
07:23
And despite checking all the boxes, the dialogue lacks soul
139
443240
4016
07:27
and those one-off quirks that make each of us who we are.
140
447280
3136
07:30
So while it might be grammatically correct
141
450440
2056
07:32
and uses all the right hashtags and emojis,
142
452520
2656
07:35
it can end up sounding mechanical and, well, a little creepy.
143
455200
4136
07:39
And we call this the uncanny valley.
144
459360
2336
07:41
You know, that creepiness factor of tech
145
461720
1936
07:43
where it's close to human but just slightly off.
146
463680
2856
07:46
And the piece will start being
147
466560
1456
07:48
one way that we test for the humanness of a conversation
148
468040
3216
07:51
and the parts that get lost in translation.
149
471280
2160
07:54
So there are other things that get lost in translation, too,
150
474560
2856
07:57
like human intuition.
151
477440
1616
07:59
Right now, computers are gaining more autonomy.
152
479080
2776
08:01
They can take care of things for us,
153
481880
1736
08:03
like change the temperature of our houses based on our preferences
154
483640
3176
08:06
and even help us drive on the freeway.
155
486840
2160
08:09
But there are things that you and I do in person
156
489560
2496
08:12
that are really difficult to translate to AI.
157
492080
2760
08:15
So think about the last time that you saw an old classmate or coworker.
158
495440
4360
08:21
Did you give them a hug or go in for a handshake?
159
501080
2480
08:24
You probably didn't think twice
160
504800
1496
08:26
because you've had so many built up experiences
161
506320
2336
08:28
that had you do one or the other.
162
508680
2000
08:31
And as an artist, I feel that access to one's intuition,
163
511440
3456
08:34
your unconscious knowing,
164
514920
1416
08:36
is what helps us create amazing things.
165
516360
3056
08:39
Big ideas, from that abstract, nonlinear place in our consciousness
166
519440
4056
08:43
that is the culmination of all of our experiences.
167
523520
2960
08:47
And if we want computers to relate to us and help amplify our creative abilities,
168
527840
4656
08:52
I feel that we'll need to start thinking about how to make computers be intuitive.
169
532520
3896
08:56
So I wanted to explore how something like human intuition
170
536440
3096
08:59
could be directly translated to artificial intelligence.
171
539560
3456
09:03
And I created a piece that explores computer-based intuition
172
543040
3216
09:06
in a physical space.
173
546280
1320
09:08
The piece is called Wayfinding,
174
548480
1696
09:10
and it's set up as a symbolic compass that has four kinetic sculptures.
175
550200
3936
09:14
Each one represents a direction,
176
554160
2056
09:16
north, east, south and west.
177
556240
2120
09:19
And there are sensors set up on the top of each sculpture
178
559080
2696
09:21
that capture how far away you are from them.
179
561800
2256
09:24
And the data that gets collected
180
564080
1816
09:25
ends up changing the way that sculptures move
181
565920
2136
09:28
and the direction of the compass.
182
568080
2040
09:31
The thing is, the piece doesn't work like the automatic door sensor
183
571360
3656
09:35
that just opens when you walk in front of it.
184
575040
2656
09:37
See, your contribution is only a part of its collection of lived experiences.
185
577720
5056
09:42
And all of those experiences affect the way that it moves.
186
582800
4056
09:46
So when you walk in front of it,
187
586880
1736
09:48
it starts to use all of the data
188
588640
1976
09:50
that it's captured throughout its exhibition history --
189
590640
2616
09:53
or its intuition --
190
593280
1816
09:55
to mechanically respond to you based on what it's learned from others.
191
595120
3560
09:59
And what ends up happening is that as participants
192
599480
2536
10:02
we start to learn the level of detail that we need
193
602040
2816
10:04
in order to manage expectations
194
604880
2016
10:06
from both humans and machines.
195
606920
2776
10:09
We can almost see our intuition being played out on the computer,
196
609720
3616
10:13
picturing all of that data being processed in our mind's eye.
197
613360
3600
10:17
My hope is that this type of art
198
617560
1656
10:19
will help us think differently about intuition
199
619240
2416
10:21
and how to apply that to AI in the future.
200
621680
2280
10:24
So these are just a few examples of how I'm using art to feed into my work
201
624480
3936
10:28
as a designer and researcher of artificial intelligence.
202
628440
3096
10:31
And I see it as a crucial way to move innovation forward.
203
631560
3496
10:35
Because right now, there are a lot of extremes when it comes to AI.
204
635080
4376
10:39
Popular movies show it as this destructive force
205
639480
2816
10:42
while commercials are showing it as a savior
206
642320
3056
10:45
to solve some of the world's most complex problems.
207
645400
2936
10:48
But regardless of where you stand,
208
648360
2536
10:50
it's hard to deny that we're living in a world
209
650920
2176
10:53
that's becoming more and more digital by the second.
210
653120
2576
10:55
Our lives revolve around our devices, smart appliances and more.
211
655720
4600
11:01
And I don't think this will let up any time soon.
212
661400
2320
11:04
So, I'm trying to embed more humanness from the start.
213
664400
3736
11:08
And I have a hunch that bringing art into an AI research process
214
668160
5136
11:13
is a way to do just that.
215
673320
1896
11:15
Thank you.
216
675240
1216
11:16
(Applause)
217
676480
2280
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