How does artificial intelligence learn? - Briana Brownell

648,705 views ・ 2021-03-11

TED-Ed


μ•„λž˜ μ˜λ¬Έμžλ§‰μ„ λ”λΈ”ν΄λ¦­ν•˜μ‹œλ©΄ μ˜μƒμ΄ μž¬μƒλ©λ‹ˆλ‹€.

λ²ˆμ—­: Jeongyeon Kim κ²€ν† : DK Kim
00:09
Today, artificial intelligence helps doctors diagnose patients,
0
9829
5167
μ˜€λŠ˜λ‚  인곡지λŠ₯은 μ˜μ‚¬λ“€μ˜ ν™˜μž 진찰,
00:14
pilots fly commercial aircraft, and city planners predict traffic.
1
14996
5042
μ‘°μ’…μ‚¬μ˜ λ―Όκ°„ λΉ„ν–‰κΈ° μ‘°μ’…, λ„μ‹œ κ³„νšκ°€λ“€μ˜ ꡐ톡 μ˜ˆμΈ‘μ„ λ•μŠ΅λ‹ˆλ‹€.
00:20
But no matter what these AIs are doing, the computer scientists who designed them
2
20038
4250
ν•˜μ§€λ§Œ 이 인곡지λŠ₯이 무엇을 ν•˜λŠ”μ§€μ— 관계 없이
개발자인 컴퓨터 κ³Όν•™μžλ“€μ€ 인곡지λŠ₯이 그것을 μ–΄λ–»κ²Œ ν•˜λŠ”μ§€ 잘 λͺ¨λ¦…λ‹ˆλ‹€.
00:24
likely don’t know exactly how they’re doing it.
3
24288
2750
00:27
This is because artificial intelligence is often self-taught,
4
27038
3875
인곡지λŠ₯이 보톡 슀슀둜 배우고
00:30
working off a simple set of instructions
5
30913
2374
κ°„λ‹¨ν•œ κ·œμΉ™λ“€λ§ŒμœΌλ‘œ
00:33
to create a unique array of rules and strategies.
6
33287
3584
νŠΉμ΄ν•œ κ·œμΉ™κ³Ό μ „λž΅λ“€μ„ λ§Œλ“€μ–΄λ‚΄κΈ° λ•Œλ¬Έμž…λ‹ˆλ‹€.
00:36
So how exactly does a machine learn?
7
36871
2625
그러면 μ»΄ν“¨ν„°λŠ” μ •ν™•νžˆ μ–΄λ–»κ²Œ λ°°μšΈκΉŒμš”?
00:39
There are many different ways to build self-teaching programs.
8
39496
3000
슀슀둜 λ°°μš°λŠ” ν”„λ‘œκ·Έλž¨μ„ λ§Œλ“œλŠ” λ°μ—λŠ” μ—¬λŸ¬ 가지 방법이 μžˆμŠ΅λ‹ˆλ‹€.
00:42
But they all rely on the three basic types of machine learning:
9
42496
3916
ν•˜μ§€λ§Œ λͺ¨λ‘κ°€ μ„Έ 가지 기본적인 λ¨Έμ‹ λŸ¬λ‹ 방법을 기초둜 ν•©λ‹ˆλ‹€.
00:46
unsupervised learning, supervised learning, and reinforcement learning.
10
46412
5000
비지도 ν•™μŠ΅, 지도 ν•™μŠ΅, 보강 ν•™μŠ΅μž…λ‹ˆλ‹€.
00:51
To see these in action,
11
51412
1959
이것이 μž‘λ™ν•˜λŠ” 방법을 보렀면
00:53
let’s imagine researchers are trying to pull information
12
53371
3250
수천 개의 ν™˜μž 정보가 μžˆλŠ” 의료 μžλ£Œμ—μ„œ
00:56
from a set of medical data containing thousands of patient profiles.
13
56621
4208
연ꡬ가듀이 μ–΄λ–€ 정보λ₯Ό μ°ΎλŠ”λ‹€κ³  μƒμƒν•΄λ΄…μ‹œλ‹€.
01:01
First up, unsupervised learning.
14
61371
3125
λ¨Όμ € 비지도 ν•™μŠ΅μž…λ‹ˆλ‹€.
01:04
This approach would be ideal for analyzing all the profiles
15
64496
3333
이 μ ‘κ·Ό 방식은 곡톡점과 μœ μ΅ν•œ κ·œμΉ™μ„ μ°ΎκΈ° μœ„ν•΄
01:07
to find general similarities and useful patterns.
16
67829
3625
λͺ¨λ“  정보λ₯Ό λΆ„μ„ν•˜κΈ°μ— μ•Œλ§žμŠ΅λ‹ˆλ‹€.
01:11
Maybe certain patients have similar disease presentations,
17
71454
3417
μ•„λ§ˆ μ–΄λ–€ ν™˜μžλ“€μ€ λΉ„μŠ·ν•œ 증상을 λ³΄μ΄κ±°λ‚˜
01:14
or perhaps a treatment produces specific sets of side effects.
18
74871
4042
ν˜Ήμ€ μ–΄λ–€ μΉ˜λ£Œλ²•μ΄ νŠΉμ •ν•œ λΆ€μž‘μš©μ„ λΆˆλŸ¬μΌμœΌν‚¬ κ²ƒμž…λ‹ˆλ‹€.
01:18
This broad pattern-seeking approach can be used to identify similarities
19
78913
4125
κ·œμΉ™μ„ μ°ΎλŠ” 이 λ°©λŒ€ν•œ 접근법은 μ‚¬λžŒμ˜ 지도 없이
01:23
between patient profiles and find emerging patterns,
20
83038
3375
ν™˜μž 정보 κ°„ 곡톡점과 κ·œμΉ™μ„
01:26
all without human guidance.
21
86413
2291
μ•Œμ•„λ‚΄λŠ” 데에 μ‚¬μš©λ  수 μžˆμŠ΅λ‹ˆλ‹€.
01:28
But let's imagine doctors are looking for something more specific.
22
88704
3209
ν•˜μ§€λ§Œ μ˜μ‚¬λ“€μ΄ 더 ꡬ체적인 것을 μ°ΎλŠ”λ‹€κ³  κ°€μ •ν•΄λ΄…μ‹œλ‹€.
01:32
These physicians want to create an algorithm
23
92371
2208
이 μ˜μ‚¬λ“€μ€ μ–΄λ–€ μƒνƒœλ₯Ό μ§„λ‹¨ν•˜λŠ”
01:34
for diagnosing a particular condition.
24
94579
2709
μ•Œκ³ λ¦¬μ¦˜μ„ λ§Œλ“€μ–΄λ‚΄κ³  μ‹Άμ–΄ν•©λ‹ˆλ‹€.
01:37
They begin by collecting two sets of dataβ€”
25
97288
2583
그듀은 두 가지 자료λ₯Ό ν™•λ³΄ν•˜λŠ” 것뢀터 μ‹œμž‘ν•©λ‹ˆλ‹€.
01:39
medical images and test results from both healthy patients
26
99871
3375
κ±΄κ°•ν•œ μ‚¬λžŒκ³Ό μ§ˆλ³‘μ„ 진단받은 ν™˜μžμ˜
01:43
and those diagnosed with the condition.
27
103246
2417
의료 μ˜μƒκ³Ό μ‹œν—˜ κ²°κ³Όμž…λ‹ˆλ‹€.
01:45
Then, they input this data into a program
28
105663
2583
κ·Έ λ‹€μŒμ— κ±΄κ°•ν•œ μ‚¬λžŒλ“€μ—κ²ŒλŠ” μ—†κ³ 
01:48
designed to identify features shared by the sick patients
29
108246
3375
ν™˜μžλ“€λ§Œμ΄ κ³΅μœ ν•˜λŠ” μš”μ†Œλ₯Ό μ•Œμ•„λ‚΄λŠ” ν”„λ‘œκ·Έλž¨μ—
01:51
but not the healthy patients.
30
111621
2125
이 자료λ₯Ό μ§‘μ–΄λ„£μŠ΅λ‹ˆλ‹€.
01:53
Based on how frequently it sees certain features,
31
113746
3083
νŠΉμ •ν•œ μš”μ†Œλ₯Ό μ–Όλ§ˆλ‚˜ 자주 λ°œκ²¬ν•˜λŠλƒμ— 따라
01:56
the program will assign values to those features’ diagnostic significance,
32
116829
4042
ν”„λ‘œκ·Έλž¨μ€ κ·Έ μš”μ†Œμ˜ 진단적 μ€‘μš”λ„μ— 값을 맀기며
02:00
generating an algorithm for diagnosing future patients.
33
120871
3708
μ•žμœΌλ‘œ ν™˜μžλ“€μ„ μ§„μ°°ν•˜λŠ” 데에 μœ μš©ν•œ μ•Œκ³ λ¦¬μ¦˜μ„ λ§Œλ“€ κ²ƒμž…λ‹ˆλ‹€.
02:04
However, unlike unsupervised learning,
34
124579
3167
κ·ΈλŸ¬λ‚˜ 비지도 ν•™μŠ΅κ³Ό λ‹€λ₯΄κ²Œ
02:07
doctors and computer scientists have an active role in what happens next.
35
127746
4625
μ˜μ‚¬λ“€κ³Ό 컴퓨터 κ³Όν•™μžλ“€μ€ κ·Έ λ‹€μŒ 단계에 더 κ°œμž…ν•©λ‹ˆλ‹€.
02:12
Doctors will make the final diagnosis
36
132371
2083
μ˜μ‚¬λ“€μ€ μ΅œμ’…μ μΈ 진단을 내리고
02:14
and check the accuracy of the algorithm’s prediction.
37
134454
2917
μ•Œκ³ λ¦¬μ¦˜μ΄ ν•œ 예츑의 정확도λ₯Ό 확인할 κ²λ‹ˆλ‹€.
02:17
Then computer scientists can use the updated datasets
38
137871
2917
κ·Έ λ‹€μŒ 컴퓨터 κ³Όν•™μžλ“€μ€ μƒˆλ‘œμš΄ μžλ£Œλ“€λ‘œ
02:20
to adjust the program’s parameters and improve its accuracy.
39
140788
3500
ν”„λ‘œκ·Έλž¨μ˜ λ§€κ°œλ³€μˆ˜λ₯Ό μ‘°μ •ν•΄ 정확도λ₯Ό 높일 수 μžˆμŠ΅λ‹ˆλ‹€.
02:24
This hands-on approach is called supervised learning.
40
144663
3166
μ΄λ ‡κ²Œ κ°œμž…ν•˜λŠ” 접근법이 지도 ν•™μŠ΅μž…λ‹ˆλ‹€.
02:27
Now, let’s say these doctors want to design another algorithm
41
147829
3167
이제 이 μ˜μ‚¬λ“€μ΄ 치료 κ³„νšμ„ μ œμ•ˆν•˜κΈ° μœ„ν•΄
02:30
to recommend treatment plans.
42
150996
1750
λ‹€λ₯Έ μ•Œκ³ λ¦¬μ¦˜μ„ λ§Œλ“ λ‹€κ³  ν•΄λ΄…μ‹œλ‹€.
02:32
Since these plans will be implemented in stages,
43
152746
2833
이 κ³„νšλ“€μ€ λ‹¨κ³„μ μœΌλ‘œ 싀행될 것이고
02:35
and they may change depending on each individual's response to treatments,
44
155579
3959
μΉ˜λ£Œλ²•μ— λŒ€ν•œ 개개인의 λ°˜μ‘μ— 따라 λ°”λ€” 수 있기 λ•Œλ¬Έμ—
02:39
the doctors decide to use reinforcement learning.
45
159538
2833
μ˜μ‚¬λ“€μ€ 보강 ν•™μŠ΅μ„ μ“°κΈ°λ‘œ ν•©λ‹ˆλ‹€.
02:42
This program uses an iterative approach to gather feedback
46
162746
3167
이 ν”„λ‘œκ·Έλž¨μ€ μ–΄λ–€ μ•½, λ³΅μš©λŸ‰, μΉ˜λ£Œκ°€ κ°€μž₯ νš¨κ³Όμ μΈμ§€μ— λŒ€ν•œ
02:45
about which medications, dosages and treatments are most effective.
47
165913
4583
ν”Όλ“œλ°±μ„ λ°›κΈ° μœ„ν•΄ 반볡 접근법을 μ‚¬μš©ν•©λ‹ˆλ‹€.
02:50
Then, it compares that data against each patient’s profile
48
170496
3000
κ·Έ λ‹€μŒμ—λŠ” 각 ν™˜μžμ˜ 졜적 μΉ˜λ£Œλ²•μ„ λ§Œλ“€κΈ° μœ„ν•΄
02:53
to create their unique, optimal treatment plan.
49
173496
2833
κ·Έ 자료λ₯Ό ν™˜μž 정보와 λΉ„κ΅ν•©λ‹ˆλ‹€.
02:56
As the treatments progress and the program receives more feedback,
50
176329
3500
μΉ˜λ£Œκ°€ μ§„ν–‰λ˜κ³  더 λ§Žμ€ ν”Όλ“œλ°±μ„ λ°›μ„μˆ˜λ‘
02:59
it can constantly update the plan for each patient.
51
179829
3292
μΉ˜λ£Œλ²•μ€ μ§€μ†μ μœΌλ‘œ κ°±μ‹ λ©λ‹ˆλ‹€.
03:03
None of these three techniques are inherently smarter than any other.
52
183121
3375
이 μ„Έ 가지 기술 쀑 μ–΄λ–€ 것이 λ‹€λ₯Έ 것보닀 더 λ‚˜μ€ 것은 μ•„λ‹™λ‹ˆλ‹€.
03:06
While some require more or less human intervention,
53
186496
2667
μ–΄λ–€ κΈ°μˆ μ€ μΈκ°„μ˜ κ°œμž…μ„ μ–΄λŠ 정도 ν•„μš”λ‘œ ν•˜μ§€λ§Œ
03:09
they all have their own strengths and weaknesses
54
189163
2416
κ·Έ 기술 λͺ¨λ‘λŠ” 각자 강점과 약점이 μžˆμ–΄
03:11
which makes them best suited for certain tasks.
55
191579
2250
νŠΉμ •ν•œ μΌμ—λŠ” λ‹€λ₯Έ 것보닀 더 μ ν•©ν•©λ‹ˆλ‹€.
03:14
However, by using them together,
56
194329
2375
κ·Έ 기술 λͺ¨λ‘λ₯Ό 같이 μ‚¬μš©ν•˜λ©΄
03:16
researchers can build complex AI systems,
57
196704
2875
μ—°κ΅¬μžλ“€μ΄ λ³΅μž‘ν•œ 인곡지λŠ₯ μ‹œμŠ€ν…œμ„ λ§Œλ“€ 수 μžˆμ–΄
03:19
where individual programs can supervise and teach each other.
58
199579
3292
각각의 ν”„λ‘œκ·Έλž¨μ΄ μ„œλ‘œλ₯Ό κ°λ…ν•˜κ³  κ°€λ₯΄μΉ  수 μžˆμŠ΅λ‹ˆλ‹€.
03:22
For example, when our unsupervised learning program
59
202871
2958
예λ₯Ό λ“€μ–΄ 비지도 ν•™μŠ΅ ν”„λ‘œκ·Έλž¨μ΄
03:25
finds groups of patients that are similar,
60
205829
2334
λΉ„μŠ·ν•œ λͺ‡λͺ‡ ν™˜μž 집단을 μ°Ύμ•„λ‚΄λ©΄
03:28
it could send that data to a connected supervised learning program.
61
208163
3375
μ—°κ²°λœ 지도 ν•™μŠ΅ ν”„λ‘œκ·Έλž¨μ— κ·Έ 정보λ₯Ό 보낼 수 μžˆμŠ΅λ‹ˆλ‹€.
03:31
That program could then incorporate this information into its predictions.
62
211829
3500
그러면 지도 ν•™μŠ΅ ν”„λ‘œκ·Έλž¨μ€ 이 μ •λ³΄λ‘œ μ˜ˆμΈ‘μ„ ν•  수 μžˆμŠ΅λ‹ˆλ‹€.
03:35
Or perhaps dozens of reinforcement learning programs
63
215871
2792
ν˜Ήμ€ μˆ˜μ‹­ 개의 보강 ν•™μŠ΅ ν”„λ‘œκ·Έλž¨μ΄
03:38
might simulate potential patient outcomes
64
218663
2291
ν™˜μžμ˜ κ²°κ³Όλ₯Ό λͺ¨μ˜ μ‹€ν—˜ν•΄μ„œ
03:40
to collect feedback about different treatment plans.
65
220954
2750
λ‹€μ–‘ν•œ μΉ˜λ£Œλ²•μ— λŒ€ν•œ ν”Όλ“œλ°±μ„ 얻을 수 μžˆμŠ΅λ‹ˆλ‹€.
03:43
There are numerous ways to create these machine-learning systems,
66
223704
3125
이 λ¨Έμ‹ λŸ¬λ‹ μ‹œμŠ€ν…œμ„ λ§Œλ“œλŠ” λ°μ—λŠ” λ§Žμ€ 방법듀이 있고
03:46
and perhaps the most promising models
67
226829
1834
μ•„λ§ˆ κ°€μž₯ μž₯λž˜μ„± μžˆλŠ” λͺ¨λΈμ€
03:48
are those that mimic the relationship between neurons in the brain.
68
228663
3416
λ‡Œ μ‹ κ²½μ„Έν¬λ“€μ˜ 연계λ₯Ό λͺ¨λ°©ν•˜λŠ” 것듀일 κ²ƒμž…λ‹ˆλ‹€.
03:52
These artificial neural networks can use millions of connections
69
232079
3292
이 인곡 신경망은 수백만 개의 μ ‘μ†μœΌλ‘œ
03:55
to tackle difficult tasks like image recognition, speech recognition,
70
235371
4417
사진 인식, μŒμ„± 인식, 심지어 μ–Έμ–΄ 톡역 같은
03:59
and even language translation.
71
239788
2041
μ–΄λ €μš΄ 일을 ν•΄λƒ…λ‹ˆλ‹€.
04:01
However, the more self-directed these models become,
72
241829
3292
그런데 이 λͺ¨λΈλ“€μ΄ 더 λ…λ¦½μ μΌμˆ˜λ‘
04:05
the harder it is for computer scientists
73
245121
2125
컴퓨터 κ³Όν•™μžλ“€μ΄ 이 슀슀둜 κ°€λ₯΄μΉ˜λŠ” μ•Œκ³ λ¦¬μ¦˜μ΄
04:07
to determine how these self-taught algorithms arrive at their solution.
74
247246
3833
μ–΄λ–»κ²Œ 결둠을 λ‚΄λŠ”μ§€λ₯Ό μ•Œμ•„λ‚΄κΈ°κ°€ 더 μ–΄λ ΅μŠ΅λ‹ˆλ‹€.
04:11
Researchers are already looking at ways to make machine learning more transparent.
75
251079
4459
μ—°κ΅¬μžλ“€μ€ 이미 λ¨Έμ‹ λŸ¬λ‹μ„ 더 투λͺ…ν•˜κ²Œ λ§Œλ“œλŠ” 법을 μ°Ύκ³  μžˆμŠ΅λ‹ˆλ‹€.
04:15
But as AI becomes more involved in our everyday lives,
76
255538
2916
ν•˜μ§€λ§Œ 인곡지λŠ₯이 μΌμƒμƒν™œμ— 더 μŠ€λ©°λ“€μˆ˜λ‘
04:18
these enigmatic decisions have increasingly large impacts
77
258454
2792
이 수수께기 같은 결정듀은 우리의 직업, 건강, μ•ˆμ „μ—
04:21
on our work, health, and safety.
78
261246
2875
점점 더 큰 영ν–₯을 미치고 μžˆμŠ΅λ‹ˆλ‹€.
04:24
So as machines continue learning to investigate, negotiate and communicate,
79
264121
4958
λ”°λΌμ„œ 기계가 쑰사, ν˜‘μƒ, μ†Œν†΅μ„ ν•™μŠ΅ν• μˆ˜λ‘
04:29
we must also consider how to teach them to teach each other to operate ethically.
80
269079
5209
μš°λ¦¬λŠ” 기계가 μ„œλ‘œλ₯Ό 윀리적이 λ˜λ„λ‘ κ°€λ₯΄μΉ˜κ²Œ ν•˜λŠ” 법을 생각해야 ν•©λ‹ˆλ‹€.
이 μ›Ήμ‚¬μ΄νŠΈ 정보

이 μ‚¬μ΄νŠΈλŠ” μ˜μ–΄ ν•™μŠ΅μ— μœ μš©ν•œ YouTube λ™μ˜μƒμ„ μ†Œκ°œν•©λ‹ˆλ‹€. μ „ 세계 졜고의 μ„ μƒλ‹˜λ“€μ΄ κ°€λ₯΄μΉ˜λŠ” μ˜μ–΄ μˆ˜μ—…μ„ 보게 될 κ²ƒμž…λ‹ˆλ‹€. 각 λ™μ˜μƒ νŽ˜μ΄μ§€μ— ν‘œμ‹œλ˜λŠ” μ˜μ–΄ μžλ§‰μ„ 더블 ν΄λ¦­ν•˜λ©΄ κ·Έκ³³μ—μ„œ λ™μ˜μƒμ΄ μž¬μƒλ©λ‹ˆλ‹€. λΉ„λ””μ˜€ μž¬μƒμ— 맞좰 μžλ§‰μ΄ μŠ€ν¬λ‘€λ©λ‹ˆλ‹€. μ˜κ²¬μ΄λ‚˜ μš”μ²­μ΄ μžˆλŠ” 경우 이 문의 양식을 μ‚¬μš©ν•˜μ—¬ λ¬Έμ˜ν•˜μ‹­μ‹œμ˜€.

https://forms.gle/WvT1wiN1qDtmnspy7