Transcript 강의 소개 - 인공지능연구실

Introduction to AI
Russell and Norvig:
Chapters 1 and 2
인공지능 접근방법
규칙에 의한 방법
통계적 방법
기계학습에 의한 방법
혼합하는 방법
Introduction to AI
2
동물의 행동은 본능인가 학습인가?
각인
언어능력
모국어
Introduction to AI
3
Human Brain
Chimpanzee Brain
Introduction to AI
4
아인슈타인의 뇌
Introduction to AI
5
보복전쟁 원인은 진화 덜 된 두뇌 탓
미국 폴 로스코 교수 주장
미국 과학진흥협회 (AAAS) annual meeting
“핵 기술을 보유한 인간의 두뇌가 여전히 석기 시대 수준에 머물러 있다”
“자신뿐 아니라 자기 종족까지 죽이는
‘복수’ 행위는
정상적인 진화의 산물이 아니다”
Introduction
to AI
6
'Speech Gene' Tied to Modern Humans
FOXP2 gene
- first identified by Monaco’s group at Oxford University (Science 2001)
- 715 amino acids, two amino acid mutation in human lineage since 6 million years ago –
fixed at 120,000 – 200,000 years ago (Svante Paabo, Nature 2002)
Evolutionary leap. One of these primates is able to talk about what
he's seeing; the other isn't. Introduction to AI
7
언어를 관할하는 FOXP2 유전자의 돌연변이
Introduction to AI
8
폭력의 생물학
Richard J. Davidson et al., Science 2000
(폭력과 전전두엽 장애 – 세로토닌 신경계 장애)
연쇄살인자 – 세로토닌 장애 – 책임문제
Introduction to AI
정신분열병 – 도파민 장애 – 살인 - 면죄부
9
뇌의 특정 유전자 과도 발현하는 형질 전환 마우스 제조
(smart mice)
T. V. P. BLISS et al., Nature,1999
transgenic mice overexpressing
the NR2B subunit of the NMDA
receptor
Introduction
to AI
10
최초의 유전자 치료술 (중증 복합 면역결핍증, SCID)
Introduction to AI
11
Found on the Web …
AI
is the reproduction of the methods of
Intelligent
behavior
human
reasoning or intuition
Computer
Using computational models
to simulate
intelligent (human) behavior and
processes
AI is the study of mental faculties
through the Humans
use computational methods
Introduction to AI
12
I personally think that AI started as a
rebellion against some form of establishment
telling us “Computers cannot perform certain
tasks requiring intelligence”
For example, for many years AI researchers
have regarded computational complexity
theory as irrelevant to their field.
They eventually had to reckon with it, but in
the meantime computational complexity had
also changed a lot.
Introduction to AI
13
What is AI?
Discipline that systematizes and automates
intellectual tasks to create machines that:
Act like humans
Act rationally
Think like humans
Think rationally
Introduction to AI
14
Act Like Humans
AI is the art of creating machines that
perform functions that require
intelligence when performed by humans
Methodology: Take an intellectual task
at which people are better and make a
computer do it •Prove a theorem
•Play chess
Turing test
•Plan a surgical operation
•Diagnose a disease
•Navigate
in
a
building
Introduction to AI
15
Chess
Name: Garry Kasparov
Title: World Chess
Champion
Crime: Valued greed
over common sense
Humans are still better at making up excuses.
© Jonathan Schaeffer
Introduction to AI
16
Perspective on Chess: Pro
“Saying Deep Blue doesn’t really think
about chess is like saying an airplane
doesn't really fly because it doesn't flap
its wings”
Drew McDermott
© Jonathan Schaeffer
Introduction to AI
17
Perspective on Chess: Con
“Chess is the Drosophila of artificial
intelligence. However, computer chess has
developed much as genetics might have if the
geneticists had concentrated their efforts
starting in 1910 on breeding racing Drosophila.
We would have some science, but mainly we
would have very fast fruit flies.”
John McCarthy
© Jonathan Schaeffer
Introduction to AI
18
Think Like Humans
How the computer performs functions
does matter
Comparison of the traces of the
reasoning steps
Cognitive science  testable theories of
the workings of the human mind
But, do we want to duplicate human imperfections?
Introduction to AI
19
Think/Act Rationally
Always make the best decision given what is
available (knowledge, time, resources)
•Connection
to economics,
A performance
measure operational
is required research,
and
control theory
 객관적인가?
청소로봇, 보상(reward)
디자이너의
•But ignores
role평가기준?
of consciousness, emotions,
 자원의 제한이 있을 때?
fear
of dying on intelligence
Perfect knowledge, unlimited resources 
logical reasoning
Imperfect knowledge, limited resources 
(limited) rationality
Introduction to AI
20
Bits of History
1956: The name “Artificial Intelligence”
was coined. (Would “computational
rationality” have been better?)
Early period (50’s to late 60’s):
Basic principles and generality




General problem solving
Theorem proving
Games
Formal calculus
Introduction to AI
21
Bits of History
1969-1971: Shakey the
robot (Fikes, Hart, Nilsson)
Logic-based planning
(STRIPS)
Motion planning (visibility
graph)
Inductive learning (PLANEX)
Computer vision
Introduction to AI
22
Bits of History
Knowledge-is-Power period (late 60’s to
mid 80’s):


Focus on narrow tasks require expertise
Encoding of expertise in rule form:
If:
the car has off-highway tires and
4-wheel drive and
high ground clearance
Then: the car can traverse difficult terrain (0.8)



Knowledge engineering
5th generation computer project
CYC system (Lenat)
Introduction to AI
23
Bits of History
AI becomes an industry (80’s – 2000):





Expert systems: Digital Equipment,
Teknowledge, Intellicorp, Du Pont, oil
industry, …
Lisp machines: LMI, Symbolics, …
Constraint programming: ILOG
Robotics: Machine Intelligence Corporation,
Adept, GMF (Fanuc), ABB, …
Speech understanding
Introduction to AI
24
Bits of History
The return of neural networks, genetic
algorithms, and artificial life (80’s – 90’s)
Increased connection with economics,
operational research, and control theory
(90’s – present)
AI becomes less philosophical, more
technical and mathematically oriented
Introduction to AI
25
Bits of History
2000-2010

Data Mining from Big Data and Applying to
AI Applications
 Trend Mining

AI Applications using Context-awareness
 Smartphone
 Ubiquitous Computing



Stochastic Approach
Human Gene Analysis
Autonomous Robots
Introduction to AI
26
Bits of History
Now, AI as a new paradigm for human
life and the new markets

Big-data and Data mining
 생산과 판매의 효율화
 의료 : healthcare
 오락: Watson

New applications with smartphones
 SIRI
 새로운 인터페이스
Introduction to AI
27
빅 데이터를 활용한 분석 영역은 무한합니다.
Smarter
Healthcare
Multi-channel
sales
Finance
Log Analysis
Homeland
Security
Traffic Control
Telecom
Search Quality
Fraud and
Risk
Retail: Churn,
NBO
Manufacturin
g
Trading
Analytics
Big Data Visualization을 통한 트렌드 분석
Spatial information flow
특정공간 안에서 정보의 흐름을 보여줌.
흐름이 많을 수록 링의 크기가 커짐
History flow
위키피디아의 경우 처럼 다수의 저자들이 문서를
수정하면서 문서가 변화된 양상을 기록한 것
Clustergram
클러스터 개수가 늘어남에 따라 각각의 데이터 셋
부분들에 클러스터가 어떻게 할당되는지를
보여주는 집약 분석 기법
Facebook transaction
Facebook 사용자의 활동을 정보의 흐름과 빈도로
표시 (지역별 사용 정도를 일목요연하게 보여 줌)
사례 : 구글 검색 트렌드와 비즈니스의 연관 관계
연관성의 주요 원인 : 제품 구매 전에 검색 엔진을 통해 정보 조사를
수행하는 인터넷 세대의 행동 심리
구글에서 ‘포드 경차’가 검색된 횟수
–
2004년 검색횟수를 100으로 했을 때 상대적인
비교
– 자료: 구글 트랜드
–
힐 배이런 UC 베클리 교수를 포드 경차가
구들에서 검색된 빈도와 판매량의 상관관계를
비교함
포드의 경차 판매량
볼보 자동차의 프로세스 개선 사례
•운행과정에서 발생되는 Data를 본사 분
수집
석 시스템에 자동 전송
•Big Data 분석을 통해 결함과 소비자의
분석
반영
잠재 Needs를 파악
•결함 및 잠재 Needs를 개발단계에 반영
•기존 50만대 판매 시점에 발견할 수 있
효과
던 결함을 1000대 판매 시점에 발견
http://www.i-cio.com/case-studies/volvo-big-data
Hadoop 활용 사례 : Yahoo & Visa
• Hadoop at Yahoo!
 25,000+ machines in 10+ clusters (largest is 3,000 machines)
 3 PBs of data (compressed, unreplicated)
 10,000+ jobs/week
• Hadoop@Visa
 2년치 raw transaction data를 이용하여 real-time risk scoring system에
사용될 데이타 요소들을 생성
 500M distinct accounts, 100M transactions per day, 200bytes per
transaction, 2 years total 73B transactions (36TB)
 Processing time : 1 months 13 minutes (3000 times faster)
빅 데이터 분석을 활용한 Watson
답을
알아내기 위해서 2억 페이지의 데이터를
분석합니다.
Watson의 고급 분석 기술 능력은 3초 내에
IBM Watson 산업별 활용 사례 – Texas Seton Healthcare
Big Data Analytics for Smarter Healthcare
 텍사스 주에서 가장 우수한 의료 시스템을 갖춘 의료기관이자, 미국에서 통합
의료 시스템을 갖춘 100개 병원 중 하나
 2011년 10월부터 상용화된 Watson ( IBM’s Content and Predictive Analytics
for Health Care) 시스템을 도입
 환자가 미래에 겪을 수 있는 질환이나 증상을 미리 예측하여 이를 예방하고
통합적이고 확장된 의료 서비스를 제공하기 위한 작업을 준비
 잠재적인 위험 요소를 미리 예측하고 질병을 예방하기 위한 비용에 대한
보험사의 보상까지 가능하도록 업무를 업그레이드
Predictions and Reality … (1/3)
In the 60’s, a famous AI professor from MIT
said: “At the end of the summer, we will have
developed an electronic eye”
As of 2002, there is still no general computer
vision system capable of understanding
complex dynamic scenes
But computer systems routinely perform road
traffic monitoring, facial recognition, some
medical image analysis, part inspection, etc…
Introduction to AI
35
Predictions and Reality … (2/3)
In 1958, Herbert Simon (CMU) predicted
that within 10 years a computer would
be Chess champion
This prediction became true in 1998
Today, computers have won over world
champions in several games, including
Checkers, Othello, and Chess, but still do
not do well in Go
Introduction to AI
36
Predictions and Reality … (3/3)
In the 70’s, many believed that computer-controlled
robots would soon be everywhere from manufacturing
plants to home
Today, some industries (automobile, electronics) are
highly robotized, but home robots are still a thing of
the future
But robots have rolled on Mars, others are performing
brain and heart surgery, and humanoid robots are
operational and available for rent (see:
http://world.honda.com/news/2001/c011112.html)
Introduction to AI
37
Mistakes …
Often, the potential of a new field is
over-estimated in its early age, but
under-estimated over the longer term
AI proponents have over-estimated the
need for smart software, and underestimated the feasibility and potential of
large software systems based on
massive coding effort
Introduction to AI
38
Mistakes …
Often, the potential of a new field is over-estimated
in its early age, but under-estimated over the longer
term
What about Bio-informatics?

줄기세포의 치료효과
 쥐의 실험에 의하면 척추장애를 완벽히 치료
 그러나 저항력이 없는 쥐에서는 바로 암으로 전이, 있을 때에도
3개월 뒤면 암으로 전이
? 탄소 큐브 (나노기술)
 완벽한 새로운 물질
 인류는 경험하지 못했으며, 가장 무서운 발암물질?
핵, 유전자변형???
Introduction to AI
39
Intelligent Agent
인간의 능력을 대신할 수 있는


부분적: 걷는? 말하는? 판단하는?
인간과 같은?
지능적 능력을 가진


인간과 다른 형태의?
말하는 개? 말하는 앵무새? 덧셈을 하는 개?
비교: SOAP, Service-oriented approach, XML,
Context, …
Introduction to AI
40
Notion of an Agent
sensors
?
environment
agent
actuators
laser range
finder
sonars
Introduction to AI
touch sensors
41
Notion of an Agent
sensors
?
environment
agent
actuators
Introduction to AI
42
Notion of an Agent
sensors
?
environment
agent
actuators
•Locality of sensors/actuators
•Imperfect modeling
•Time/resource constraints
•Sequential interaction
•Multi-agent
worlds
Introduction to AI
43
단순화한 인공지능 문제
에이전트의 조건




성공을 평가한 판단기준
환경 또는 응용영역에 대한 사전지식
Agent가 행할 수 있는 행위
현재까지 한 행위와 결과에 대한 인식
문제의 어려움






Fully observable vs. Partially observable
Deterministic vs. Stochastic
Episodic vs. Sequential
Static vs. Dynamic
Discrete vs. Continuous
Single agent vs. Multiagent
Introduction to AI
44
Example: Tracking a Target
• The robot must keep
the target in view
• The target’s trajectory
is not known in advance
• The robot may not know
all the obstacles in
advance
• Fast decision is required
robot
Introduction to AI
target
45