인공지능과 딥러닝에 대한 소개
TRANSCRIPT
-
?
?
?
1
/()
, : - ()
-
1. ?
-
vs , , (2016 )
2011 IBM Watson
Jeopardy 100
Watson Watson
Chef Watson (recipe)
Watson ,
() ( )
581 80% 472 80% A
IBM Watson on Jeopardy Show
-
?
( + Aldebaran Robotics) 200 2015 7 1,000
Pepper, Softbank
-
-
(classification)
(real-time bidding)
90% (high frequency trading)
UBIC (Paralegal)
UBIC e-Discovery Service
-
SF ?
() 1,000 .
.2016 () () 4 1 1
()
-
AI
Google 2013 1 (Geoffrey E. Hinton) DNN 2014 4 (CEO: Demis Hassabis)
AlphaGoFacebook 2013 (: Yann LeCun )
() 2014 Institute of Deep Learning() 3 , : (Andrew Ng) ( )
IBM Watson
10 (1) 2,000 1 Watson
Dwango (Dial-up Wide-Area Network Game Operation) 2014 : ()
PFI 2014 PFI(Preferred Infrastructure) Preferred Networks
Deep-learning , IoT NTT 2
Preferred Networks
-
( )
? 2014 (Deloitte) 35% 20
IT 20 50%
2014 (Transcendence) 2014 (Her) , ( )2015 (Imitation game) (Alan Turing) 1968 (Stanley Kubrick) 2001 (2001: Space Odyssey) HAL 9000
1984 (Terminator)
-
(Technological Singularity)
?(singularity) (Ray Kurzweil) 2045
(Elon Musk) . .
(Bill Gates)
Google DeepMind
2014
?
Singularity is imminent?
-
2. ?
-
. ?
.
.
-
, ?
.
.
.(Turing)
. (Turing machine)
(Marvin Lee Minsky)
(Roger Penrose)
(Hubert Lederer Dreyfus)
. .
-
() ?
.
, (), ()
(),
() (),
(),
(),
(),
(), ()
, , (), ()
, (),
(Marvin Minsky)
Artificial intelligence is the science of making machines do
things that would require intelligence if done by men
-
, AI
(, AlphaGo)
AI , , (), (, ), AI
, ( ) (Marvin Minsky) : AI
-
?
?
(Stuart Russell), ()
1 : , ,
2 / ,
3 ,
4 (Deep Learning)
Stuart Russell
A definition of intelligence needs to be formala property of the systems input, structure, and outputso that it can support analysis and synthesis. The Turing test does not meet this requirement
-
, , ,
4 1 () , ()
, ()
2 ,
, ,
3 , , , , ,
4 . .
-
AI AI
AI AI (Strong AI as Artificial General Intelligence)
AI (Strong AI as Computational Theory of Mind)
(John Rogers Searle) ( )
AI
AI John Searle AI (John Searl )
. . . . . AI .
-
- (John Searle) 1984
Axioms 1. = (, Syntactic) 2. = (Semantics) 3. (syntax) (semantics) .
axiom Axiom
1 .
. () . . . , .
Axiom
4. .
2 (causal power) .
3 . . ( 1,2)
4 .
John Rogers Searle
-
3. 1 AI
-
AI
1 1950 1960
1970 (toy problem)
2 1980
. 1990
3 (Watson, , AlphaGo)
1960 1970 1980 1990 2000 2010
1 AI 2 AI
3 AI
Watson
-
1956 1970 1980 20151995 20101 (, )
2 ()
3 (, )
ELIZA SIRI
botCALO
WatsonLOD(Linked Open Data)
DENDRAL
-
,
, AI
-
1956 : Dartmouth Summer Research Project on Artificial Intelligence : (John McCarthy) 1956 2 10 .
. , , , . .
John McCarthy (1927~2011)
Marvin Minsky1927~
Allan Newell1927~1992
Herbert Simon1916~2001
4 Turing Simon
-
1 AI : : , : (ex. Depth-first iterative-deepening)
() ()
AA
BB CC
DDEE
FF
GG
HH
II
AA
BB CC
DDEE
FF
GG
HH
II
start
goal goal
start start
goal
start
AAEE
CCBBFFGG
II HHDD goal
-
. : 1883 (douard Lucas)
(dome) . 1 . 64 . , . . . , .
: .
64 18,445,744,073,709,551,616 1 5849 4241 7355
-
Planning (condition)
(action) (result)
(condition) (action) (result)
STRIPS (Stanford Research Institute Problem Solver) 1971 (Richard Fikes) (Nils Nilsson)
SHRDLU (Terry Winograd) Planner
(Sergey Brin), (Larry Page)
Terry Winograd
-
, (brute-force search, exhaustive search)
: 1060 / : 10120 / : 10220 / : 10360
1080
Minimax
(maximum loss)
i : / -i : ai : , a-i : vi :
-
Minimax
.4 1. ( ) 2. 3. 4. 5.
.
( ) Minimax
Minimax search
-
, 1997 IBM (DeepBlue)
(Garry Kasparov) 2012 ( ) ( )
() () 2016 AlphaGo
.
.
()
-
Perceptron (Frank Rosenblatt)1957 Mark I Perceptron
IBM 704
20x20 400 Mark I Perceptron
-
1
1960
, , =
(Marvin Minsky) Perceptron Perceptron XOR XOR ?
1970 1
-
(Papert) , Perceptrons: an introduction to computational geometry
Marvin Minsky and Seymour Papert, 1969.
1 (Bronx High School of Science)
.
XOR
(symbolic)
: XOR , . / Boolean McCulloch-Pitt
3 feed-forward
connectedness
-
4. 2 AI
-
1980
2 AI =
1 AI 1964 ELIZA
1970 (MYCIN)
1960 (Edward Albert Feigenbaum) DENDRAL
Dendritic Algorithm Heuristic-DENDRAL: performance Meta-DENDRAL: learning
MYCIN, MOLGEN, MACSYMA, PROSPECTOR, XCON DENDRAL
Edward FeigenbaumThe Father of Expert System
DENDRAL , 1994. Turing
-
(description)
(domain)
? 4 ? 10 ?
-
(semantic network) (node) ( )
Cyc 1984 (Douglas Lenat) (#$isa #$BillClinton #$UnitedStatesPresident) Bill Clinton belongs to the collection of U.S. President (#$genls #$Tree-ThePlant #$Plant) All trees are plants
Commonsense Reasoning and Commonsense Knowledge in Artificial Intelligence, Ernest Davis and Gary Marcus, Communications of the ACM, 58(9): 92-103, 2015
-
(ontology)
(ontology) () Towards Ontology Engineering, Technical Report AI-TR-96-1, I.S.I.R., Osaka University Theory of Existence An explicit representation of conceptualization / A Theory of vocabulary/concepts used in building artificial system
(, transitive law) Is-a
is-a ^ is-a is-a
Part-of part-of ^ part-of
part-of
part-of ^ part-of part-of
(JAIST)
Ontology example
-
(Ontology)
Linked Data Applications: There is no One-Size-Fits-All Formula,Asuncion Gomez-Perez,The 9th Summer School on Ontology Engineering and Semantic Web, 2012
-
(heavyweight ontology)
(lightweight ontology) ,
Thesaurus vs ontology in Why an ontological approach?, Oreste Signore, http://www.weblab.isti.cnr.it/talks/2009/iccu/slides.html#(1)
http://www.ai-one.com/tag/lightweight-ontology/
-
IBM Watson
2011 Jeopardy -
: 1871 :
/
? O O O X X
() ? X O X O O
1871 ? X O X O X
1300 500 200 150 10
2% 92% 20% 6% 0%
Watson - , -
-
He saw a woman in the garden with a telescope. 1. .2. .3. .
2 , . ?
Time flies like an arrow . .
Time flies, Dwarf4r, http://dwarf4r.deviantart.com/art/Time-Flies-332901940
-
(frame problem)
: (John McCarthy) (Patric Hayes) (1969) (Daniel Dennett)
Mission: Robot1: . . . . . . () .Robot2: . . . . . . . time over. . . Robot3: ? ? .
, , , ,
, , .
-
Frame problem John McCarthy
open / on open(t), on(t)
, !open(0)!on(0)True open(1) : : !locked(0) open(1) : for all t, execute(t) ^ true open(t+1)
3 !open(0), !on(0) open(1), !on(1) !open(0), !on(0) open(1), on(1)
?Frame axiom: nothing else changes
Frame axiom
-
(Symbol grounding) (Stevan Harnard) 1990 (referent)
Tony Blair The prime minister of the UK during the year 2004 Cherie Blairs husband
? . .
? (sensor) ? .
Aaron Sloman, Univ. of Birmingham( )
-
2 AI
2 AI .
AI
1. 1966 ALPAC
2. 1969 Symbolic reasoning
3. 1974 Lighthill Lighthill AI
4. 70 DARPA AI 1969 mission-orient direct research
5. SUR debacleDARPA CMU speech understanding research ( ) CMU
6. 1987 LISP 7. 90 8. 5
5 8.5
AI ?
-
5 3 AI
-
2 AI
AI 1989 Berners-Lee (web ) 1990 1993 1998
/
(Google) - : $100,000 (1998)2015: $517,170,000,000
-
?
YES/NO (binary classifier)
?
.
.
-
(rule) , (frequent)
( )
(Nearest neightbor) (Nave Bayes) (Decision tree) (Support vector machine) (Neural networks)
-
Nave
Nave Bayes .
: = 1:10 : = 100:1
log(10/1) log(1/1000)
How To Build a Naive Bayes Classifierhttps://bionicspirit.com/blog/2012/02/09/howto-build-naive-bayes-classifier.html
-
Y xi
, https://ko.wikipedia.org/wiki/_
-
(SVM, Support Vector Machine)
(maximum-margin hyperplane)
SVM (x y ) (w ) (X+,X-)
X+ X- (margin) = 2/||W||
Yi = 1 Yi = -1
-
(Neural Network)
(sigmoid)
(threshold) on/off
( )
on/off :
-
, ,
- : MNIST 28x28 pixel (784-pixel) 7
( 100) 784x100 + 100x10 = 79,400 ()
Backpropagation
()
MNIST data
0 1 2 3 4 5 6 7 8 90.05 0.05 0.07 0.40 0.05 0.10 0.06 0.03 0.14 0.02
-
feature() feature
168 2500
155 7000
183 12000
175 4000
174 1800
163 50000
. ?
=
. .
-
symbol grounding . . (feature design) .
?
(signfi, ) (signifiant, ) ( )
.
.
-
6. 3 AI
-
2012 ILSVRC(ImageNet Large Scale Visual Recognition Challege) (SuperVision)
Team name Filename Error (5 guesses) Description
SuperVision test-preds-141-146.2009-131-137-145-146.2011-145f. 0.15315 Using extra training data from ImageNet Fall 2011 release
SuperVision test-preds-131-137-145-135-145f.txt 0.16422 Using only supplied training data
ISI pred_FVs_wLACs_weighted.txt 0.26172Weighted sum of scores from each classifier with SIFT+FV, LBP+FV, GIST+FV, and CSIFT+FV, respectively.
ISI pred_FVs_weighted.txt 0.26602 Weighted sum of scores from classifiers using each FV.
ISI pred_FVs_summed.txt 0.26646 Naive sum of scores from classifiers using each FV.
ISI pred_FVs_wLACs_summed.txt 0.26952Naive sum of scores from each classifier with SIFT+FV, LBP+FV, GIST+FV, and CSIFT+FV, respectively.
Task 1()
?
SuperVision : -
SuperVision:
Deep Learning
Geoffrey Hinton
-
(Deep learning)
minor change
-
(auto-encoder)
Code =
(code) .
-
: 10 . A B 5 . B .
10
9
8
11
12
15
3
17
16
13
A
10
9
8
11
12
B
1 . B
10
12
15
3
17
2 1
?
9.5
12
15
3
15.3
?
-
(auto-encoder) A B
: 28x28 (784) 100
(representation)
= (representation learning) (Geoffrey Hinton) - ()
Auto-encoder
- - -
-
() End-to-end learning
1: 100 2: 1 100 2 20
20 (code) (represent)
784 nodes
784 nodes
100 nodes
100 nodes
100 nodes
20 nodes
-
2 = 1
3 = 2
4 = 3
-
=
(, signfi) (signifiant)
= () : , ()
-
,
Noise
Drop-out
Regularization
.