인공지능 입문 제7강 : 학습하는 기계 I 서울대학교 컴퓨터학부 & 인지과학/뇌과학 협동과정 담당 교수: 장병탁
목차 7- 머신러닝 3 7-2 머신러닝의종류 0 7-3 감독학습 4 7-4 무감독학습 30 7-5 요약 39 7-6 숙제 4
7- 머신러닝
7- 머신러닝 ) 학습시스템 환경 E 와의상호작용으로부터획득한경험적인데이터 D 를바탕으로모델 M 을자동으로구성하여스스로성능 P 를향상하는시스템 - 환경 E - 데이터 D P - 모델 M ML: D M 장교수의딥러닝, 홍릉과학출판사, 207
7- 머신러닝 or 입력값 x 입력 x 입력 x 2 입력 x 3 비례조정함수 오류역전파 (Error Backpropagation) w w + w i 정보전파가중치 입력층은닉층출력층 활성함수 i i, E wi = w 활성함수 E d i ( w ) 출력 출력비교 2 k k outputs o = ( t f (x) o k 2 ) 목표출력값 t= (Apple) t=0 (Orange)
7- 머신러닝 2) 머신러닝과인공지능 인공지능, 머신러닝, 딥러닝 인공지능 : 사람처럼생각하고사람처럼행동하는기계를만드는연구머신러닝 : 기계가학습을할수있도록하는인공지능연구의한분야딥러닝 : 깊은신경망구조기반의머신러닝 Artificial Intelligence Machine Learning Deep Learning https://rapidminer.com/blog/artificial-intelligence-machine-learning-deep-learning/ 209.0.5
7- 머신러닝 3) 프로그래밍방식과의차이점 일반적인컴퓨터프로그램 사람이알고리듬설계및코딩주어진문제 ( 데이터 ) 에대한답을출력 2 머신러닝프로그램 기계가알고리듬을자동설계 (Automatic Programming) 주어진문제 ( 데이터 ) 에대한답을주는프로그램을출력 장교수의딥러닝, 홍릉과학출판사, 207
7- 머신러닝 4) 머신러닝의중요성 머신러닝이필요한문제 명시적문제해결지식의부재 ( 알고리듬부재 ) 프로그래밍이어려운문제 ( 예 : 음성인식 ) 지속적으로변화하는문제 ( 예 : 자율이동로봇 ) 2 머신러닝더욱중요해지는이유 빅데이터의존재 ( 학습의소재 ) 컴퓨팅성능의향상 ( 고난도학습이가능 ) 서비스와직접연결 ( 비지니스적효과 ) 비즈니스가치창출 ( 회사가치향상 ) 장교수의딥러닝, 홍릉과학출판사, 207
7- 머신러닝 5) 활용사례 머신러닝의다양한활용분야 장교수의딥러닝, 홍릉과학출판사, 207
7- 머신러닝 6) 역사와발전동향 980 985 990 995 2000 2005 200 205 Algorithm: MLP DT SVM PGM CNN Model: 신경망모델 확률통계적모델 딥러닝모델 Data: MNIST PASCAL ImageNet IT Infra: PC 의보급 웹, 데이터마이닝정보검색, 전자상거래 스마트폰 자율주행차 MLP = Multilayer Perceptron, DT = Decision Tree, SVM = Support Vector Machine PGM = Probabilistic Graphical Model, CNN = Convolutional Neural Network
7-2 머신러닝의종류
7-2 머신러닝의종류 ) 머신러닝의종류 Supervised Learning Estimate an unknown mapping from known input and target output pairs Learn f w from training set D = {(x, y)} s.t. f w ( x) = y = f ( x) Classification: y is discrete Regression: y is continuous 2 Unsupervised Learning Only input values are provided Learn f w from D = {(x)} s.t. Density estimation and compression Clustering, dimension reduction fw ( x) = x
7-2 머신러닝의종류 ) 머신러닝의종류 3 Reinforcement Learning Not target, but rewards (critiques) are provided sequentially Learn a heuristic function f w from D t = {(s t, a t, r t ) t =, 2, } s.t. With respect to the future, not just past Sequential decision-making Action selection and policy learning fw s (, a, r ) t t t Zhang, B.-T., Next-Generation Machine Learning Technologies, Communications of KIISE, 25(3), 2007
7-2 머신러닝의종류 장병탁, 차세대기계학습기술, 정보과학회지, 25(3), 2007 학습방법감독학습무감독학습강화학습 학습문제의예인식, 분류, 진단, 예측, 회귀분석군집화, 밀도추정, 차원축소, 특징추출시행착오, 보상함수, 동적프로그래밍 모델구조표현기계학습모델예 논리식 명제논리, 술어논리, Prolog 프로그램 Version Space, 귀납적논리프로그래밍 (ILP) 규칙 If-Then 규칙, 결정규칙 AQ 함수 트리 Sigmoid, 다항식, 커널 유전자프로그램, Lisp 프로그램 신경망, RBF 망, SVM, 커널머신 결정트리, 유전자프로그래밍, 뉴럴트리 그래프방향성 / 무방향성그래프, 네트워크확률그래프모델, 베이지안망, HMM
7-3 감독학습
7-3 감독학습 ) 감독학습문제 ( 분류 ) 데이터 x가주어졌을때해당되는레이블 y를찾는문제 ex) x: 사람의얼굴이미지, y: 사람의이름 ex2) x: 혈당수치, 혈압수치, 심박수, y: 당뇨병여부 ex3) x: 사람의목소리, y: 목소리에해당하는문장 x: n차원벡터, y: m차원벡터대표적인감독학습 ( 분류 ) 알고리듬 Multi-Layer Perceptron (Artificial Neural Network; 인공신경망 ) Decision Tree K-Nearest Neighbor Support Vector Machine
7-3 감독학습 퍼셉트론의구조 https://kr.deductiontheory.com/207/03/blog-post_3.html
7-3 감독학습 퍼셉트론과패턴분류 x2 w x + w2 x2 + b = 0 > 0: < 0: w x w w2 b = w0 x x x2
7-3 감독학습 퍼셉트론학습알고리듬 start: The weight vector w is generated randomly test: A vector x P N is selected randomly, If x P and w x > 0 goto test, If x P and w x 0 goto add, If x N and w x < 0 go to test, If x N and w x 0 go to subtract. add: Set w = w + x, goto test subtract: Set w = w x, goto test P: set of positive examples N: set of negative examples http://ocw.snu.ac.kr/sites/default/files/note/iml_lecture%20%2806%29.pdf,209.0.5
7-3 감독학습 Perceptron Learning x x 2 y x 0 0 0 0 x 0 0 0-0.06 0 0-0. x 2 0.05-0.06 0 0 207, SNU BioIntelligence Lab, http://bi.snu.ac.kr/ RIGHT
7-3 감독학습 Perceptron Learning x x 2 y x 0 0 0 0 x 0 0 0-0.06 0 0-0. x 2 0.05-0.0 0 207, SNU BioIntelligence Lab, http://bi.snu.ac.kr/ RIGHT
7-3 감독학습 Perceptron Learning x x 2 y x 0 0 0 0 x 0 0-0.06 0 0-0. x 2 0.05-0.6 0 0 207, SNU BioIntelligence Lab, http://bi.snu.ac.kr/ RIGHT
7-3 감독학습 Perceptron Learning x x 2 y x 0 0 0 0 x 0 0-0.06 0 0-0. x 2 0.05-0. 207, SNU BioIntelligence Lab, http://bi.snu.ac.kr/ 0 WRONG
7-3 감독학습 Perceptron Learning Fails to fire, x x 2 y so add proportion, x 0, to weights. 0 0 0 x 0 0-0.06 0 0-0. x 2 0.05 207, SNU BioIntelligence Lab, http://bi.snu.ac.kr/
7-3 감독학습 Perceptron Learning x x 2 y = 0.0 x 0 0 0 0 x -0. + 0.0x -0.06 + 0.0x 0 0 0 0 x 2 0.05 + 0.0x 207, SNU BioIntelligence Lab, http://bi.snu.ac.kr/
7-3 감독학습 Perceptron Learning x x 2 y x 0 0 0 0 x 0 0-0.05 0 0-0.09 x 2 0.06 207, SNU BioIntelligence Lab, http://bi.snu.ac.kr/
7-3 감독학습 Perceptron Learning x x 2 y x 0 0 0 0 x 0 0 0-0.05 0 0-0.09 x 2 0.06 0.0 Decrease! 207, SNU BioIntelligence Lab, http://bi.snu.ac.kr/
7-3 감독학습 Perceptron Learning x x 2 y = 0.0 x 0 0 0 0 x -0.05 + 0.0x 0-0.09 + 0.0x0 0 0 0 0 x 2 0.06 + 0.0x 207, SNU BioIntelligence Lab, http://bi.snu.ac.kr/
7-3 감독학습 Perceptron Learning x x 2 y x 0 = 0.0 0 0 0 x -0.04-0.09 0 0 0 0 x 2 0.07-0.06 0 Increase! 207, SNU BioIntelligence Lab, http://bi.snu.ac.kr/
7-3 감독학습 Perceptron Learning x x 2 y = 0.0 x 0 0 0 0 x -0.09 + 0.0x -0.06 + 0.0x 0 0 0 0 x 2 0.06+0.0x 207, SNU BioIntelligence Lab, http://bi.snu.ac.kr/
7-4 무감독학습
7-4 무감독학습 무감독학습
7-4 무감독학습 왜곡 (Distortion) v : data point, X : a set of points Distance from v to X d(v, X) as distance from v to the closest point from X. V = v v n, Squared Error Distortion d V, X = d v i, X 2 /n ( i n) V X
7-4 무감독학습 K- 평균클러스터링 Input: A set, V, consisting of n points and a parameter k Output: A set X consisting of k points (cluster centers) that minimizes the squared error distortion d(v, X) over all possible choices of X K-means clustering algorithm ) Pick a number (k) of cluster centers 2) Assign every data point (e.g., gene) to its nearest cluster center 3) Move each cluster center to the mean of its assigned data points (e.g., genes) 4) Repeat 2-3 until convergence.
expression in condition 2 7-4 무감독학습 5 v: Data points 4 3 x X: Cluster centre 2 x 2 v 0 0 2 3 4 5 x 3 expression in condition Iteration 0
7-4 무감독학습 expression in condition 2 5 4 3 2 x 2 x 0 0 2 3 4 5 x expression in condition 3 Iteration
7-4 무감독학습 expression in condition 2 5 4 3 x 2 0 x 2 0 2 3 4 5 x expression in condition 3 Iteration 2
7-4 무감독학습 expression in condition 2 5 4 3 x 2 0 x 2 0 2 3 4 5 x expression in condition 3 Iteration 3
7-4 무감독학습 Example: 4-cluster data and 4 iterations https://slideplayer.com/slide/239944/, 209.0.5
7-5. 요약
제 7 강요약및정리 학습시스템은환경과의상호작용을통해서관측된데이터로부터모델을자동으로구축함으로써경험을통해서스스로성능이향상되는시스템이다. 머신러닝은기계가경험을통해서학습할수있게함으로써자동으로인공지능시스템을개발하는기술이다. 머신러닝은학습하는문제에따라서감독학습, 무감독학습, 강화학습으로구분된다. 감독학습은입력과목표출력의쌍으로된학습데이터를이용하여입력에서출력으로의사상을학습한다. 퍼셉트론학습알고리듬은주어진입력에대해서출력이오류를범할경우이를교정하도록연결가중치를조정함으로써학습한다. 무감독학습알고리듬은목표출력이없으므로오류함수를사용할수없다. 대신왜곡함수를이용하여주어진데이터점들간의거리 ( 유사성 ) 를측정한다. K-평균클러스터링은무감독학습의예로서, 주어진데이터셋을 K개의클러스터 ( 군 ) 로자동으로그루핑해준다. 클러스터내에있는데이터점들간의거리 ( 왜곡 ) 는클러스터간데이터점들사이의거리보다가깝다. 결과적으로주어진데이터점들을서로유사한패
7-6 과제
Reading (Watching) Assignments ) Learning Machine Learning in 3 Months, Video Lecture, 208. https://www.youtube.com/watch?v=cr6vqtrov0 Siraj Raval, Learn Machine Learning in 3 Months (with curriculum), 208.03.02 Q: 머신러닝을공부하기위한인터넷비디오강좌와정보소스는무엇이있는가? 이비디오에나오는관련정보를조사하여기술하시오.