실습 강의 개요와 인공지능, 기계학습, 신경망 <인공지능 입문> 강의 허 민 오 Biointelligence Laboratory School of Computer Science and Engineering Seoul National University
실습강의개요와인공지능, 기계학습, 신경망
<인공지능입문>강의
허 민 오
Biointelligence LaboratorySchool of Computer Science and Engineering
Seoul National University
실습강의개요
노트북을꼭지참해야하는강좌
© 2018, 인공지능입문, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr
신경망소개(2주, 허민오)
Python
(프로그래밍언어)
(2주, 김준호)
Python으로
신경망 다뤄보기
(2주 , 김준호)
딥러닝소개(2주 , 허민오)
Tensorflow
(딥러닝라이브러리)
(3주, 류제환)
Tensorflow로
분류문제 풀어보기
(2주, 허유정)
실습을진행할어벤저스
© 2018, 인공지능입문, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr
허민오 김준호 류제환 허유정
한동식
Python과신경망
담당자:김준호
3~6주차계획1. 환경세팅및 Python Introduction
-실습과제: Codecademy2. Numpy
-실습과제: Codecademy3. Perceptron
-실습과제:주어진데이터 classification4. Multi-layer Perceptron (MLP)& Backpropagation
-실습과제: Backpropagation코드구현
© 2018, 인공지능입문, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr
https://www.codecademy.com
© 2018, 인공지능입문, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr
Tensorflow
담당자: 류제환
10~12주차계획1. Tensorflow란무엇인가?2. Tensorflow의구성요소3. 기본적인 Tensorflow의연산들
1. 기계학습평가방법론소개2. Tensorboard소개
1. Tensorflow로 MLP 만들기
© 2018, 인공지능입문, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr
Tensorflow로분류해보기: MNIST 데이터
담당자: 허유정
13~14주차계획1. MNIST 데이터2. matplotlib으로영상데이터확인하기3. Convolutional Neural Network
1. Tensorflow로MNIST 분류기코드읽기 / 사용하기
© 2018, 인공지능입문, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr
Final Project: CIFAR-10 물체사진분류기
딥러닝실전프로젝트Tensorflow를써서 10 가지물체사진분류하는도구를만들고분석하기데이터집합: CIFAR-10 (https://www.cs.toronto.edu/~kriz/cifar.html)
© 2018, 인공지능입문, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr
실습평가방법
실습점수는수업전체평가점수중 30%매시간평가 ( 20% )신경망소개, 딥러닝소개구글서베이를통해퀴즈풀어제출(수업종료 5분전에 링크공개, 5분동안문제풀기수업마치고 5 ~ 20분후에제출불가로변환됩니다.)기타실습
코드작성후에조교확인받기
Final 프로젝트 ( 10% )© 2018, 인공지능입문, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr
질문있나요?
© 2018, 인공지능입문, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr
인공지능, 기계학습, 신경망
© 2018, 인공지능입문, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr
인공지능(Artificial Intelligence)인공지능(AI): “사람처럼생각하고사람처럼행동하는기계”(컴퓨터, SW, 로봇)사람이기계보다잘하는일을기계가할수있도록하는연구
지능을필요로하는일을기계가할수있도록하는연구
1950: Turing Test, 1956: “Artificial Intelligence (AI)”
13
1970-1980년대: 붐전문가/지식기반 시스템
1982-1992:제5세대 컴퓨터계획 (FGCS)
1990년대: 암흑기뉴럴넷, 유전자 알고리즘, 퍼지로직
1990대 후반:인터넷, 웹, 전자상거래정보검색, 데이터마이닝아마존, 이베이, 야후, 구글
2010년대: 부흥기• 지능형 에이전트• 머신러닝/딥러닝
IBM “Deep Blue” Chess Machine Beats Human Champion (1997)
AI의역사적흐름
Grand Challenges of AI: Thinking Machines
14
Deep Blue Watson AlphaGo
1997 2011 2016
Why is AI difficult?
환경과의상호작용에필요한것은?-적절한행동 (Decision making + body manipulation)-지각능력 (Perception)
A thinking machine? An acting machine?
Self-driving Cars: Acting machine?
16
DARPAGrand Challenge
GoogleSelf-driving
Car
RHINOMuseum Tour Guide
1997 2005 2010
핵심인공지능기술:기계학습(Machine Learning)
사람처럼 “경험으로부터학습하는기계”를개발축적되는데이터로부터스스로성능을향상하는시스템
데이터로부터모델(프로그램,패턴/규칙,지식)을자동생성하는기술자동프로그래밍,패턴인식,지식발굴/습득
17
AI, 기계학습, 딥러닝
Artificial Intelligence
Machine Learning Knowledge Representation- Memory- Reasoning- …
Action (Body Manipulation)- Decision making- Planning- …
Perception- Vision- Language- …
Deep Learning
18
What is Changed?
IDC’s Data Age 2025 study
19
What is Changed?
20출처: https://en.wikipedia.org/wiki/Big_data
Where does big data come from?
What is Changed?
Deep learningMajor advantage of deep learning: scalability
(C) 2007-2018, SNU Biointelligence Lab, http://bi.snu.ac.kr/ 21
What is Changed?
GPU(Graphics Processing Unit)Many slow cores (thousands) Originally for graphicsGood at parallel computation
22
딥러닝성공사례 -물체인식
• 심층 컨볼루션 신경망 (Deep Convolutional Neural Network, CNN)
• 이미지에서특징(feature)을자동으로추출함• 높은층으로갈수록더복잡하고종합적인인식
• ImageNet• CNN으로이미지에서다양한종류의물체를인식함• 약 6천만개의매개변수(parameter), 65만여개의인공신
경세포를이용해 1천종류이미지약 120만장을분류• 인간 수준(이상)의 물체인식
23
24ex) 120 breeds of dogs
https://arxiv.org/abs/1409.0575
Face Identification (Facebook)
(C) 2007-2017, SNU Biointelligence Lab, http://bi.snu.ac.kr/ 25
[Y. Taigman et al., CVPR 2014]
음성인식
~2010 GMM-HMM (Dynamic Bayesian Models)~2013 DNN-HMM (Deep Neural Networks)~Current LSTM-RNN (Recurrent Neural Networks)
26
Image Captioning
(C) 2007-2017, SNU Biointelligence Lab, http://bi.snu.ac.kr/ 27
[X. Kelvin et al., ICML 2015]
Lip Reading in the Wild
(C) 2007-2017, SNU Biointelligence Lab, http://bi.snu.ac.kr/ 28
Neural machine translation
(C) 2007-2017, SNU Biointelligence Lab, http://bi.snu.ac.kr/ 29
김상경, Naver Labs, DEVIEW 2016
[Ilya Sutskever et al., NIPS 2014]
Image-to-image translation
Conditional Adversarial Networks
(C) 2007-2017, SNU Biointelligence Lab, http://bi.snu.ac.kr/ 30
Visual Question-Answering
(C) 2007-2017, SNU Biointelligence Lab, http://bi.snu.ac.kr/ 31
Question
ImageAnswer
J.-H. Kim et al., NIPS 2016
딥러닝의성공적적용을위한요소
데이터다루는문제의복잡도를충분히채울만큼의많은데이터
SW 기술: 딥러닝기술 + 알고리즘기술
하드웨어CPU / GPU병렬연산기술 / 분산컴퓨팅기술
32
딥러닝이잘다루는문제
딥러닝이잘다루는문제데이터를표현하는인자들내에복잡성요소가포함됨
예) 영상데이터, 음성데이터, 언어데이터, ……
큰분량의데이터확보가가능한문제
상당한노이즈가있어도데이터분량이크면다룰수있음
표지(label)가있는데이터현재기술수준에서는아직까지는 supervised learning을더잘함Label이일반적인분류문제의 label일필요는없음cf) Image captioning, Neural machine translation(NMT), image-to-image translation
33
딥러닝이뭐길래?딥러닝: Deep Neural Networks를이용한기계학습방법 차후수업시간에다룸
기존접근법과의차이
기존방법: 데이터전처리및가공을통해문제해결에적합한특징추출후이를학습데이터로패턴분류기를훈련
딥러닝: 특징추출을위한전처리단계를 (무감독학습) 전체학습프로세스에포함. 특징맵(feature map) 또는표상(representation)을자동으로학습함
34
AI, 기계학습, 딥러닝
Artificial Intelligence
Machine Learning Knowledge Representation- Memory- Reasoning- …
Action (Body Manipulation)- Decision making- Planning- …
Perception- Vision- Language- …
Deep Learning
35
AI, 기계학습, 딥러닝
Artificial Intelligence
Machine Learning Knowledge Representation- Memory- Reasoning- …
Action (Body Manipulation)- Decision making- Planning- …
Perception- Vision- Language- …
Deep Learning
36
AI, 기계학습, 딥러닝
Artificial Intelligence
Machine Learning Knowledge Representation- Memory- Reasoning- …
Action (Body Manipulation)- Decision making- Planning- …
Perception- Vision- Language- …
Deep Learning
Neural language modelWord2vec
GloveThought vector
37
AI, 기계학습, 딥러닝
Artificial Intelligence
Machine Learning Knowledge Representation- Memory- Reasoning- …
Action (Body Manipulation)- Decision making- Planning- …
Perception- Vision- Language- …
Deep Learning
Neural Language modelWord2vec
GloveThought vector
38
딥러닝과인공지능
주변사용자와의상호작용에필요한것은? 사람과의상호작용에필수적인기술
보기
읽기/듣기
보여주기
쓰기/말하기
사람을대신할수도있게되는가?
39
Robot & Communication
(C) 2007-2016, SNU Biointelligence Lab, http://bi.snu.ac.kr/ 40
Human Need Not Apply
www.CGPGrey.com
41
질문있나요?
© 2018, 인공지능입문, SNU CSE Biointelligence Lab., http://bi.snu.ac.kr