New era Paradigm Shift Intelligence Intelligence Intelligence Service

목차 New era Paradigm Shift Intelligence 정의 Intelligence의 중요성 Intelligence Service Platform

New era is coming. .

Computer outperforms human 2014. 8. 20 Hongkong University 최근 결과 99. 15% (Dubbed Gaussian. Face 알고리즘 성능) 인간의 인식률 97. 53% 3/57

Computer starts to hear from you 6/57

Paradigm Shift

지식의 증폭 지식 = f (Knowledge Data, Algorithm) Engine #N Algorithm Engine #3 KB Repo. Engine #2 Engine #1 KB Repo Algorithm User Knowledge Common Sense Knowledge 15/57

지식 기반 가치 창출 Intelligence Platform, Knowledge graph, Machine learning, Data Mining, Analytics Value Creation 17/57

IBM Watson “Watson에 IBM의 미래가 있다” – Ginni Rometty, IBM CEO – - Jeopardy Show - 18/57

Intelligence란?

Big Data와 Intelligence의 다른점

Big Data vs Intelligence Big Data Analytics Cognition Inference 25/57

Enabling Technology l Natural Language Understanding § SVM (Support Vector Machine) § CRF (Conditional Random Field) l Machine Learning § GMM (Gaussian Mixture Model) § MLP (Multi-Layer Perceptron) § ELM (Extreme Learning Machine) l Deep Learning § DBN (Deep Belief Network) § DNN (Deep Neural Network) § CNN (Convolutional Neural Network) 27/57

Deep Learning = Learning Representations/Features ■ 전통적인 Pattern Recognition 접근 방법 - Fixed/engineered features + trainable classifier - 주로 Computer Vision 기반의 접근 방향 기존 Hand-crafted Feature Extractor “Simple” Trainable Classifier “이미지의 특징을 찾는 알고리즘은 사람이 직접 연구해서 개발함” ■ Deep Learning의 접근 방법 Deep Learning - Trainable features + trainable classifier - Machine Learning 기반의 접근 방향 Trainable Hand-crafted Feature Extractor “Simple” Trainable Classifier “특징(feature/representation)을 찾는 방법을 Training을 통해서 찾음” 28/57

Gartner Hype Cycle | Neural Network Expectations SVM Speech Recognition (DNN+DBN) DBN (Geoffrey Hinton) Technology Trigger 1950~70 1980 Peak of Inflated Expectation 1990 Trough of Disillusionment 2000 2006 Slope of Enlightenment 2009 Plateau of Productivity 2014 Time 29/57

Intelligence의 응용 분야

Intelligence의 응용 | Io. T 31/57

Intelligence의 응용 | Object Recognition 32/57

Intelligence의 응용 | Personal Assistant Meeting 33/57

Intelligence의 응용 | Product Recommendation 34/57

Intelligence의 응용 | Machine Translation MS, at Code Conference, May 27 2014 Real-time machine translation for Vo. IP 35/57

Technical Enabler: Cloud Real-Time Service 및 전문화/정밀화 향상 Cloud Computing 41/57

Technical Enabler: Big Data Growth 800배 증가, 저장비용 250배 감소 42/57

Technical Enabler: Io. T Connected Device 50배 증가, 가격 75% 감소 43/57

Technical Enabler: 고성능CPU Tianhe 2 (Milkyway-2) 中國 No 1 since 2013 28. 5 Peta. Flops Titan: Oak Ridge National Laboratory No. 1 in Nov 2012 44/57

Intelligence Service Platform Grand Bleu (그랑블루)

Intelligence 시대 Value Intelligence : next gen. paradigm Personalized Service OS Platform 시대 Service Expert Service Data Knowledge TV, Phone Smart Device Intelligent Device + Internet of Things Past Present Future HW/Embedded 시대 46/57

Grand Bleu Architecture Online System Dialog Management Multi. Fan-in Out Offline System Proactive Event Decision NLU Model Generation Domain KB Extraction NLP/NLU Q&A Search Device Online Offline Data Comput. Data Storage Data Acquisition Data Processing System Service/App FW Runtime Adaption FW Personal Knowledge Extraction Management Device Pipeline FW Data Computation Data Storage Data Cluster 48/57

Grand Bleu Data Flow User 사용 Online System Offline System 생성 Raw Data Source Data Processing System (Computation & Storage) Grand Bleu 49/57

Grand Bleu Example | 음성인식 50/57

Grand Bleu Example | 전문가 시스템 52/57

Grand Bleu | Open Innovation Intelligence 자문단 ML/DL 전문가 Cloud 자문단 삼성전자 소프트웨어센터 해외연구소 (Allen Institute, Fraunhofer, KIT등) 산학 프로그램 / 과제 Visiting Scholar 53/57

Grand Bleu Vision User Behavior 기업 Big Data 서비스 지식증폭 Cloud Knowledge 54/57

별첨 1 - Machine Learning Model 구분 Shallow Deep 감사합니다 56/57