AI = Alien Intelligence

Transcription

1 AI = Alien Intelligence Dr. Aljoscha Burchardt German Research Center for Artificial Intelligence AI Monday Berlin, September 17, 2018

2 Changes (c) Burchardt, DFKI AI Monday Berlin 2

3 The temporal dimension (c) Burchardt, DFKI AI Monday Berlin 3

4 The Four Phases of AI Research Degree of Intelligence Phase 4 : since 2015 Cognitive Systems Combination of machine learning with knowledge technologies Phase 3: until 2015 Learning Systems Machine learning from mass data AI Winter Phase 2: until 1995 Phase 1: until 1970 Knowledge-Based Systems Heuristic Systems Knowledge processing with manually creatied knowledge bases Heuristic search and reasoning methods Autor: W. Wahlster, mod.: A. Burchardt (c) Burchardt, DFKI AI Monday Berlin 4

5 DFKI as the World s Largest PPP for AI Research Deutschland GmbH Bremen Osnabrück Berlin 512 R&D Staff, 29 Professors Revenues 55 Million Euro in 2017 Kaiserslautern Saarbrücken 914 Employees 295 Ongoing Projects DFKI as a career springboard: > 80 CEOs of AI Start-ups, > 90 AI Professors, > 30 Managers of Industrial AI Departments (c) Burchardt, DFKI AI Monday Berlin 5 W. Wahlster

6 The R&D Departments and Groups of DFKI Innovative Factory Systems Embedded Intelligence Interactive Textiles Agents and Simulated Reality Multilingual Technologies Cyber-Physical Systems Digital Business Robotics Innovation Center Intelligent User Interfaces Intelligent Analytics for Massive Data Intelligent Networks Language Technology Augmented Vision Educational Technology Smart Data & Knowledge Services Plan-Based Robot Control (c) Burchardt, DFKI AI Monday Berlin 6

7 DFKI Has Created 3200 New Jobs in IT Industry and More than 80 Startup and Spin-Off Companies schwartz&stahl (c) Burchardt, DFKI AI Monday Berlin 7

8 SELECTED SPIN-OFFS OF DFKI BERLIN (c) Burchardt, DFKI AI Monday Berlin 8

9 acrolinx (c) Burchardt, DFKI AI Monday Berlin 9

10 Leverton (c) Burchardt, DFKI AI Monday Berlin 10

11 Debatoo DEBATOO FOR EVERYONE DEBATOO OPEN DEBATOO ENTERPRISE DEBATOO PUBLIC DEBATOO CUSTOM Kostenfreies Diskussionstool für jedermann. Egal ob im Familienkreis, unter Freunden, Kollegen, im Chor oder Fußballverein: Debatoo for Everyone ermöglicht Diskussionen und Abstimmungen mit bis zu 50 Teilnehmern Diskussionsplattform im Internet. Ständig neue Diskussionen zum aktuellen Zeitgeschehen Tool für strukturierte Diskussionen in Teams. Automatisierte Zusammenfassung und Entscheidungsvorbereitung in Echtzeit Online-Plattform zur verlässlichen und transparenten Unterstützung öffentlicher Infrastrukturvorhaben und anderer Formen der Bürgerbeteiligung Online-Plattform und Tools, die zielgenau auf den Bedarf der Kunden zugeschnitten sind bzw. besonderen Vorgaben entsprechen Illustrationen: Anatoliy; snyggg; alexyndr; martialred; Marc / istockphoto r DEBATOO-Diskussionsplattform (c) Burchardt, DFKI AI Monday Berlin 11

12 Disrupting German Economy by Injecting AI: Transforming Premium Products Into Smart Products and Smart Services Medical Equipment Cars Manufacturing Equipment Agricultural Machinery (c) Burchardt, DFKI AI Monday Berlin 12 Home Appliances W. Wahlster

13 End-to-End Demonstration in Seven Living Labs (c) Burchardt, DFKI AI Monday Berlin 13

14 Today Goal: The 2nd Wave of Digital Transformation Means: Artificial Intelligence Method: Maschine Learning (c) Burchardt, DFKI AI Monday Berlin 14

15 DIGITAL TRANSFORMATION (c) Burchardt, DFKI AI Monday Berlin 15

16 AI for the Second Wave of Digital Transformation First Wave: Data are digitally - acquired - stored - transmitted - processed Second Wave: Data are digitally - interpreted - enhanced - actively used - monetized Machine readable data: Internet and cloud technologies Machine unterstandable data: AI and machine learning (c) Burchardt, DFKI AI Monday Berlin 16 Author: W. Wahlster

17 ARTIFICIAL INTELLIGENCE (c) Burchardt, DFKI AI Monday Berlin 17

18 Working Definition Artificial Intelligence (AI) is the property of IT systems to demonstrate abilities similar to human cognition. The single ability can exist only to some extent, such as the dialog capability of today's smartphones. However, it can also show a performance way beyond human capabilities as in scanning tens of thousands of MRI scans. AI systems have different proportions of certain core skills such as situational perception, planning, communication, acting, reasoning or learning. (Pre-)Translated with (c) Burchardt, DFKI AI Monday Berlin 18

19 Assistance for physical and mental work Assistance with physical work by collaborative AI robots Assistance with mental work through cognitive AI systems e.g.: AILA collaborative DFKI robot Man-robot collaboration, standby robots, team robotics e.g.: DFKI Spin-Off LEVERTON using Deep Learning in Legal Tech Banking advice, legal advice, personnel consulting, etc (c) Burchardt, DFKI AI Monday Berlin 19

20 Digital Transformation vs. AI I never want my decisions to be dominated by a machine (c) Burchardt, DFKI AI Monday Berlin 20

21 From Information to Knowledge: Who is Wolfgang Dauner? (c) Burchardt, DFKI AI Monday Berlin 21

22 MACHINE LEARNING (c) Burchardt, DFKI AI Monday Berlin 22

23 How to make machines smart Output Output Output Computer Program Algorithms, Heuristics Programmer creates Software Knowledge Processing Search, Inference, Planning Knowledge Base Facts, Rules, Models Deep Machine Learning Feature Extraction, Pattern Recognition, Feature Mapping Mass Data Training Data, Meta Data, Test Data Input Input Input Bottleneck: Programmer high development costs restriced adaptability weak explanation capabilities Bottleneck: Knowledge Base huge development effort high maintenance costs good explanation capabilities Bottleneck: Training Data minimal development effort good adaptability no explanation capabilities W. Wahlster (c) Burchardt, DFKI AI Monday Berlin 23

24 1. Programming Output Computer Program Algorithms, Heuristics Programmer creates Software Input Bottleneck: Programmer high development costs restriced adaptability weak explanation capabilities Bild: (c) Burchardt, DFKI AI Monday Berlin 24

25 2. Building up Expert Knowledge Towards Self-Learning Systems Output Output Computer Program Algorithms, Heuristics Programmer creates Software Input Bottleneck: Programmer high development costs restriced adaptability weak explanation capabilities (c) Burchardt, DFKI AI Monday Berlin Output Knowledge Processing Knowledge Base Deep Mach Learnin Search, Inference, Planning Facts, Rules, Models Feature Extrac Pattern Recog Feature Map Input Bottleneck: Knowledge Base huge development effort high maintenance costs good explanation capabilities Input Bottleneck: T minimal de good adap no explana 25

26 3. Self-Learning Machine Learning wards Systems Output Output Computer Program Algorithms, Heuristics Programmer creates Software Output Knowledge Processing Knowledge Base Deep Machine Learning Search, Inference, Planning Facts, Rules, Models Feature Extraction, Pattern Recognition, Feature Mapping Input Input Automatic clustering for recommendations Machine Translation Control car Bottleneck: Knowledge Base ottleneck: Programmer costs Go Play huge development effort high development Composing like Bach high maintenance costs restriced adaptability weak explanation capabilities good explanation capabilities Mass Data Training Data, Meta Data, Test Data Input Bottleneck: Training Data minimal development effort good adaptability no explanation capabilities (c) Burchardt, DFKI AI Monday Berlin 26 W. Wahlster

27 Learning from humans flower Drive carefully due to the traffic situation Hallo Welt! Hello world! (c) Burchardt, DFKI AI Monday Berlin 27

28 Deep Machine Learning flower Drive carefully due to the traffic situation Hallo Welt! Hello world! (c) Burchardt, DFKI AI Monday Berlin 28

29 Machine Learning Style Imitation A Neural Algorithm of Artistic Style, L. Gatys, A. Ecker, M.Bethge, Universität Tübingen, (c) Burchardt, DFKI AI Monday Berlin 29

30 EXAMPLE: LANGUAGE TECHNOLOGY (c) Burchardt, DFKI AI Monday Berlin 30

31 Language Technology (Almost) Everywhere Spelling Correction OCR (optical character recognition) Web search with Google Language learning programs Machine Translation Recommendations, Did you mean, Voice control and dialogue: Apple Siri, Amazon Alexa,... IBM Watson (c) Burchardt, DFKI AI Monday Berlin 31

32 Chatbots for Customer Support Collavorative development of a chat bot to support call center agents through suggestions from FAQs and (product-) data bases Text analysis for the development of knowledge graphs about products, components, customer problems.? Customers! Human Operator Knowledge Bases Customer Hi Hi Paul How can I help you? Can you explain me how to setup an account on my Moto Z? It s a Gmail account Yes, thanks! System In order to help you I need to know your provider? Ok. Here s what you need to do Did this solve your problem? (c) Burchardt, DFKI AI Monday Berlin 32

33 Ressources and training data Necessary resources Existing dialogues, the more the better Annotated dialogs (e.g. customer requests categorized by CSA) Existing knowledge bases (e.g. documentation, FAQs, call center knowledge base) And training data in the Lenovo project for correct answer from FAQ existing dialogs for intent user turns labelled (ca. 60 types) for question type user turns labelled (3 types) same for slots, etc (c) Burchardt, DFKI AI Monday Berlin 33

34 The Moli Chatbot Types of Dialogs How To Questions Troubleshooting Facts (c) Burchardt, DFKI AI Monday Berlin 34

35 How can I help you? Q1 I dropped my phone and the screen shattered, and the battery hasn't been holding a charge well for a couple of months. I have all of the information to send it in for repair, but I was wondering if a reset would have to be done on it. There are some things on the phone I need to keep if at all possible (c) Burchardt, DFKI AI Monday Berlin 35

36 Central Concepts Digital Transformation Second Wave: From Information Storage to Knowledge Source Artificial Intelligence Learning, language comprehension, planning, interference, situational context Machine Learning Recognizing characteristics and patterns Often on manually edited/created training data (supervised) Powerful, but "reactive", no explanations / decision support, no dynamic interaction (c) Burchardt, DFKI AI Monday Berlin 36

37 Thank you! Dr. Aljoscha Burchardt DFKI, Language Technology Lab Alt-Moabit 91c, Berlin Tel (c) Burchardt, DFKI AI Monday Berlin 37