Menschliche und künstliche „Intelligenz“ · Digital Transformation Enables + Needs AI 60/70s:...
Transcript of Menschliche und künstliche „Intelligenz“ · Digital Transformation Enables + Needs AI 60/70s:...
Berlin / Bremen / Kaiserslautern / Oldenburg / Osnabrück / Saarbrücken
German Research Center for
Artificial Intelligence
Menschliche und künstliche „Intelligenz“
Zur Natur menschlichen Denkens und den Möglichkeiten technischer Systeme. Wie wirken künstliche Systeme auf die „Natur“ menschlichen Denkens und Handels zurück?
Prof. Dr. habil. Jana Koehler
1
DFKI
The Beginnings of Artificial Intelligence
2
Alan Turing: Can Machines Think?
A. M. Turing (1950)
Computing Machinery and Intelligence
Mind 49: 433-460
DFKI
My Working Definition
3
Intelligence
… is the ability to define and set goals
and to develop behavior to achieve these
goals …
Build
Intelligent Agents
Understand
Human Intelligence
Artificial
Intelligence
Cognitive
Science
DFKI
Continuous Progress in Key AI Fields
4
https://beamandrew.github.io/deeplearning/2017/02/23/deep_learning_101_part1.html
DFKI
1992
BackgammonTicTacToe/Checkers Chess
1996/971940s-1970s 2012
Jeopardy
2016
Go Poker Starcraft II
Watson AlphaGo Libratus
Reinforcement
Learning
Monte-
Carlo Tree
Search
Deep
Learning
2019
AlphaStar
2017
Deep Blue
Discounted
Regret
Minimization
Alpha Beta
Pruning Long-/Shortterm
Planning
Policy
Distillation
105/1031 10144 10123 10360Imperfect Information Games
Temporal
Difference
Learning
TD-Gammon
Conflict
Learning
Heuristic Search
and Knowledge
Systematic
Search
© Jana Koehler
> 20 People > 40 People
5
Population-based/
Multi-Agent RL
DFKI
The PENDULUM of AI Methods
6
Nonmonotonic Reasoning
Satisfiability
Mathematical
Optimization
Reasoning about
ActionComputational
Science
Statistics
Constraint Reasoning Model Checking
Planning/Scheduling
Probability
Theory
Bayesian Networks
Neural Networks
Deep Learning
Markov Models
Reinforcement Learning
Ontologies/Knowledge Graphs
DFKI
Digital Transformation Enables + Needs AI
60/70s: selected data in small amounts
- support business processes
since 00s: arbitrary data in massive amounts
- revolutionize business models and processes
= Business Transformation leveraging digital capabilities
= Customer experience + Operational excellence
= Efficiency + Flexibility
= Talents and Data as strategic assets
7
DFKI
DFKI – Public-Private Partnership for Innovation
8
Est. 1988
34 Professors
> 1100 People
50 M€ in Projects
25 % from Industry
372 Ongoing Projects
Turnover: 80.2 M€ in 2018
Operating Income: 55.2 M€ in 2018
Saarbrücken
Berlin
Bremen
Osnabrück
Kaiserslautern
OldenburgDeutschland GmbH
DFKI
Algorithmic Business and Production @ DFKI
9
“Big data” is high-volume, -velocity and -variety information assets that demand
cost-effective, innovative forms of information processing
for enhanced insight and decision making
Gartner 2011,2016
“Algorithmic business is pivotal to competitive advantage … causing the disruption of entire industries.”
DFKI
Major Technologies & Application Areas
10
Natural
Language
Understanding
Intelligent
Agents
Knowledge
Representation
& Reasoning
Search
Algorithms
Game &
Decision
Theory
Machine
Learning
Planning &
Reasoning
about
Action
Understanding
Human
Behavior
Image/
Video
Analysis
Adaption &
Flexibility
Automatic
Translation
Search &
Information
Retrieval
Insights &
Discovery
OptimizationPrediction
Assistants
& Robots
Decision
Making
DFKI
Hype and Reality: Intelligent Autonomous Systems Intelligent Agents
11
Sensors Intelligence
Actuators
DFKI
Hype Today: Last Mile Autonomous Deliveryfor Everybody
Failsafe systems with humans in the loop
Operate flawlessly in the wild
Understand the world
Make complex decisions in real time
Perform on par with humans, exceed them
Needs more sophisticated on-board sensors and processors artificial intelligence (AI) algorithms controllers and communication systems battery charging systems (ground/flight)
12
DFKI
Reality Today: Drone-Assisted Photography & Surveying
Construction, insurance, facility management, agriculture, …
More data Cheaper
Faster
Better
Precise Decisions
Controlled Actions
13
Future: Last-mile logistics for high value commodities
DFKI
Discovery and Action in Areas Inaccessible to Humans
FlatFish from DFKI Bremen
In operation for Shell forpipeline inspection in Brasil
Many such areas at themacro and micro level!
14
DFKI
Current Limitations of Language-based Interaction
«Understanding» questions that require context
Conducting dialogues for clarification
Recognizing arbitrary names
Differentiating between concrete and abstract terms
15
DFKI
Future: Mastering Context in Dialogue
16
DFKI
Challenges in Deep Learning
17
Need to resolve:
Hidden dependency on training data
Non-calibrated confidence values
Intransparancy of learned model
DFKI
Reinforcement Learning
18
20 % success
rate for
arbitrary cube
60 % success
rate for 15
moves
problem
80 % dropped
cubes
DFKI
Transfer Learning
19
https://www.youtube.com/watch?v=iaF43Ze1oeI
DFKI
Finding Unusual Strategies
Atari Breakout after 600 training episodes
20
DFKI21
https://www.wired.com/story/the-genderless-digital-voice-the-world-needs-right-now/
DFKI
Secondary Effects of Innovation are more Disruptive than the initial Digital Change
22
https://www.youtube.com/watch?v=MnrJzXM7a6o http://disruptiveinnovation.se/?cat=16
DFKI
Secondary Effects: The Challenge of Resource Efficiency
23
DFKI
Freedom to Operate in a Competitive AI Market
24
DFKI
Europe’s Greatest Risk: The No-Sayers
25
The doctor´s
opinion is much
more differentiated
than the press
headline
Tagesanzeiger
15.11.2018
DFKI
The Future of Work – Are You Ready?
26
DFKI
Five Key Success Factors for a Digital Nation
Infrastructure
Social Stability and Equity
Life-Accompagnying Education and Learning
High-Quality Affordable Health Care
Business-friendly public processes predictable, fair, efficient
27
DFKI
Current Research Trends
Multi-AI Architectures
Mastering Context
Embodiment
Scientific Discovery with AI
28
Integrated Intelligence
• Contextualized AI• Open knowledge repositories
• Understanding human
intelligence
Self-Aware Learning
• Robust and trustworthy learning
• Deeper learning for challenging tasks
• Integrating symbolic andnumeric representations
• Learning in integratedAI/robotic systems
Meaningful Interaction
• Collaboration
• Trust and responsibility• Diversity of interaction channels
• Improving online interaction
A 20-Year Community
Roadmap for Artificial
Intelligence Research
in the US
Yolanda Gil (USC) and Bart
Selman (Cornell), co-chairs
DFKI
AI Will Master the Four Big „A“
Abstraction
Analogy
Argumentation
Arbitrary Common Sense
29
Thank you very much for your attention.
Berlin / Bremen / Kaiserslautern / Oldenburg / Osnabrück / Saarbrücken
German Research Center for
Artificial Intelligence GmbH
Registered Office Kaiserslautern Site
Trippstadter Straße 122
D-67663 Kaiserslautern
Germanyphone: +49 631 205 75 0
email: [email protected]
www.dfki.de