Embed Size (px)
DESCRIPTION
Contents Requirements for Human Visual Processing Two Extreme Kinds of Dynamical System Topics for Further Investigation The Variety of Possible Architectures Dynamical Systems for Symbolic AI Types of Architecture Layered Perception and Action How Evolution Produced Mathematicians? Conclusion 3
Citation preview
Chapter 12. Some Requirements for Hu-man-Like Robots
in Creating Brain-Like Intelligence, Aaron Solman.
Course: Robots Learning from Humans
Hur, Woo-Sol
Vehicle Intelligence LaboratorySchool of Electrical and Computer Engineering
Seoul National University
http://vi.snu.ac.kr
VEHICLE INTELLIGENCE LAB
2
Contents
Introduction
The Seduction of Embodiment
Fallacies in Nouvelle AI
Limitations of Symbolic AI
Meta-semantic and Exosomatic Ontologies
Morphology and Development
3
Contents Requirements for Human Visual Processing
Two Extreme Kinds of Dynamical System Topics for Further Investigation
The Variety of Possible Architectures Dynamical Systems for Symbolic AI Types of Architecture Layered Perception and Action
How Evolution Produced Mathematicians?
Conclusion
Some Questions
What animals did you see?
What was in the last picture?
What was in the picture taken at dusk?
Did you see any windows?
Did you see a uniformed official?
Did anything have horns? What?
Did anything have hands on the floor?
Requirements forHuman Visual Processing
Human process photographs
at a rate of one per second.
No known mechanism exists
for human visual processing.
Reflection on a wide range of phenomena has led to
a hypothesized architecture with a complex system.
Two Extreme Kinds ofDynamical System
■
1 / 26
a simpler one
assumed to build
“biologically inspired” robots
key feature is close coupling
between internal states and
the environment
Two Extreme Kinds ofDynamical System
more complex one
Topics for further investigation
how to build dynamical systems
how those systems are constructed over many years
how the constraint-propagation works
how the processes are influenced by various aspects
how the resulting percepts continue to be driven
how all of that is used
The Variety of Possible Architectures
Reminder Symbolic AI
Symbol grounding Kant’s empiricism Learning of robots to discovery of statistical
patterns relating sensory and motor signals
Nouvelle AI (Brook) Emphasis embodiment and sensory-motor interactions with the
environment Dispense with symbolic representations (But not fully rejected) Use morphology to reduce the SW sophistication
The Variety of Possible Architectures:Dynamical Systems for Symbolic AI
Existing symbolic AI systems are clearly nowhere near
human competence except in very narrow domains: chess (not Go), Mathematica, Matlab, compiler optimization, …
Symbolic AI will suffice for everything often fail to attend
to the kinds of intelligence required for controlling continuous
actions in a 3-D structured environment: maintaining balance while pushing a broom,
drawing a picture with a pencil, …
The Variety of Possible Architectures:Dynamical Systems for Symbolic AI
Nevertheless,
it should be clear that the “traditional”
sense→think/decide→act loop
(presented in several AI textbooks) is
much too restrictive to accommodate
the requirements presented before.
The Variety of Possible Architectures:Types of Architecture
The CogAff schema
The Variety of Possible Architectures:Types of Architecture
The H-CogAff architecture
The Variety of Possible Architectures:Layered Perception and Action
Vision operates at different levels of abstraction:
How Evolution Produced Mathemati-cians?
How Evolution Produced Mathemati-cians?
How Evolution Produced Mathemati-cians?
From Empirical to Mathematical Truths
turning around, counting objects, rubber band, …
Empirical discovery becomes a necessary truth.
It’s not a matter of probabilities.
different ways can be exist for human
to discover useful affordances
Conclusion
Much of the work on embodied cognition in robots
has focused on the terribly narrow problem of learning.
Symbolic computation was a human
competence long before AI began, so
there must be biological mechanisms
that make it possible.
Conclusion
It will be necessary for AI researchers to abandon
factional disputes and stop claiming of one key mechanism.
A synthesis of complementary approaches
might be essential to progress in the long run
Q & A
■
26 / 26
References 인공지능은 뇌를 닮아 가는가 , 유신 저 , 컬처룩 , 2014
Aaron Sloman, “Some Requirements for Human-Like Robots:
Why the Recent Over-Emphasis on Embodiment Has Held Up
Progress”, Creating Brain-Like Intelligence, pp. 248-277,
Springer (2009)
“Artificial Intelligence”, Wikipedia, Sep. 2015, https://
en.wikipedia.org/wiki/Artificial_intelligence
