Difference between revisions of "Four principles of pattern recognition in complex systems"

From Santa Fe Institute Events Wiki

Line 1: Line 1:
Melanie Mitchell
Portland State University and Santa Fe Institute
Modern-day computers have been programmed to do some remarkable things
that most people would consider to be "intelligent": beat grandmasters
at chess, diagnose diseases, crack cryptographic codes, and predict
trends in financial markets, to name a few examples.  So what is
lacking in today's artificial intelligence systems? 
Most practitioners agree that the missing link is sophisticated
pattern recognition, both sensory and conceptual.  The ability to
recognize abstract patterns is a core feature of human intelligence;
the most surprising result coming out of artificial intelligence is
how hard human-level pattern recognition has turned out to be for
computers.  While computer programs have beaten world chess champions
and are better than Ph.D.s at doing probabilistic inference, they
still cannot interpret visual or auditory data , or perceive abstract
similarities, with anywhere near the ability of a two-year-old child.
This is the paradox behind AI pioneer Marvin Minsky's dictum, "easy
things are hard".
Sophisticated pattern recognition is not limited to the brain.  In
this talk I will describe the mechanisms underlying pattern
recognition in three different biological systems, and abstract four
general principles that I claim are key to adaptive pattern
recognition in decentralized systems such as these.  These principles
deal with the representation and transmission of information, the
essential role of randomness, the importance of fine-grained parallel
architectures, and the interplay of bottom-up and top-down processes
in all such systems.  Finally, I will describe how these principles
are inspiring new efforts to develop artificial intelligence systems
that can achieve robust and fluid pattern recognition and learning.
Back to [ Program]

Latest revision as of 15:39, 12 September 2007