TR on Attractor Networks

[David Noelle]

The following technical report is now available via both the World
Wide Web and anonymous FTP:

      http://www.cse.ucsd.edu/users/dnoelle/publications/tr-s97/
      ftp://ftp.cs.ucsd.edu:/pub/dnoelle/tr-s97.ps.Z

Note that the web version includes a link to the PostScript version at
the bottom of the page.


Extreme Attraction:  The Benefits of Corner Attractors
------------------------------------------------------
by David C. Noelle, Garrison W. Cottrell, and Fred R. Wilms
Technical Report CS97-536
Department of Computer Science & Engineering
University of California, San Diego

Connectionist attractor networks have played a central role in many
cognitive models involving associative memory and soft constraint
satisfaction.  While early attractor networks used step activation
functions, permitting the construction of attractors for only binary
(or bipolar) patterns, much recent work has focused on networks with
continuous sigmoidal activation functions.  The incorporation of
sigmoidal processing elements allows for the use of expressive real
vector representations in attractor networks.  The empirical studies
reported here, however, reveal that the learning performance of
sigmoidal attractor networks is best when such general real vectors
are avoided -- when training patterns are explicitly placed in the
extreme corners of the network's activation space.  Using binary (or
bipolar) patterns produces benefits in the number of attractors
learnable by a network, in the accuracy of the learned attractors, and
in the amount of training required.  These benefits persist under
conditions of sparse patterns.  Furthermore, these experiments show
that the advantages of extreme-valued patterns are not solely effects
of the large separation between training patterns afforded by corner
attractors.


Thank you for your consideration.


-- David Noelle ----- Department of Computer Science & Engineering --
--------------------- Department of Cognitive Science ---------------
--------------------- University of California, San Diego -----------
-- noelle at ucsd.edu -- http://www.cse.ucsd.edu/users/dnoelle/ --------