Binding - interim report

[Jerome Feldman]

 A while back, I posted a query about Jackendoff's four challenges
in neural binding. In addition to the responses that were posted
to the whole group, the three other replies are included in this message.
Only one response showed any evidence of having looked at Jackendoff's
problems and formulating a response.

 It isn't obvious (at least to me) how to use any of the standard 
techniques to specify a model that meets Jackendoff's criteria. I
privately encouraged the respondents to outline how they would do
this and hope that they and others will respond.

Jerry F.


~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Jerry,

Connectionist models that learn to process sentences have had to solve
the binding problem and have done so for many years.  Among the models
that solve this problem I can list St. John and McClelland (Artificial
Intelligence, 1990), the Miikkulainen and Dyer work, Elman's simple
recurrent networks, and a recent CMU-CS dissertation by Doug Rohde.
The latter is available at doug's home page:

   http://tedlab.mit.edu/~dr/Thesis

Best wishes,

 - Jay McClelland

~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Here are some papers describing a mostly localist approach toward
the variable binding problem:


A. Browne and R. Sun,
Connectionist inference models.
{\it Neural Networks},
Vol.14, No.10, pp.1331-1355,
December 2001.

A. Browne and R. Sun, Connectionist variable binding.
{\it Expert Systems},
Vol.16, No.3, pp.189-207. 1999.

R. Sun,
Robust reasoning: integrating rule-based and similarity-based reasoning.
{\it Artificial Intelligence} (AIJ).
Vol.75, No.2, pp.241-296.  June, 1995.


R. Sun,
On schemas, logics, and neural assemblies. {\it Applied Intelligence}.
Vol.5, No.2. pp.83-102. 1995.

R. Sun,
Beyond associative memories: logics and variables in
connectionist networks. {\it Information Sciences},
Vol.70, No.1-2.  pp.49-74. 1993.

R. Sun,
On variable binding in connectionist networks. {\it Connection Science},
Vol.4,
No.2, pp.93-124. 1992.

Most of these papers can be downloaded from my web page.

Professor Ron Sun, Ph.D  


~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
        
Dear Jerome,

like DeLiang Wang, who recently posted a response to your request
on the connectionists, I would say that there are now powerful
neural binding architectures available.

I have been working on a spatial model of neural binding, which employs
spatial coactivation of neurons for the representation of feature
bindings. It is called competitive layer model, and has been shown
to perform well on perceptual grouping tasks like contour grouping,
texture and greyscale segmentation:  

H. Wersing, J. J. Steil, and H. Ritter. A competitive layer model for
feature binding and sensory segmentation. Neural Computation
13(2):357-387 (2001).


http://www.techfak.uni-bielefeld.de/ags/ni/publications/papers/WersingSteilRitter2001-ACL.ps.gz

H. Wersing. Spatial Feature Binding and Learning in Competitive Neural
Layer Architectures PhD Thesis. Faculty of Technology, University of
Bielefeld,
March 2000. Published by Cuvillier, Goettingen

http://www.techfak.uni-bielefeld.de/~hwersing/dissertation.ps.gz

One particular feature of the model is, that it can be trained in very
efficient way, solving a simple quadratic optimization problem. Our 
applications so far concentrated on segmentation problems, but the 
framework could be easily applied to other problem domains:

H. Wersing. Learning Lateral Interactions for Feature Binding and
Sensory Segmentation. Advances in Neural Information Processing Systems
NIPS 2001, Vancouver

http://www.techfak.uni-bielefeld.de/ags/ni/publications/papers/Wersing2001-LLI.ps.gz


With kindest regards,

Heiko Wersing


~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~