Preprint Available: Binding Problem

[Bartlett Mel]

The following preprint is now available via our web page:

http://lnc.usc.edu
25 pages, 230K gzipped postscript

=========================================================

"Seeing with Spatially-Invariant Receptive Fields: 
When the `Binding Problem' Isn't"

Bartlett W. Mel
Biomedical Engineering Department
University of Southern California

Jozsef Fiser
Department of Brain and Cognitive Sciences
University of Rochester


ABSTRACT

We have studied the design tradeoffs governing visual representations
based on complex spatially-invariant receptive fields (RF's), with an
emphasis on the susceptibility of such systems to false-positive
recognition errors---Malsburg's classical ``binding'' problem.  We begin
by deriving an analytical model that makes explicit how recognition
performance is affected by the number of objects that must be
distinguished, the number of features included in the representation,
the complexity of individual objects, and the clutter load, i.e. the
amount of visual material in the field of view in which multiple objects
must be simultaneously recognized, independent of pose, and without
explicit segmentation.  Using the domain of text as a convenient
surrogate for object recognition in cluttered scenes, we show that, with
corrections for the non-uniform probability and non-independence of
English text features, the analytical model achieves good fits to
measured recognition rates in simulations involving a wide range of
clutter loads, word sizes, and feature counts.  We then present a greedy
algorithm for feature learning, derived from the analytical model, which
grows a visual representation by choosing those features most likely to
distinguish objects from the cluttered backgrounds in which they are
embedded.  We show that the representations produced by this algorithm
are decorrelated, heavily weighted to features of low conjunctive order,
and remarkably compact.  Our results provide a quantitative basis for
understanding when, and under what conditions, spatially-invariant
RF-based representations can support veridical perception in
multi-object scenes, and lead to several insights regarding the
properties of visual representations optimized for specific visual
recognition tasks.


-- 

 Bartlett W. Mel                               (213)740-0334, -3397(lab)
 Assistant Professor of Biomedical Engineering (213)740-0343 fax
 University of Southern California, OHE 500     mel at lnc.usc.edu,
http://lnc.usc.edu
                             
     US Mail: BME Department, MC 1451, USC, Los Angeles, CA 90089
       Fedex: 3650 McClintock Ave, 500 Olin Hall, LA, CA 90089