Channel Associative Networks for Multiple Valued Mappings

Per-Erik Forssén, Björn Johansson, Gösta Granlund
ICVW06, Graz, Austria
2nd International Cognitive Vision Workshop
Pages 4-11
May, 2006

Abstract

This paper introduces a novel artificial neural network (ANN) structure which can learn multiple valued, non-linear mappings. This is accomplished by expanding both input and output domains using a set of localised functions called channels. In the channel space the learning problem becomes a linear mapping, which can be made sparse using a non-negative constraint. By applying this ANN to an object view recognition problem, we demonstrate that the network is able to learn efficiently under perceptual aliasing. This has applications for cognitive vision systems where learning has to occur at several abstraction levels simultaneously. If a subsystem is supplied with ambiguous inputs, learning will not break down, instead the subsystem will learn to pass the ambiguity to the output side, where the next subsystem can hopefully resolve it using additional context.

Full Paper

Portable document format file PDF

Bibtex entry

@InProceedings{fjg06,
  author = 	 {Per-Erik Forss\'en and Bj\"orn Johansson and G\"osta Granlund},
  title = 	 {Channel Associative Networks for Multiple Valued Mappings},
  booktitle =    {2nd International Cognitive Vision Workshop},
  pages = 	 {4--11},
  year = 	 {2006},
  address = 	 {Graz, Austria},
  month = 	 {May}
}