A Cognitive Vision Architecture Integrating Neural Networks with Symbolic Processing

Gösta Granlund
K\"unstliche Intelligenz
Künstliche Intelligenz
Number 2, Pages 18-24
2005

Abstract

A fundamental property of cognitive vision systems is that they shall be extendable, which requires that they can both acquire and store information autonomously. The paper discusses organization of systems to allow this, and proposes an architecture for cognitive vision systems. The architecture consists of two parts. The first part, step by step learns a mapping from percepts directly onto actions or states. In the learning phase, action precedes perception, as action space is much less complex. This requires a semantic information representation, allowing computation and storage with respect to similarity. The second part uses invariant or symbolic representations, which are derived mainly from system and action states.

Through active exploration, a system builds up concept spaces or models. This allows the system to subsequently acquire information using passive observation or language. The structure has been used to learn object properties, and constitutes the basic concepts for a European project COSPAL, within the IST programme.

Bibtex entry

@Article{g05,
  author = 	 {G\"osta Granlund},
  title = 	 {A {C}ognitive {V}ision {A}rchitecture {I}ntegrating 
                  {N}eural {N}etworks with {S}ymbolic {P}rocessing},
  journal = 	 {K\"unstliche Intelligenz},
  year = 	 {2005},
  OPTkey = 	 {},
  OPTvolume = 	 {},
  number = 	 {2},
  pages = 	 {18--24},
  OPTmonth = 	 {},
  note = 	 {ISSN 0933-1875, B\"ottcher IT Verlag, Bremen, Germany},
  OPTannote = 	 {}
}