EPE Association: Pelincec 2005 - 06: Generalized Brain-State-in-a-Box (GBSB) Neural Network

Pelincec 2005 - 06: Generalized Brain-State-in-a-Box (GBSB) Neural Network
You are here: EPE Documents > 05 - EPE Supported Conference Proceedings > Pelincec 2005 > Pelincec 2005 - 06: Generalized Brain-State-in-a-Box (GBSB) Neural Network

[return to parent folder]

The Generalized Brain-State-in-a-Box (gBSB) Neural Network: Model, Analysis, and Applications
By Cheolhwan Oh, Stefen Hui, Stanislaw H. Zak [View]
[Download]

Abstract: The generalized Brain-State-in-a-Box (gBSB) neural network is a generalized version of the Brain- State-in-a-Box (BSB) neural network. The BSB net is a simple nonlinear autoassociative neural network that was proposed by J. A. Anderson, J. W. Silverstein, S. A. Ritz, and R. S. Jones in 1977 as a memory model based on neurophysiological considerations. The BSB model gets its name from the fact that the network trajectory is constrained to reside in the hypercube Hn = [−1, 1]n. The BSB model was used primarily to model effects and mechanisms seen in psychology and cognitive science. It can be used as a simple pattern recognizer and also as a pattern recognizer that employs a smooth nearness measure and generates smooth decision boundaries. Three different generalizations of the BSB model were proposed by Hui and ÿZak, Golden, and Anderson. In particular, the network considered by Hui and ÿZak, referred to as the generalized Brain-State-in-a-Box (gBSB), has the property that the network trajectory constrained to a hyperface of Hn is described by a lower-order gBSB type model. This property simplifies significantly the analysis of the dynamical behavior of the gBSB neural network. Another tool that makes the gBSB model suitable for constructing associative memory is the stability criterion of the vertices of Hn. Using this criterion, a number of systematic methods to synthesize associative memories were developed. In this paper, an introduction to some useful properties of the gBSB model and some applications of this model are presented first. In particular, the gBSB based hybrid neural network for storing and retrieving pattern sequences is described. The hybrid network consists of autoassociative and heteroassociative parts. In the autoassociative part, where patterns are usually represented as vectors, a set of patterns is stored by the neural network. A distorted (noisy) version of a stored pattern is subsequently presented to the network and the task of the neural network is to retrieve (recall) the original stored pattern from the noisy pattern. In the heteroassociative part, an arbitrary set of input patterns is paired with another arbitrary set of output patterns. After presenting hybrid networks, neural associative memory that processes large scale patterns in an efficient way is described. Then the gBSB based hybrid neural network and the pattern decomposition concept are used to construct a neural associative memory. Finally, an image storage and retrieval system is constructed using the subsystems described above. Results of extensive simulations are included to illustrate the operation of the proposed image storage and retrieval system.

Pelincec 2005 - 06: Generalized Brain-State-in-a-Box (GBSB) Neural Network
You are here: EPE Documents > 05 - EPE Supported Conference Proceedings > Pelincec 2005 > Pelincec 2005 - 06: Generalized Brain-State-in-a-Box (GBSB) Neural Network
	[return to parent folder]

	The Generalized Brain-State-in-a-Box (gBSB) Neural Network: Model, Analysis, and Applications By Cheolhwan Oh, Stefen Hui, Stanislaw H. Zak	[View] [Download]
Abstract: The generalized Brain-State-in-a-Box (gBSB) neural network is a generalized version of the Brain- State-in-a-Box (BSB) neural network. The BSB net is a simple nonlinear autoassociative neural network that was proposed by J. A. Anderson, J. W. Silverstein, S. A. Ritz, and R. S. Jones in 1977 as a memory model based on neurophysiological considerations. The BSB model gets its name from the fact that the network trajectory is constrained to reside in the hypercube Hn = [−1, 1]n. The BSB model was used primarily to model effects and mechanisms seen in psychology and cognitive science. It can be used as a simple pattern recognizer and also as a pattern recognizer that employs a smooth nearness measure and generates smooth decision boundaries. Three different generalizations of the BSB model were proposed by Hui and ÿZak, Golden, and Anderson. In particular, the network considered by Hui and ÿZak, referred to as the generalized Brain-State-in-a-Box (gBSB), has the property that the network trajectory constrained to a hyperface of Hn is described by a lower-order gBSB type model. This property simplifies significantly the analysis of the dynamical behavior of the gBSB neural network. Another tool that makes the gBSB model suitable for constructing associative memory is the stability criterion of the vertices of Hn. Using this criterion, a number of systematic methods to synthesize associative memories were developed. In this paper, an introduction to some useful properties of the gBSB model and some applications of this model are presented first. In particular, the gBSB based hybrid neural network for storing and retrieving pattern sequences is described. The hybrid network consists of autoassociative and heteroassociative parts. In the autoassociative part, where patterns are usually represented as vectors, a set of patterns is stored by the neural network. A distorted (noisy) version of a stored pattern is subsequently presented to the network and the task of the neural network is to retrieve (recall) the original stored pattern from the noisy pattern. In the heteroassociative part, an arbitrary set of input patterns is paired with another arbitrary set of output patterns. After presenting hybrid networks, neural associative memory that processes large scale patterns in an efficient way is described. Then the gBSB based hybrid neural network and the pattern decomposition concept are used to construct a neural associative memory. Finally, an image storage and retrieval system is constructed using the subsystems described above. Results of extensive simulations are included to illustrate the operation of the proposed image storage and retrieval system.