Properties and characteristics of self-organizing neural networks for unsupervised pattern recognition
Recently, neural network models have become more popular and are now being applied to numerous fields. The author proposes several new algorithms for pattern classification based on extensions of Kohonen's self-organizing feature map model. His models are based on a one dimensional ring-structure constructed such that every functional unit (neuron) is connected to its two nearest neighbors. Each neuron receives the same scalar-valued input signal {xi} parallel. One of the very important features of the Self-Organizing Neural Network (SONN) system is self-organizing behavior. This system works without any information about the number of clusters or cluster centers. It adjusts weights between the input layer and the distance layer until it is stabilized. Then it classifies each input according to the distance between the weights and the normalized input using Bezdek's (BEZ81) membership value equation. These new algorithms are examined for convergence characteristics and performance using a new learning rule. In addition, performance comparisons with other methods are studied and comparative convergence characteristics are investigated. A distributed parallel algorithm is also developed and implemented.
- Research Organization:
- South Carolina Univ., Columbia, SC (United States)
- OSTI ID:
- 5918957
- Resource Relation:
- Other Information: Thesis (Ph.D)
- Country of Publication:
- United States
- Language:
- English
Similar Records
Self-organizing neural network as a fuzzy classifier
New learning and control algorithms for neural networks
Related Subjects
NEURAL NETWORKS
USES
PATTERN RECOGNITION
COMPUTERIZED SIMULATION
ALGORITHMS
CONVERGENCE
DISTRIBUTED DATA PROCESSING
MAPPING
MATHEMATICAL MODELS
PARALLEL PROCESSING
PERFORMANCE
TASK SCHEDULING
DATA PROCESSING
MATHEMATICAL LOGIC
PROCESSING
PROGRAMMING
SIMULATION
990200* - Mathematics & Computers