New learning and control algorithms for neural networks
Neural networks offer distributed processing power, error correcting capability and structural simplicity of the basic computing element. Neural networks have been found to be attractive for applications such as associative memory, robotics, image processing, speech understanding and optimization. Neural networks are self-adaptive systems that try to configure themselves to store new information. This dissertation investigates two approaches to improve performance: better learning and supervisory control. A new learning algorithm called the Correlation Continuous Unlearning (CCU) algorithm is presented. It is based on the idea of removing undesirable information that is encountered during the learning period. The control methods proposed in the dissertation improve the convergence by affecting the order of updates using a controller. Most previous studies have focused on monolithic structures. But it is known that the human brain has a bicameral nature at the gross level and it also has several specialized structures. In this dissertation, the author investigates the computing characteristics of neural networks that are not monolithic being enhanced by a controller that can run algorithms that take advantage of the known global characteristics of the stored information. Such networks have been called bicameral neural networks. Stinson and Kak considered elementary bicameral models that used asynchronous control. New control methods, the method of iteration and bicameral classifier, are now proposed. The method of iteration uses the Hamming distance between the probe and the answer to control the convergence to a correct answer, whereas the bicameral classifier takes advantage of global characteristics using a clustering algorithm.
- Research Organization:
- Louisiana State Univ. and Agricultural and Mechanical Coll., Baton Rouge, LA (United States)
- OSTI ID:
- 5533869
- Resource Relation:
- Other Information: Thesis (Ph. D.)
- Country of Publication:
- United States
- Language:
- English
Similar Records
Neural-network algorithms and architectures for pattern classification
Artificial Neural Network Training with Conjugate Gradients for Diagnosing Transients in Nuclear Power Plants
Related Subjects
NEURAL NETWORKS
ALGORITHMS
PERFORMANCE
CONTROL
CONVERGENCE
DISTRIBUTED DATA PROCESSING
FAULT TOLERANT COMPUTERS
IMAGE PROCESSING
INFORMATION
LEARNING
OPTIMIZATION
ROBOTS
SPEECH SYNTHESIZERS
USES
COMPUTERS
DATA PROCESSING
DIGITAL COMPUTERS
ELECTRONIC EQUIPMENT
EQUIPMENT
MATHEMATICAL LOGIC
PROCESSING
990200* - Mathematics & Computers