A team-decision theoretic approach to learning in neural networks
A novel approach to neural network learning is developed through the use of team decision theoretic techniques. By considering a neural network as a cooperative team, a learning rule is derived which can adapt with a change in nature's statistics. Since this algorithm is a function of nature's statistics and not a particular pattern, an entire statistical class of patterns is learned. Furthermore, this learning is optimal in the sense of minimum risk. Since the learning is for an entire statistical class of patterns, care must be taken in retrieving the stored data. Specifically, when the input to the network is a noise corrupted pattern from nature, the network's output will be affected by the noise. Therefore, an algorithm for retrieving a pattern from the statistical class of patterns which has been learned is also derived. This algorithm-a sequential MAP estimator for observations corrupted by colored noise-is able to retrieve a pattern from a clean as well as a noisy input. The algorithm is simulated on a set of 5 x 5 binary patterns which are stored as the class of patterns having a normal distribution with zero mean and unity variance. Results of recalling this set of patterns from both clean and noisy inputs are presented.
- Research Organization:
- Arkansas Univ., Fayetteville, AR (USA)
- OSTI ID:
- 6046430
- Resource Relation:
- Other Information: Thesis (Ph. D.)
- Country of Publication:
- United States
- Language:
- English
Similar Records
Neural network recognition and classification of aerosol particle distributions measured with a two-spot laser velocimeter
Model neural networks