Target Detection Using a Three-Layered Neural Network Trained by Supervised Back-Propagation
Abstract
A neural network should be considered as a computer program that has the ability to improve its performance at a defined task by making use of a 'learning phase' before it is applied to actual data. During this learning phase a large amount of representative data is repeatedly presented to the network. The network automatically adjusts its internal parameters to optimize its performance as measured by a simple test at the network's output. No information on the internal parameters are presented to the user. This ability of neural networks to use a priori knowledge often leads them to being incorrectly referred to as 'intelligent systems' but it does not give them the potential to outperform more conventional techniques that do not involve prior training. However, it should be borne in mind that this potentially enhanced performance is only available for the previously defined task. To be as successful at a different task the network must be retrained. To date, in the realm of sonar, neural networks have been used almost exclusively for classification where it is perceived that operator experience is a vital factor. They can, however, operate as a detection system and still make use of any available a priori knowledge. In this memorandum a so called three-layered network has been applied to the detection problem using simulated data. The performance of networks as a function of signal-to-noise-ratio, absolute signal level and network complexity has been briefly evaluated. The results indicate that such a network can perform at least as well as a conventional detection system when simulated data is used.
Document Details
- Document Type
- Technical Report
- Publication Date
- May 01, 1990
- Accession Number
- AD1114290
Entities
People
- H. Meek
Organizations
- SACLANT ASW Research Centre