Feedforward Neural Network Design

Feedforward neural network design is critical to the success of this technology. Good design is most important when addressing visual pattern recognition problems. This talk first gives a brief review of neural networks and their abilities, then presents important guidelines to their design. Rules for intelligent choice of data representations, architectures, sizes, and post-processing schemes are also presented. An example is given describing, at a high level, my work on passive sonar signal recognition using neural nets. The passive sonar problem is described, only with my choice of data representation, pre-processor, and neural network architecture. Results of the work in terms of correct classification performance conclude the presentation.