Layered fractal neural net: computational performance as a classifier1

摘要

A fractal connection structure among different layers of multi-layer neural network has been proposed here and its computational performance as a pattern classifier has been studied in comparison to a more conventional and widely used multi-layer perceptron classifier (MLP). A good classifier should consider two aspects: correct classification and perfect generalization. After an appropriate learning period, a good classifier should be able to classify any unknown sample outside the training set with the same accuracy as any sample from the training set. The performance of the proposed neural network architecture as compared to a fully connected multi-layer perceptron and any randomly connected architecture with same average connectivity in terms of proper classification and good generalization has been studied by simulation with sonar data collected from underwater target classification problems using sonar signals. It was found that with equal average connectivity, a fractally connected net performs better than a randomly connected net, and with the same number of neurons a fractally connected net performs better than the fully connected net.