On the problem of binary detection with neural networks

A neural structure is purposed for binary detection applications such as radar or sonar detection. Most of the experiments in this field use typical back propagation algorithm that minimizes sum-of-square-error, although this is not the optimal solution to this problem. For this reason, performances of several nets whose basic differences lay in their criterion functions are compared. Their Receiver Operating Characteristics (ROC) and detection curves are presented, and dependencies on training signal-to-noise ratio, structure, and threshold value are analyzed.