Performance of BSDT Decoding Algorithms Based on Locally Damaged Neural Networks

Abstract

Traditional signal detection theory (SDT) and recent binary signal detection theory (BSDT) provide the same basic performance functions: receiver operating characteristic (ROC) and basic decoding performance (BDP) curves. Because the BSDT may simultaneously be presented in neural network (NN), convolutional, and Hamming distance forms, it contains more parameters and its predictions are richer. Here we discuss a formal definition of one of specific BSDT parameters, the confidence level of decisions, and demonstrate that the BSDT’s ROCs and BDPs, as functions of the number of NN disrupted links, have specific features, though rather strange at first glance but consistent with psychophysics experiments (for example, judgment errors in cluttered environments).