Abstract
Error correcting codes were used previously to encode the output of feed-forward neural nets. We study the effect of additive noise on the performance of a coded net and compare it to an uncoded net. Some necessary analytical tools are developed to estimate the performance of a neural net in the presence of noise. Simulation examples (isolated word utterances recognition) are also included to show the advantage of coding in reducing the probability of classification error due to noise. In addition we point the use of the estimated performance as a lower limit to the performance of a multilayer neural net.