A new definition of similarity for the bipolar WTA neural network model and its implementation are presented. A threshold is added to the system for distinguishing input patterns that are not one of the memory patterns or do not have enough ioformation to be recognized. The threshold value is related to the requirement of error tolerance.