Classification problems in a sequential framework is a very important field in pattern recognition. One of biggest concerns in classification problems is overtraining. In metacognitive neural network (McNN), overtraining can be avoided by using the confidence of classifier (CoC) measure, which assigns a value between [0,1] to class label output neuron. By testing CoC we found the value is not accurate enough for this measure. Another issue is about hingloss error function in McNN. This error function doesn’t have a particular probabilistic interpretation. Due to this deficiency, first we proposed Soft meta-cognitive neural network (SMcNN) algorithm by changing CoC measure and activation function of output layer via softmax function, then we applying cross-entropy function instead of hingloss error to achieve better accuracy. SMcNN applied to well-known UCI datasets and the reasult compared with McNN, SVM and some other classifier. Experimental results showed we improve the classification performance.

Keywords