A Soft Target Learning Method of Posterior Pseudo-probabilities Based Classifiers with Its Application to Handwritten Digit Recognition

This paper proposes a soft target discriminative learning method for posterior pseudo-probabilities based classification. The empirical loss is measured based on two soft targets which are corresponding with positive samples and negative samples of the class. The learning objective is to minimize empirical loss and maximize the difference between two soft targets. Consequently, we obtain unknown parameters in posterior pseudoprobabilities based classifiers by optimizing the objective using the gradient descent algorithm. We apply the proposed soft target method to handwritten digit recognition. Experimental results on MNIST database show the effectiveness of our method.