Scale conjugate gradient based learning applied to handwritten digit classification

Most of the existing neural network algorithms treat mean square error function as the standard cost function. In our proposed system, we used cross-entropy as a cost function. For this Softmax transfer function is used. Consequently, a lot of improvement has been accomplished on the training method. For training, we used scale conjugate gradient algorithm. We have used 42,000 images of handwritten digit dataset and performed testing on 28,000 images. From this method, we achieved an accuracy rate of about 98.8 percent. The training results are sensibly evaluated to test the performance of a proposed system. There are three essential factors and they are considered from high to low priority: Receiver Operating Characteristics, which actually gives the classification stability result, confusion matrix, and gradient value of proposed system.