Convolution Neural Network Learning Features for Handwritten Digit Recognition

The classical six-layer neural network is considered. This network is used to recognize handwritten digit patterns from the MNIST database. The influence of the size of mini-batches on the learning speed and pattern recognition accuracy is analyzed. The optimal sizes of mini-batch are obtained. The relationship between the accuracy of training and the accuracy of test samples is considered. The change in the values of the radius vector of the scales during training is shown. Conclusions are drawn about the influence of the initial value of the balance on the recognition accuracy. A more accurate formula is obtained for the limits, in which the initial values of the weights of the neural network are generated.