Recognition of handwritten numbers on the basis of convolutional neural networks

Abstract

It is known that the use of a multilayer perceptron with a traditional structure in solving real problems of image recognition and classification causes certain difficulties, in particular, associated with a large image dimension (this significantly increases the number of neurons and synaptic connections in neural connections of the network and, therefore, significantly increases the training sample and training time). In addition, the topology of the input data is ignored. The components of the input layer can be presented in any order, regardless of the learning goal. However, images have a strict two-dimensional structure, in which there is a relationship between neighboring pixels in space. These shortcomings are deprived of the so-called convolutional neural networks, which are a special class of multilayer perceptrons specifically designed to recognize two-dimensional surfaces with a high degree of invariance to scaling, displacement, rotation, angle change, and other spatial transformations. This article discusses the problem of practical implementation of handwritten digit recognition based on convolutional neural networks (CNN). The CNN architecture is presented, for which it is recommended to use cross-entropy as the learning loss function, and the Softmax function as the activation function of the last CNN layer. It is also proposed to use a modification of the well-known error backpropagation algorithm to implement the CNN learning algorithm. To do this, the article presents the main ratios for errors at each level. Ref. 14, pic.2.

Keywords

• artificial intelligence
• convolution neural networks
• handwritten digit recognition
• MNIST database
• architecture of convolution neural networks
• learning algorithm