Increasing the Receptive Field of Neurons in Convolutional Neural Networks

The convolutional neural network architectures for classifying 1D and 2D signals are analyzed. The authors have found that for a high-dimensional input signal, one can ensure an adequate classification accuracy only by using a large number of layers. It is impossible to achieve the required accuracy with limited computing resources. However, if the number of layers is limited, the accuracy decreases, starting from some critical dimensionality value. A method for modifying a convolutional neural network with relatively small number of layers to solve this problem has been proposed. Its effectiveness has been experimentally proved.