An original architecture of convolutional neural network for handwritten digits 
recognition is developed. This network includes four layers. The first two 
layers are composed of neurons with local receptive fields and sigmoid activation function. 
The last two layers form a classifier based on radial basis functions. 
A backpropagation algorithm specifically adapted to the network architecture was used. 
A character distortion method applied to increase the diversity of the training data. 
The network was tested on the MNIST database.