Handwritten digit recognition based on shunting inhibitory convolutional neural networks

This paper presents the application of a new class of convolutional neural networks based on the mechanism of shunting inhibition for handwritten digit recognition. With a three layer network architecture and the use of shunting inhibitory neurons as information processing elements, the network consists of 1926 trainable parameters which are adapted by a first-order gradient training algorithm derived from Rprop, Quickprop, and SuperSAB. Trained on a dataset of 10,000 samples and evaluated on the entire test set of the MNIST database, these networks achieve classification accuracies above 93% with the best performance obtained from those networks with partial connection schemes.