Clustered class-dependant training method for digit recognition classifiers

This paper presents a convolutional neural network clustering approach for handwritten digits recognition. Neural networks were trained individually, using the same training set and combined into clusters, depending on the training method used. These clusters formed a layered architecture, where each layer attempted to recognize the given digit, when the previous layers were not able to do so with sufficient certainty. We examine various ways of combining such clusters and training their constituent networks.