A multi-sieving neural network architecture that decomposes learning tasks automatically

This paper presents a multi-sieving network (MSN) architecture and the multi-sieving learning (MSL) algorithm for it. The basic idea behind MSN architecture is that patterns are classified by a rough sieve at the beginning and done by finer ones gradually. MSN is constructed by adding a sieving module (SM) adaptively with progress of training. SM consists of two different neural networks and a simple logical circuit. MSL algorithm starts with a single SM, then does the following three phases repeatedly until all the training samples are successfully learned: 1) the learning phase in which the training samples are learned by the current SM; 2) the sieving phase in which the training samples that have been successfully learned are sifted out from the training set; and 3) the growing phase in which the current SM is frozen and a new SM is added in order to learn the remaining training samples. The performance of MSN architecture is illustrated on two benchmark problems