A new neural network structure with cooperative modules

A new neural network structure is proposed in order to solve complex (large) classification problems. At first, the problem is divided into a number of simpler sub-problems using an unsupervised network. Then, each sub-problem is solved using one module of the proposed network. The new addition to this traditional "divide-and-conquer" structure is a voting layer above the outputs of the modules. In this layer, the modules cooperate in deciding the final decision by training every module on some additional classes: each class represents a group of classes of another module. During testing, the average "vote" of the modules for each other is calculated and multiplied by the classes output. This scheme gave higher accuracy and faster learning speed than the backpropagation scheme. A recursive hierarchy of this network can be built-up for more enhancement to the performance.<>