Low-Interference Output Partitioning for Neural Network Training

Abstract

Abstract-This paper presents a new output partitioning approach with the advantages of constructive learning and output parallelism. Classification error is used to guide the partitioning process so that several smaller sub-dimensional data sets are divided from the original data set. When training each sub-dimensional data set in parallel, the smaller constructively trained sub-network uses the whole input vector and produces a portion of the final output vector where each class is represented by one unit. Three classification data sets are used to test the validity of this algorithm, while the results show that this method is feasible.