Deep classifiers require lots of computations due to the use of multi-layer structure and processing of bulk input samples. Selecting effective features that have the same accuracy as full features, has recently been considered. Using feature selection, less calculation is needed to update the weights and consequently the speed of the training/testing process is increased. In this research, a new deep feature selection method called Discriminative Deep Feature Selection using Signed Laplacian Restricted Boltzmann Machine (DDFS-SLRBM) is proposed. In the approach, full training samples are fed into a SLRBM model. In the updates, the weight matrix is recomputed using the neighborhood matrix of similar and dissimilar classes which results in the discriminative property of the approach. Then, selected features are identified based on the minimum reconstruction error criterion. The efficiency of the proposed algorithm is demonstrated by performing different experiments on MNIST, GISETTE and Protein datasets. The experiments show that the proposed method is able to improve the classification accuracy and the generalizability of the approach while reducing the processing time. Also, the proposed approach shows scalability on the type of the problem, number of original features, number of samples and the number of final selected features.
Read full abstract