Improvement of GRBM Based on Activation Function

摘要

In this paper, inspired by ReLu and Softplus activation function, we propose two improved models of GRBM, called SPC-GRBM and RPC-GRBM, to obtain better recognition results. Different from the traditional activation-function-improved models, SPC-GRBM and RPC-GRBM focus on the visual layer activation function, which is trained by CBCL database and is finally used for image classification with the help of the k-Nearest Neighbor (KNN) method. Experimental results show that the recognition accuracy of SPC-GRBM and RPC-GRBM are both enhanced and SPC-GRBM has achieved the highest recognition rate among the several models particularly, of which the recognition accuracy is 20.10% higher than the original GRBM. In addition, the reconstruction error is apparently reduced and its performance keeps well.