Robust Deep Neural Network Using Fuzzy Denoising Autoencoder

Abstract

Deep neural network (DNN) has been applied in many fields and achieved great successes. However, DNN suffers from poor robustness for uncertainties because of its characteristic of the deterministic representation. To overcome this problem, a novel robust DNN (RDNN) is designed in this paper. First, a fuzzy denoising autoencoder (FDA) is developed to replace the general base-building unit in DNN. Then, the proposed RDNN is able to extract the robust representations to weaken the uncertainties. Second, a compact parameter strategy (CPS) is designed to reconstruct the parameters of FDA. Then, the computational burden of FDA can be alleviated to speed up the learning process. Third, an adaptive back-propagation (ABP) algorithm, with an adaptive learning rate strategy, is proposed to update the parameters of RDNN. Then, the performance of RDNN can be improved. Finally, the results on the benchmark problems and real applications demonstrate the effectiveness of RDNN.