A Fast Convolutional Denoising Autoencoder Based Extreme Learning Machine

Abstract

The convolutional autoencoder (CAE) was proposed on convolutional neural network (CNN) and denoising autoencoder (DAE). CAE can address the corrupted input samples and high dimensional problem. However, CAE has a shortcoming involving a large training timescale because the parameters of network are commonly tuned by gradient descent (GD) learning method. In order to alleviate this problem, this paper proposed a fast convolutional denoising autoencoder based extreme learning machine (ELM), called fast convolutional denoising autoencoder (FCDA). In FCDA, the random convolutional hidden nodes are used to reduce the dimension of input data. After that, the proposed denoising ELM autoencoder is used to reconstruct the cleaned data. The experimental results indicate that the proposed method not only speeds up the traditional CAE, but it also outperforms the CAE algorithm in terms of reconstruction error. Moreover, we applied the proposed method FCDA as the pre-processing method for ML-ELM classifier. The results illustrate the combination of the proposed FCDA, and ML-ELM achieves the classification performance better than the comparative methods.