Impact of Input Data Randomness on Training Performance of Basic Autoencoder and Stacked Autoencoder

In this paper, we examine the impact of the input data randomness on the traning performance of the autoencoder. Assuming the input data follow a Gaussian distribution, we mathematically analyze that the randomness of the input data and the training performance of the autoencoder have a linear relationship, and demonstrate this finding through experiments. Experiments on the basic autoencoder and the stacked autoencoder confirm that the input data randomness and the training performance of the autoencoder have a linear relationship regardless of the number of hidden layers. In addition, we examine the training performance of autoencoder according to the mean and standard deviation of the input data. The results support that the mean of the input data has a negligible effect on the training loss of the autoencoder. Therefore, we ensure consistency between mathematical analysis and the experimental results that the number of nodes in the hidden layer increases, the training loss of the autoencoder and the impact of input data randomness decrease.