I. Introduction
Deep learning methods have brought revolutionary progress to computer vision and machine learning. The traditional deep neural network has been successfully applied in many domains, including very sensitive domains, such as medical health[1][2], security[3], fraudulent transactions[4] and so on. These fields depend heavily on the prediction accuracy of the model, and even an overconfident decision may lead to an unacceptable result, while the current neural network architecture lacks uncertainty measurement in prediction. The introduction of Bayesian method[5], provides uncertainty measurement in prediction for neural network, namely Bayesian neural network. The addition of priors provides a constraint and regularization to the network, which is more robust to over-fitting. Compared with the single point estimation neural network, Bayesian neural network takes random variables that obey a certain distribution as weights, rather than a fixed value. Forward propagation is sampled from the weight distribution and then calculated. Therefore, Bayesian network provides uncertainty estimation through its parameters of probability distribution form, which is widely used in medical diagnosis[6], text and image classification[7], decision support system[8][9], bioinformatics[10] and other fields. However, the uncertainty in Bayesian neural network decreases with the increase of data volume. In large-scale data sets, the importance of a priori is reduced, and the measurement significance of this uncertainty becomes less obvious. Therefore, the design of a better performance Bayesian network is more worthy of our study. Nevertheless, manually designing a powerful architecture is tedious and formidable, which relies heavily on human expertise and may be very time-consuming. Consequently, automatic design of neural network architecture has become an effective means.