Uncertainty Analysis Of Bayesian Deep Network

With the advent of the era of artificial intelligence,deep learning technology is constantly being updated.In the field of computer vision,the use of deep networks is increasingly widespread.However,there are two challenges facing the development of deep learning technology at present: One is the lack of uncertainty measurement of prediction results in its network architecture,so in practical applications,decision-making mistakes are often caused by the model's overconfidence.The second is the high cost of acquiring high-quality and labeled data in real life,while deep learning relies on a large amount of training data.This thesis directly confronts the above challenges,carries out the theoretical research of Bayesian deep network,and puts it into practice,conducting experiments on image denoising tasks and image classification tasks respectively.The main results of the thesis are as follows:(1)From a mathematical point of view,the theoretical framework of Bayesian deep network is derived systematically and in detail.(2)In order to understand the uncertainty information in the model,this paper designs bayesian deep networks based on cognitive uncertainty,accidental uncertainty,and mixed uncertainty,respectively,and conducts denoising experiments.The realization of the deep denoising model outputs a clean image and gives a schematic diagram of uncertainty,thus guiding the direction of subsequent optimization.Compared with the traditional network,the accuracy of the network model after modeling has been improved.(3)In order to train a high-precision model with a small amount of labeled data,this paper combines an active learning framework and a Bayesian deep network.Based on the uncertainty theory,the maximum entropy acquisition function,interactive information acquisition function and maximum change rate acquisition function for Bayesian deep network are derived and compared experiments are conducted.Compared with deterministic active learning deep networks,Bayesian active learning deep networks can realize the use of a limited number of labeled data to train high-precision models.