Research And Applications Of Uncertainty In Bayesian Deep Label Distribution Learning

Though widely used,most deep convolutional neural networks fail to capture pre-diction uncertainty,which can be crucial in scenarios such as automotive applications and disease diagnosis.The proposed Bayesian deep learning(BDL)model based on dropout provides a new idea for obtaining the uncertainty of the deep learning model.Aleatoric and epistemic uncertainties have been proposed in the BDL framework for single-label tasks,which require training images with unambiguous labels for success.Some situations do not have precise labels by nature,such as age estimation or lesion contour annotation by different physicians in the real world.Label distribution learn-ing(LDL)has been proposed to account for the label ambiguity.However,uncertainty estimation has not been studied for LDL.This study presents a Bayesian deep label distribution learning(BDLDL)to obtain the uncertainties of LDL tasks.We define se-mantic uncertainty to account for the incompleteness of data acquisition and inaccurate data labeling and unify the three types of uncertainties' mathematical expressions for classification and LDL.We apply the proposed semantic uncertainty and the other two uncertainties to many fields,which prove that uncertainties are valuable in practice.The specific research content and innovations mainly include the following four points:1.The Bayesian deep label learning distribution learning model(BDLDL)is pro-posed to obtain the uncertainties of LDL tasks.We present a calculation method for aleatoric uncertainty and model uncertainty in label distributed learning tasks.We de-fine semantic uncertainty to account for the incompleteness of data acquisition and inac-curate data labeling and unify the three types of uncertainties' mathematical expressions for classification and LDL.2.We apply three uncertainties to single-label learning tasks,study the relationship between three uncertainties and model performance in classification and segmentation tasks.Besides,we propose three types of uncertainty loss on single-label tasks.The ex-perimental results show that the performance of the single-label learning task gets better as the three uncertainties decrease.The model achieves state-of-art performance on the two public classification datasets(Adience and ICCV ChaLearn LAP 2015)and two public segmentation datasets(CamVid and e_ophtha_EX segmentation)with training with three uncertainties losses,which proves that three uncertainty losses can effectively improve model performance.Further,aleatoric,model and semantic uncertainties can be used as features to train a support vector classifier to evaluate the correctness of the prediction results.3.Three uncertainties are applied to LDL tasks,study the correlation between the three uncertainties and model performance on LDL tasks.To improve the performance of LDL tasks,we proposed three uncertainty losses on LDL tasks,which are experi-mentally proved on two public LDL datasets including ICCV ChaLearn LAP 2015 and SCUT-FBP datasets and one private ROP2020 dataset.The difference between model uncertainty and semantic uncertainty is showed by the relationship between uncertain-ties and the standard deviation of labels.At the same time,we use the support vector regression model with three uncertainties as features to evaluate the correctness of the prediction results.4.Based on the relationship between the three uncertainties obtained and the re-liability of the prediction results,we can apply them to other fields.For example,us-ing three types of uncertainty as a criterion to select the training samples substantially promotes active learning.Combining uncertainties with the policy strategy in rein-forcement learning helps the convergence of the agent.Besides,three uncertainties are explored to distinguish between adversarial images and true images.We can prove that three applications could obtain better performance with three uncertainties,especially semantic uncertainty.