Collaborative Query Of Adversarial And Uncertain Samples For Deep Active Learning

In recent years,deep learning has made breakthroughs in many fields such as image understanding,speech recognition,and natural language processing.However,in practical applications,deep learning often relies on large-scale annotated data for model training.It has become urgent to reduce the amount of data annotation required for deep learning.Traditional active learning methods manually design query strategies so that the machine can select the most informative samples for model training one by one or in batches for labeling,thereby significantly reducing the cost of sample labeling.The single-sampling stratege is difficult to use for large-scale deep learning due to its low efficiency,while the batch-sampling stratege is easy to train on a large scale.However,the batch-sampling stratege also exposes the following two problems in deep learning scenarios.Firstly,the existing batch-mode methods are mostly designed for specific data or models,thus lacking wide applicability;secondly,these methods often fail to consider both the uncertainty and representativeness of the sample,which easily leads to a biased selection.To address the issues above,the main work of this thesis includes:(1)To learn a representative query strategy with general applicability,this thesis proposes an active learning method based on variational adversarial learning.This method is motivated by adversarial learning between variational encoder and query assistant model.After training in the variational adversarial manner,the query assistant model dynamically learns the distribution difference between the labeled samples and the unlabeled samples,thereby can select representative unlabeled samples for active learning.Thanks to the learnability of the query assistant,this method can adaptively adjust the query strategy according to the different states of the model and data.(2)To avoid the selective bias of representativeness and uncertainty,this thesis introduces model uncertainty into the above-mentioned active learning method derived from variational adversarial learning.On the one hand,the model prediction results and the variational latent vectors are feature-fused,and the new feature with the information of model uncertainty is obtained as the input of the query assistant model;on the other hand,the prediction of task loss is used as the uncertainty measure of the corresponding sample,which will be imposed on the query assistant by regularization.In this way,the proposed method can comprehensively select those samples that are more helpful for model learning by considering both representativeness and uncertainty.(3)To justify the advantages of the proposed active learning method,this thesis conducts comparative experiments on four common image classification data sets.Compared with existing active learning methods,the proposed method shows significant superiority in reducing labeling costs and improving sampling performance.Beside that,the visualization experiment and ablation experiment of the sampling results verify the rationality and effectiveness of the method design.