Yoon-Yeong Kim, Research on Improving Generalization of Deep Active Learning by Perturbing Input Space, PhD Dissertation, Department of Industrial and Systems Engineering, KAIST, 2023 

Abstract

Deep learning models have recently shown outstanding performance in various tasks, but they require a large amount of datasets in the training process. However, labeling a dataset requires a large among of compuational cost and human resources. Among various research directions to deal with the problem of annotation, active learning, which selects informative data instances via acquisition function, has been proposed for an efficient resource utilization. Although active learning efficiently trains a model with a small amount of data, it is vulnerable to generalization issue due to the bias inherent in the selectively collected dataset. Therefore, this dissertation proposes a study on improving generalization ability of deep active learning by perturbing input space to find informative instances. The first study applies data augmentation to perturb data space. In order to find informative data instances, we propose an acquisition function that looks ahead the future efficacy of data augmentation in advance, thereby increasing the efficacy of the virtual instances to be generated during training of the model. The study differentiate itself from the conventional pipelined combination of data augmentation and active learning by matching the goals of data augmentation and active learning. Furthermore, the study trains the augmentation policy with an objective of maximizing the acquisition score of active learning, so that we can maximize the informativeness of virtual data instances. The second study perturbs parameter space and uses the sharpness of the loss function emitted by the model with the transformed parameters as a criterion for judging the informativeness of the data. It has been proven through various studies that the loss sharpness has a strong correlation with the generalization ability of the deep learning model. Therefore, this study acquires the data instances that have the maximum value of the loss sharpness, by measuring the maximally perturbed loss value with the perturbed parameter. The data collected in this way is the most detrimental to the model from the point of view of generalization, and it can be expected that the generalization ability of the model will be strengthened by including this data in training. However, it is inaccessible to calculate the maximally perturbed loss because the dataset that is given in active learning is unlabeled. Therefore, the study calculates the acquisition score, or the maximally perturbed loss, using the pseudo label which is predicted by the current model. Also, the proposed acquisition function conservatively evaluates in that the acquired function calculated in this way is the lower bound of the acquisition function calculated using the ground-truth label. Moreover, through theoretical analysis, the acquisition function proposed in the study is shown to include the recently proposed active learning methods. Finally, by applying the two proposed studies to the vision-based tasks, which include object classification and object detection, this dissertation shows that the generalization performance is improved compared to the existing methodologies.

@phdthesis{Kim:2023,

author = {Yoon-Yeong Kim},

advisor ={Il-Chul Moon},

title = {Research on Improving Generalization of Deep Active Learning by Perturbing Input Space},

school = {KAIST},

year = {2023}

}