Active learning using Generative Adversarial Networks for improving generalization and avoiding distractor points

In supervised computer vision tasks, convolutional neural networks (CNNs) have demonstrated superiority over alternative methods. However, training and validating these models requires large-scale labeled datasets, which are generated using specific expert knowledge and are expensive. With the limitation of a limited labeling budget, active learning is a promising approach for generating labeled datasets. The conventional methods used for active learning only consider a pool of unlabeled data that are relevant to the target task. However, in several real-world tasks such as web search results, this pool includes several data points that are irrelevant to the target task, which are called distractor points. These tend to largely degrade the effectiveness of active learning. To address this problem, we propose a method to avoid distractor points. The proposed method utilizes criteria that incorporates existing active learning methods and an effective training strategy using Generative Adversarial Networks (GANs) to eliminate distractor points from the area including the selection criteria. As a result, the proposed method successfully divides the data points into data points utilized in active learning and distractor points to improve generalization and decrease the accuracy degradation of the system. In the benchmark verification process, we show that the proposed method is effective when the number of distractor points in the considered datasets is small or large.

Automated summary

In supervised computer vision tasks, CNNs have proven to be superior to alternative methods. However, generating large-scale labeled datasets for training and validating these models is costly and requires specific expert knowledge. Active learning is a promising approach to generate labeled datasets with limited labeling budgets. This study addresses the issue of distractor points in active learning, proposing a method that effectively eliminates distractor points using a combination of existing active learning methods and a training strategy with GANs.