On Robust Grouping Active Learning

Abstract

Early active learning, in a common paradigm, usually selects representative samples for human annotating. This aligns with the goal of minimizing the overall reconstruction error in an unsupervised manner. While existing methods mainly focus on data samples that are drawn from individual yet high-dimensional feature space, they can hardly handle the real-world scenario where samples are often represented by low-dimensional features drawn from multiple groups (subspaces). In this case, how to leverage the grouping structure to select most representative samples becomes the key point to success. In this paper, we propose an unsupervised active learning framework, called Robust Grouping Active Learning (RGAL), to achieve this goal. The key idea is to take into account of different degrees of information shared across data groups. Specifically in RGAL, we assume data from some group can be embedded in a low-dimensional space, as well as that the data distributions of different groups can overlap with each other to a certain degree. And RGAL controls such group overlaps by imposing sparsity constraints on a matrix of reconstruction coefficients. To encourage a smooth coefficient space, we also enforce a robust loss with Laplacian regularization for noise suppression. We perform extensive experiments on multiple tasks which normally require costly human annotation, including facial age estimation, video action recognition and medical image classification. Results on benchmark datasets clearly demonstrate the efficacy of our RGAL method compared state-of-the-art methods.