Uncertainty-Based Active Learning via Sparse Modeling for Image Classification.

Area of Science:

• Machine Learning
• Computer Vision

Background:

• Active learning aims to improve classifier performance by selecting informative samples.
• Batch-mode active learning selects multiple samples simultaneously, often prioritizing high uncertainty.
• Existing methods may select redundant samples, ignoring inter-sample relationships.

Purpose of the Study:

• To propose a novel batch-mode active learning method that integrates uncertainty, diversity, and density.
• To address the redundancy issue in sample selection by incorporating sample relationships.
• To develop an efficient sparse modeling approach for sample selection.

Main Methods:

• A novel method combining uncertainty, diversity, and density using sparse modeling.
• Utilizing sparse linear combination with Gaussian kernels to represent unlabeled data uncertainty.
• Incorporating selective sampling before optimization to minimize representation error.
• Employing two approximated approaches for efficient optimization of the L0 norm constraint.

Main Results:

• The proposed method effectively combines uncertainty, diversity, and density in sample selection.
• Experiments on four image classification datasets demonstrate the method's advantages.
• Extensive analysis confirms benefits across various parameters like batch size and feature space.

Conclusions:

• The novel sparse modeling approach enhances batch-mode active learning by reducing sample redundancy.
• The method offers improved performance and efficiency compared to traditional uncertainty sampling.
• This work provides a robust framework for informative sample selection in machine learning.