Comparative Analysis of Active Learning Algorithms for Data-Efficient Image Classification

Abstract

The increasing demand for labelled data in supervised image classification tasks poses significant challenges in terms of time, cost, and scalability. This paper investigates the application of active learning techniques to enhance data efficiency in image classification. Using the MNIST benchmark dataset, we implement and evaluate three query strategies Random Sampling, Uncertainty Sampling, and Margin Sampling within a unified active learning framework. Experimental results demonstrate that Uncertainty and Margin Sampling strategies can achieve 90% test accuracy while requiring approximately 33% fewer labelled instances than Random Sampling. Visual analyses of model confidence distributions and learning curves further validate the superior label efficiency of these strategies. The findings highlight the potential of active learning as a practical solution for reducing annotation overhead in data-constrained machine learning scenarios.