Various Active Learning Strategies Analysis in Image Labeling: Maximizing Performance with Minimum Labeled Data

The use of active learning in supervised machine learning is proposed in this study to reduce the expenses associated with labeling data. Active learning is a technique that includes iteratively selecting the most informative unlabeled data points and asking a human expert to label them. Active learning can achieve high accuracy while utilizing fewer labeled examples than typical supervised learning algorithms by selecting the most informative data points. This study conducts and provides an in-depth examination and analysis of numerous active learning algorithms and their applications to various machine learning labeling problems, especially focusing on image classification. The experiments are carried out using Fashion MNIST as a benchmark dataset. This study compares the performance of five popular active learning methods BALD, DBAL, coreset, least confidence and ensemble varR for the given problem. The best performing algorithm was BALD with a mean classification accuracy of 91.31%, when 50% of the data is considered labeled, closely followed by all other techniques, making each suitable for specific use cases. The trials conducted by the study illustrates how active learning may lower the time and cost of data labeling while also maintaining high accuracy.