In this paper, we present a novel methodology for evaluating the performance of active learning (AL), the Performance Analysis of Active Learning Models (PALM). To overcome the limitations of existing AL evaluation methods that focus only on final accuracy, PALM presents an integrated and interpretable mathematical model that characterizes the dynamics of AL learning through four key parameters: achievable accuracy, coverage efficiency, initial performance, and scalability. It predictively explains the behavior of AL from partial observations, allowing for estimation of future performance and principled comparisons between different strategies. It is validated through extensive experiments using various AL methods and self-supervised learning embeddings on the CIFAR-10/100 and ImageNet-50/100/200 datasets, and shows generalization performance that accurately predicts the entire learning curve from limited labeled data. PALM provides important insights into the learning efficiency, data space coverage, and scalability of AL methods, enabling cost-effective strategy selection and performance prediction under budget constraints. It thus lays the foundation for systematic, reproducible, and data-efficient evaluation of AL in both research and real-world applications. The source code is available on GitHub.
