Active Learning with Adaptive Reinforcement for Informative Sample Selection in Few-Shot Image Classification

Few-shot learning (FSL) has emerged as a pivotal area in machine learning due to its ability to perform well with a limited number of labeled samples, addressing common constraints in various domains, such as agriculture. While metric-based FSL methods, such as prototypical networks, leverage prior knowledge for effective classification, they often struggle with suboptimal performance due to the challenge of identifying and utilizing highly informative labeled samples. Active learning offers a potential solution by strategically selecting informative samples, but existing strategies often struggle in FSL due to their limited ability to capture data complexity. This paper introduces a strategy integrating reinforcement learning with hyperparameter optimization using distributed genetic algorithms for active learning in few-shot image classification. Applied to rice pest image classification, this approach demonstrates significant improvements in classification accuracy and sample efficiency, as evidenced by t-SNE visualizations and metrics such as accuracy, macro precision, macro recall, and macro F1-score. The method outperforms traditional query strategies, including uncertainty sampling, SVM-based approaches, query-by-committee, and random selection. The broader implications suggest that this research could be adapted to other fields where data scarcity is a significant challenge. Future work will explore the application of this method to different classifiers and domains, ensuring its scalability and adaptability.