Committee-Based Threshold Classification Algorithm with Reinforcement Active Learning

• Published in: 2024 IEEE 2nd International Conference on Control, Electronics and Computer Technology (ICCECT) pp. 63 - 67
• Main Authors: Zhao, Dan, Shen, Hujun
• Format: Conference Proceeding
• Language: English
• Published: IEEE 26.04.2024
• Subjects: Active learning; classification algorithm; Classification algorithms; Couplings; Data preprocessing; deep learning; Prediction algorithms; Reinforcement learning; sampling method; Sampling methods; Training
• DOI: 10.1109/ICCECT60629.2024.10546099

The deep learning-based computing paradigm plays a pivotal role in contemporary society. Nevertheless, current methods are plagued by the challenges of demanding extensive labeled data and excessive training resource consumption. Active learning (AL) diminishes the quantity of required training and labeling data by selecting representative samples. Nonetheless, traditional AL approaches demand a significant number of iterations and sample labels, making them susceptible to data bias and excessive human intervention. In response to these challenges, this paper presents a threshold classification algorithm based on query by committee (TC-QBC) that integrates reinforcement learning techniques. Unlike conventional sampling methods, TC-QBC assigned dynamic weights to labels during committee voting. Subsequently, a resampling operation was executed by establishing a threshold to attain the data classification objective. Following validation in pertinent experiments, TC-QBC exhibits outstanding classification prowess. Our work has shown that the proposed method can improve model performance while reducing the resources necessary for training.