计算机科学
人工智能
情绪识别
语音识别
模式识别(心理学)
噪音(视频)
特征(语言学)
匹配(统计)
心理学
人工神经网络
计算机视觉
作者
Songha Kim,Soyeon Bak,Jun-Mo Kim,Woohyeok Choi,Yebin Choi,Tae‐Eui Kam
出处
期刊:
[Institute of Electrical and Electronics Engineers]
日期:2026-04-21
卷期号:: 7191-7195
标识
DOI:10.1109/icassp55912.2026.11461997
摘要
EEG-based emotion recognition has gained attention for its ability to capture implicit and non-verbal emotional responses. Due to the cost and effort of EEG annotation, labeled data are limited, thereby motivating subject-independent semi-supervised learning. Particularly, prototype-based methods have shown promise using class-wise feature centroids to guide pseudo-label assignment. However, in early training, prototypes computed from the sparse and noisy samples are often unreliable, leading to inaccurate pseudo-labels and confirmation bias. To address these challenges, we propose CASPER, a confidence-aware adaptive semi-supervised framework that jointly enhances the reliability of pseudo-labels and robustness of prototypes. CASPER introduces a confidence bank that accumulates class-wise confidence scores over time and dynamically guides the two-stage pseudo-label adjustment process. During early training, a debiased constraint emphasizes low-confidence samples to encourage broad class exploration, while later, a stabilized constraint sharpens decision boundaries using reliable predictions. In parallel, CASPER applies distance-weighted prototype updates, where each sample is weighted based on its distance from the prototype, reducing the impact of outliers and yielding more representative class centers. Experiments on the SEED and SEED-IV datasets demonstrate that CASPER achieves state-of-the-art performance, validating the complementary effects of confidence-aware pseudo-label adjustment and robust prototype updates.
科研通智能强力驱动
Strongly Powered by AbleSci AI