Detecting Concept Shifts Under Different Levels of Self-awareness on Emotion Labeling

초록

Generalizing deep learning for all requires individual self-assessment. However, the quality of ground-truth labels depends on the annotators’ self-awareness. Real-world datasets inevitably experience the Concept Shift problem. Recent advances in Out-of-distribution (OOD) detection have received much attention due to its ability to alleviate distribution shift problems by distinguishing between anomalous and in-distribution(ID) data samples. Existing approaches underlie pre-trained ID models learned with class-balanced data. However, this assumption makes the methods incapable when the ID models are trained with inter- and intra-class variance depending on user characteristics, such as gender, culture, and genetics. We present an OOD detection framework. Our system builds a generalized ID model by extracting high-quality data from high-dimensional neural activities considering individuals’ cognitive and perceptional ability to evaluate self-assessments. The proposed system detects and removes abnormal pairs of data and labels to enhance model performance by considering the maximum softmax probability approach. Experimental results on public EEG datasets in emotion recognition demonstrate the superiority of our method despite the non-stationary nature of EEG signals. The codes are available at https://github.com/affctivai/coglier. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2025.

키워드