3 September 2024 in IEEE Transactions on Big Data
Accurate sleep staging is crucial for the diagnosis of diseases such as sleep disorders. Existing sleep staging models with excellent performance are usually large and require a lot of computational resources, limiting their application on wearable devices. Therefore, it is a key issue to distil the knowledge embedded in large models into small heterogeneous models for better deployment. In the process of knowledge distillation of heterogeneous models for sleep EEG signals, we mainly deal with three major challenges: 1) There are large structural differences between heterogeneous sleep staging models; 2) What kind of knowledge should be conveyed in sleep EEG signals in the knowledge distillation of heterogeneous models; 3) Significant scale differences exist between heterogeneous models. To address these challenges, we design a generic heterogeneous model knowledge distillation framework for sleep staging. Specifically, we first propose a knowledge distillation strategy for heterogeneous models that addresses the large structural differences between heterogeneous models. Then, a multi-level knowledge distillation module is designed to effectively transfer important multi-level feature knowledge. In addition, the teacher assistant module is introduced to ease the scale difference between the heterogeneous models which further enhances the knowledge distillation performance. Experimental results on both Sleep-EDF and ISRUC datasets show that our distillation framework achieves state-of the-art performance.
7 May 2026 in IEEE Transactions on Affective Computing
Emotion recognition based on multi-modal physiological signals has recently attracted significant attention. However, challenges remain in multimodal fusion, handling individual differences, and managing the large model sizes in this domain. To address these challenges, we propose a lightweight cross-subject multi-modal emotion recognition model, tiny-MMDNet. This model reduces model parameters by employing a dynamic low-rank approximation technique within a multi-level disentangling framework. Specifically, tiny-MMDNet consists of a modality-level disentangling module and a subject-level disentangling module. The modality-level disentangling module projects the raw multi-modal physiological signals into modality-invariant and modality-specific subspaces, capturing both modality-invariant and modality-specific features. Meanwhile, the subject-level disentangling module separates subject-shared and subject-private features within the modality-level features to enhance cross-subject emotion recognition. Furthermore, tiny-MMDNet dynamically compresses the model through dynamic low-rank approximation, reducing model parameters while maintaining performance. Experimental results demonstrate that tiny-MMDNet achieves superior cross-subject recognition accuracy and effective model compression on public available multi-modal emotion datasets, outperforming current state-of-the-art models.