A lightweight multimodal framework for emotion-aware voice interaction on edge devices, integrating EEG and facial-expression signals with model compression to support adaptive and efficient on-device responses.

ASH-MIL is a weakly-supervised framework for thoracic disease detection in chest X-rays that combines parallel anatomy-structured branches with hierarchical MIL, injecting anatomical priors into decoder cross-attention to enable anatomically grounded localization without bounding box supervision.

We propose a two-stage active learning pipeline for semantic segmentation under extremely low labeling budgets, which leverages a pre-trained diffusion model for diverse feature extraction and introduces an uncertainty score to select the most informative pixels for annotation.