Scholarlink

<p>Autism Spectrum Disorder is a neurological condition characterized by symptoms affecting communication and social interaction. Analysing the facial expressions of autistic children is crucial for understanding their emotional states, which is essential for both diagnosis and effective intervention. This work proposes a specialized transformer architecture with a teacher–student knowledge distillation framework to enhance emotion detection in autistic children. A context specific multi-head attention mechanism was introduced in the teacher model. The student model was built based on dynamic window partitioning. The uncertainty- aware knowledge distillation is introduced to learn from the teacher model. Before knowledge distillation, the model achieved an accuracy of 0.77 in emotion detection. The student model was applied after applying the uncertainty-based knowledge distillation, and the accuracy of the student model increased by 12%, to 0.89. The precision, recall, and F1 score of the model were shown to be significantly improved. Combining uncertainty-aware knowledge distillation from the teacher to the student and context-aware multi-head attention in the baseline teacher model, the performance of the student model was significantly improved.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Contextual Multi-Head Transformer with Uncertainty-Aware Distillation for Emotion Recognition in Autistic Children

  • 期刊论文

  • 发表日期：

    2026年4月9日

查看全文
  • J. Christina,
  • J. Dhalia Sweetlin

摘要

Autism Spectrum Disorder is a neurological condition characterized by symptoms affecting communication and social interaction. Analysing the facial expressions of autistic children is crucial for understanding their emotional states, which is essential for both diagnosis and effective intervention. This work proposes a specialized transformer architecture with a teacher–student knowledge distillation framework to enhance emotion detection in autistic children. A context specific multi-head attention mechanism was introduced in the teacher model. The student model was built based on dynamic window partitioning. The uncertainty- aware knowledge distillation is introduced to learn from the teacher model. Before knowledge distillation, the model achieved an accuracy of 0.77 in emotion detection. The student model was applied after applying the uncertainty-based knowledge distillation, and the accuracy of the student model increased by 12%, to 0.89. The precision, recall, and F1 score of the model were shown to be significantly improved. Combining uncertainty-aware knowledge distillation from the teacher to the student and context-aware multi-head attention in the baseline teacher model, the performance of the student model was significantly improved.