Scholarlink

This paper presents GViT, a hybrid CNN-Transformer architecture designed to improve EEG-based gaze prediction by leveraging spatial-temporal representations of brain signals. GViT integrates convolutional layers to extract local spatial features with a transformer encoder that models global temporal dependencies, enabling robust performance on noisy EEG data. Evaluated on the EEGEyeNet dataset, GViT consistently achieves the lowest gaze prediction error among all tested models, outperforming baseline CNN, GRU, and transformer variants. By bridging neuroscience-inspired design and deep learning advances, this work demonstrates the effectiveness of hybrid architectures for brain signal decoding and introduces a modular framework applicable to a broad range of neurophysiological time-series tasks.

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

GViT: Combining Convolutional and Transformer Layers for Spatial-Temporal EEG Analysis

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Chenxu Zhu,
  • Yiming Xu,
  • Xiaodong Qu

摘要

This paper presents GViT, a hybrid CNN-Transformer architecture designed to improve EEG-based gaze prediction by leveraging spatial-temporal representations of brain signals. GViT integrates convolutional layers to extract local spatial features with a transformer encoder that models global temporal dependencies, enabling robust performance on noisy EEG data. Evaluated on the EEGEyeNet dataset, GViT consistently achieves the lowest gaze prediction error among all tested models, outperforming baseline CNN, GRU, and transformer variants. By bridging neuroscience-inspired design and deep learning advances, this work demonstrates the effectiveness of hybrid architectures for brain signal decoding and introduces a modular framework applicable to a broad range of neurophysiological time-series tasks.