Hits:

Title of Paper:A Lightweight Semantic Decoding Network with Group to Individual Transfer Learning for EEG-Based Visual Recognition

Journal:Neurocomputing

Key Words:EEG-based visual stimulus classification, Deep learning, Transfer learning, Multi-scale feature fusion, Redundant downsampling, Group to individual cognitive transfer

Summary:Decoding visual semantic features from electroencephalograph (EEG) signals is crucial for cognitive neuroscience and brain-computer interfaces (BCI). However, this task remains challenging due to the tension between group-level commonality and individual-specific variability in neural responses: existing deep learning models capture generic patterns from large-scale data but fail to adapt to individual differences, while their excessive complexity hinders deployment in resource-constrained environments. Additionally, pronounced inter-subject variability and data acquisition difficulties aggravate data scarcity and limit generalization. To address these challenges, this paper proposes a lightweight Multi-scale Spatial Pyramid and Temporal U-shaped (MSPTU) network for efficient EEG semantic feature extraction. Furthermore, a new Group-to-Individual Transfer Learning (G2I-TL) framework is introduced that leverages general group cognitive concepts to obtain individual-specific decoding models. Our method significantly reduces the number of parameters while improving accuracy, circumvents the inter-individual variability issue, and effectively mitigates data scarcity. Experiments on two public datasets (OCED and EEG200) demonstrate state-of-the-art (SOTA) performance; for example, on the OCED dataset, compared with the previous SOTA method, MSPTU-TL improves accuracy by 1.67% on the 6-class task and by 7.06% on the 72-exemplar task. Using the proposed masked channel attention method, even with only 20% of channels remaining, the accuracy still reaches 56.03% for the 6-Category task and 35.76% for the 72-Exemplar task, surpassing previous methods. This work provides a scalable solution for EEG-based visual recognition with practical BCI applications.

All the Authors:Doudou Zhang,Qimin Xu,Xinyu Cui,Rongkai Zhang,Yiming Jiang,Fu Li,Yang Li,Guangming Shi

First Author:Xiaotian Wang

Correspondence Author:Xiaotian Wang

Document Code:132377

Document Type:J

Issue:Q1, 二区, IF=6.0

Translation or Not:No

Date of Publication:2025-12-06

Date:2025-12-16

Impact Factor:6.0