EEGEncoder: Advancing BCI with Transformer-Based Motor Imagery Classification

TL;DR

This paper innovatively proposes a fusion architecture, namely Dual-Stream Temporal-Spatial Block (DSTS), to capture temporal and spatial features, improving the accuracy of Motor Imagery classification task and uses multiple parallel structures to enhance the performance of the model.

Abstract

Brain-computer interfaces (BCIs) harness electroencephalographic signals for direct neural control of devices, offering a significant benefit for individuals with motor impairments. Traditional machine learning methods for EEG-based motor imagery (MI) classification encounter challenges such as manual feature extraction and susceptibility to noise.This paper introduces EEGEncoder, a deep learning framework that employs modified transformers and TCNs to surmount these limitations. We innovatively propose a fusion architecture, namely Dual-Stream Temporal-Spatial Block (DSTS), to capture temporal and spatial features, improving the accuracy of Motor Imagery classification task. Additionally, we use multiple parallel structures to enhance the performance of the model. When tested on the BCI Competition IV-2a dataset, our model results outperform current state-of-the-art techniques.

Figures (3)

• Figure 1: Architecture of the EEGEncoder. The figure illustrates the data processing pipeline within the EEGEncoder, highlighting the novel application of parallel dropout layers to enrich the diversity of the hidden state representations.
• Figure 2: Architecture of the Downsampling projector. The figure provides a detailed schematic of the Downsampling projector's architecture. It includes three convolutional layers, with the second and third layers each followed by a batch normalization (BN) layer and an ELU activation layer. Additionally, two average pooling layers and two dropout layers are incorporated to foster model generalization. Specific parameters, such as the kernel size and stride for the convolutional layers, and the kernel size for the average pooling layers, are also depicted. For example, "Conv 1x16, (64,1)" signifies a convolutional layer transitioning from an input channel depth of 1 to an output channel depth of 16, with a stride of 64 along the width and 1 along the height of the input feature map.
• Figure 3: Architecture of the DTDS Block. The DTDS Block integrates a TCN for local temporal feature extraction with a self-attention block for global spatial context analysis, enabling a detailed examination of EEG signals for MI classification tasks.