Bidirectional Time-Frequency Pyramid Network for Enhanced Robust EEG Classification

TL;DR

The paper tackles poor cross-paradigm generalization in EEG-BCI systems by introducing Bite, a unified architecture that simultaneously processes temporal and spectral information via aligned STFT-based streams, a Pyramid Time-Frequency Attention module, and Bidirectional Temporal Convolutional Networks. Bite fuses multi-scale spectral-temporal features and bidirectional context through adaptive fusion, achieving state-of-the-art accuracy and robust cross-subject generalization across MI and SSVEP datasets with efficient computation. Key contributions include a dual-stream feature extractor, symmetric multi-scale PTFA, BiTCN with last-moment feature extraction, and learnable forward/backward fusion, validated on four divergent datasets. The results demonstrate that paradigm-aligned spectral-temporal processing significantly improves reliability and practicality of BCI systems, enabling plug-and-play deployment with strong within-subject performance and cross-subject transfer.

Abstract

Existing EEG recognition models suffer from poor cross-paradigm generalization due to dataset-specific constraints and individual variability. To overcome these limitations, we propose BITE (Bidirectional Time-Freq Pyramid Network), an end-to-end unified architecture featuring robust multistream synergy, pyramid time-frequency attention (PTFA), and bidirectional adaptive convolutions. The framework uniquely integrates: 1) Aligned time-frequency streams maintaining temporal synchronization with STFT for bidirectional modeling, 2) PTFA-based multi-scale feature enhancement amplifying critical neural patterns, 3) BiTCN with learnable fusion capturing forward/backward neural dynamics. Demonstrating enhanced robustness, BITE achieves state-of-the-art performance across four divergent paradigms (BCICIV-2A/2B, HGD, SD-SSVEP), excelling in both within-subject accuracy and cross-subject generalization. As a unified architecture, it combines robust performance across both MI and SSVEP tasks with exceptional computational efficiency. Our work validates that paradigm-aligned spectral-temporal processing is essential for reliable BCI systems. Just as its name suggests, BITE "takes a bite out of EEG." The source code is available at https://github.com/cindy-hong/BiteEEG.

Figures (4)

• Figure 1: Performance comparison across different approaches. The bubble plot provides a visualization of the trade-offs among performance, efficiency, and training time (bubble size). The results emphasize evaluations across different paradigms. The reported results are a composite of four datasets to ensure robustness and generalizability.
• Figure 2: Overview of the proposed Bite architecture including PTFA and Bi-TCN.
• Figure 3: t-SNE of feature embeddings for Subject 3 in BCICIV-2A.Bite exhibits superior separability between classes (distinct colors) and tighter clustering within classes compared to alternatives. Sub3 is selected due to convention.
• Figure 4: Ablation study results. Ablation study results showing per-subject accuracy (solid lines) and mean accuracy (right-side bar chart) across configurations. Different configurations are represented by lines of different colors, and their mean accuracies are displayed as bars with matching colors.