Abstract: Generating high-fidelity EEG data at scale is essential for overcoming dataset scale limitation and satisfying privacy requirements in computational neuroscience. However, the predominant reliance on discrete denoising formulations fails to adequately capture the continuous temporal evolution and frequency-domain characteristics intrinsic to EEG recordings. Consequently, such approaches often compromise long-range temporal coherence and introduce structural discrepancies in both spectral and temporal domains.
In this work, We argue that faithful EEG synthesis demands generative models that directly characterize the continuous dynamics underlying neural signals. To this end, we propose a conditional flow matching framework that treats EEG as raw waveform sequences traversing continuous-time trajectories. Rather than relying on discretized denoising steps or handcrafted signal representations, our method learns a smooth velocity field mapping noise distributions to the target EEG manifold, thereby naturally preserving temporal continuity and transient neural phenomena. To enforce fidelity to fundamental EEG characteristics, we incorporate principled constraint terms that maintain spectral consistency, temporal stationarity, and signal-level statistical properties. Evaluated on large-scale benchmarks, our approach establishes new state-of-the-art results compared to competitive baselines. Comprehensive analyses further confirm that the proposed framework faithfully recovers core structural attributes of neural dynamics, offering a scalable and theoretically grounded solution for high-fidelity EEG generation.
Speaker: Yifan Wang
Location: https://stonybrook.zoom.us/j/
Meeting ID: 953 0027 9172
Passcode: 418725