Generating Visually Evoked Potentials Using a Diffusion Probabilistic Model

We present a diffusion probabilistic model (DPM) capable of generating high-quality electroencephalography (EEG) signals. In particular, we investigate the fidelity of synthesized visually evoked potentials and the performance of the novel class-conditioned continuous-time DPM that generates the signals. We carefully select our neural network architecture to prevent generated frequency artifacts in the multi-channel synthesized EEG signals. With progressive distillation, we are able to achieve a single-step diffusion model that still has a competitive performance. With generative metrics and statistical analysis, we show that using DPMs for brain signal generation is feasible.