Generating realistic neurophysiological time series with denoising diffusion probabilistic models
Julius Vetter, Jakob H. Macke, Richard Gao
Abstract
Denoising diffusion probabilistic models (DDPMs) have recently been shown to accurately generate complicated data such as images, audio, or time series. Experimental and clinical neuroscience also stand to benefit from this progress, as the accurate generation of neurophysiological time series can enable or improve many neuroscientific applications. Here, we present a flexible DDPM-based method for modeling multichannel, densely sampled neurophysiological recordings. DDPMs can generate realistic synthetic data for a variety of datasets from different species and recording techniques. The generated data capture important statistics, such as frequency spectra and phase-amplitude coupling, as well as fine-grained features such as sharp wave ripples. Furthermore, data can be generated based on additional information such as experimental conditions. We demonstrate the flexibility of DDPMs in several applications, including brain-state classification and missing-data imputation. In summary, DDPMs can serve as accurate generative models of neurophysiological recordings and have broad utility in the probabilistic generation of synthetic recordings for neuroscientific applications.