The paper explores the potential of Fourier analysis in generative modelling, specifically within time series diffusion models. The authors investigate if representing time series in the frequency domain is beneficial for score-based diffusion models. The paper introduces ‘frequency diffusion models’, showing they can better represent training distribution than traditional time diffusion models, particularly within healthcare and finance datasets. The researchers attribute this to time series data being more localized in the frequency domain. The study suggests potential synergies between Fourier analysis and diffusion models.
Publication date: 8 Feb 2024
Project Page: https://arxiv.org/abs/2402.05933v1
Paper: https://arxiv.org/pdf/2402.05933
