Domain Specific Denoising Diffusion Probabilistic Models for Brain Dynamics

Yiqun Duan, Jinzhao Zhou, Zhen Wang, Yu-Cheng Chang, Yu-Kai Wang, Chin-teng Lin

Keywords: Denoising Diffusion Probalistic Models, EEG Signal, Domain Variance Generation, Subject Difference, Deep Learning

Abstract: The distribution differences in brain dynamics according to different human subjects, which is a kind of human-subject noise, as referred to as human artifacts, have severely limited the generalized ability of brain dynamics recognition. Previous human artifact removal methods normally utilize traditional spectrum filtering or blind Source Separation techniques, based on a simple assumption of prior distributions, which limit the capacity of learning domain variance of each subject. We propose a new approach to model the removal of the human artifacts as a generative denoising process, which can generate and learn subject-specific domain variance and the invariant brain signals, simultaneously. We propose Domain Specific Denoising Diffusion Probabilistic Model (DS-DDPM) to decompose the denoising process into the subject domain variance and invariant content at each step. Subtle constraints and probabilistic design are proposed to formulate domain variance and invariant content into orthogonal spaces and further supervise the domain variance with the subject classifier. This method is the first work to explicitly separate human subject-specific variance through generative denoising processes, which outperforms previous methods in two aspects, 1) DS-DDPM could learn more accurate subject-specific domain variance by domain generative learning rather than previous filtering methods 2) DS-DDPM is the first work could explicitly generate subject noise distribution. Comprehensive experimental results suggest that DS-DDPM could help alleviate domain distribution bias for cross-domain brain dynamics signal recognition.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Neuroscience and Cognitive Science (e.g., neural coding, brain-computer interfaces)