From Rest to Action: Adaptive Weight Generation for Motor Imagery Classification from Resting-State EEG Using Hypernetworks
Keywords: Brain-Computer Interfaces (BCIs), Motor Imagery, HyperNetworks, Data driven learning, Adaptive weights
TL;DR: We propose an architecture that leverages resting state EEG data for generating weights via hypernetworks for downstream task like Motor Imagery
Abstract: Existing EEG-based brain-computer interface (BCI) systems require long calibration sessions from the intended users to train the models, limiting their use in real-world applications. Additionally, despite containing user-specific information and features correlating with BCI performance of a user, resting-state EEG data is underutilized, especially in motor imagery decoding tasks. To address the challenge of within and across-user generalisation, we propose a novel architecture, HyperEEGNet, which integrates HyperNetworks (HNs) with the EEGNet architecture to adaptively generate weights for motor imagery classification based on resting-state data. Our approach performs similarly in a Leave-Subject-Out scenario using a dataset with 9 participants, compared to the baseline EEGNet. When the dataset size is scaled, with 33 participants' datasets, the model demonstrates its generalisation capabilities using the information from resting state EEG data, particularly when faced with unseen subjects. Our model can learn robust representations in both cross-session and cross-user scenarios, opening a novel premise to leverage the resting state data for downstream tasks like motor imagery classification. The findings also demonstrate that such models with smaller footprints reduce memory and storage requirements for edge computing. The approach opens up avenues for faster user calibration and better feasibility of edge computing, a favourable combination to push forward the efforts to bring BCIs to real-world applications.
Primary Area: applications to neuroscience & cognitive science
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: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
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.
Submission Number: 14090
Loading