Invited Research Talk: Measuring Generalization in EEG Foundation Models
Episode details
Episode description
Over the past few years, there has been an increasing interest towards building EEG foundation models that provide generalized representations for EEG signals, which are able to provide noticeable performance improvements over supervised baseline methods. While there have been numerous models that have been proposed, limited work has been performed in evaluation and interpretability of these foundation models, which could further help guide future model building. This talk will discuss a multi-dimensional evaluation framework for EEG foundation models, based on parameter, sample and channel efficiency that reflect real-world constraints in EEG deployment. Through this evaluation, we show that in a multitude of aspects, EEG foundation models fail to show performance gains over supervised baselines. Additionally, through a simulation based framework, synthetic EEG signals with controllable oscillatory and aperiodic components are used as inputs to EEG foundation models, to identify whether EEG foundation models capture both aperiodic and oscillatory components. We find that most reconstruction based EEG foundation models tend to have an aperiodic low frequency bias which is reflected in the models capability to better capture subject specific constructs over task specific information. Altogether, our analysis highlights deficiencies in current EEG foundation models and sheds light on missing components that could help improve generalization. See more at https://www.microsoft.com/en-us/research/video/invited-research-talk-measuring-generalization-in-eeg-foundation-models/
Watch also from the same channel
More episodes from this channel
You might be interested in
