|M.Sc Student||Even-Chen Nir|
|Subject||Intrinsic Modeling of Repetitive Time Series with|
Applicarion for Analysis of Event-Related
|Department||Department of Electrical Engineering||Supervisors||Professor Ron Meir|
|Professor Miriam Zacksenhouse|
|Full Thesis text|
Modeling and analyzing biological time series (e.g. Electroencephalography (EEG)) are challenging tasks due to their non-stationary, nonlinearity, low signal-to noise ratio (SNR) and the absence of a clear ground truth.
In this work we are interested in analyzing EEG in order to detect event-related potentials (ERP). Most analyses of EEG dynamics are based on parametric linear models (e.g. AR, RLS etc.) which may not reflect the true dynamics. Standard ERP analysis provides limited results of the grand average. However, single-trial analysis is much more challenging.
We propose a manifold-learning approach to analyze data from repetitive trials and to reveal the main intrinsic processes behind the measurements' dynamics. Manifold-learning is mostly applied to data based on independent samples, and was recently extended to take temporal information into account. We further extend this approach to account for trial-to-trial variability. Accounting for both temporal and trial-to-trial variations holds the promise for a suitable local metric to reveal the underlying invariant dynamics.
Using a manifold-learning approach on the averaged ERP revealed an intrinsic component of P300 that distinguishes between different stimuli in a speller task. A toy example and EEG recordings demonstrate the advantage of our extension in the case of multiple trials. Our approach reveals the invariant dynamics shared by repetitive trials and improves the tradeoff between temporal resolution and noise robustness.