Investigating the Effect of Forgetting Factor on Tracking Non-Stationary Neural Dynamics

Abstract

Neural dynamics can be non-stationary in many neural applications such as brain-machine-interfaces (BMIs) and long-term brain stimulation. Adaptive modeling is a useful approach in tracking the neural non-stationarities. Our prior work has tracked time-variant linear state-space models (LSSM) to describe human electrocorticogram (ECoG) dynamics. One key design parameter in adaptive LSSM identification is the forgetting factor, which could significantly influence the accuracy of the fitted model. However, this influence has not been systematically studied yet. Here, we use comprehensive numerical simulations to investigate the effect of the forgetting factor on identifying time-variant LSSMs. We simulate non-stationary neural activity using time-variant LSSMs and use different forgetting factors to track time-variant LSSMs and predict the simulated activity. We find that the prediction accuracy of the fitted models strongly varied with the choice of the forgetting factor. We also find that the optimal forgetting factor that led to the highest prediction accuracy varied as a function of various properties of the simulated time-variant LSSMs. Our results have implications for building more accurate adaptive models to track non-stationary neural network dynamics and can facilitate the investigation of neural non-stationarity using the adaptive models.