Constructing Robust Neural Decoders Using Limited Training Data

One of the essential components of a neuromotor prosthetic device is a neural decoder that translates the activity of a set of neurons into an estimate of the intended movement of the prosthetic limb. Wiener filter style approaches model this transformation as a linear function of the number of spikes observed from a set of neurons and over a range of distinct time bins. More recently, researchers have employed recursive Bayesian estimation techniques, such as Kalman filters, and have reported substantially better performance than with the Wiener filter. It is argued that this improvement in performance is due to the compact nature of these Bayesian models. Our results show that the poor performance of the Wiener filter is restricted to cases in which small training data sets are used, leading to substantial model overfitting. However, when training data sets are larger, we show that the Wiener filter is able to make appropriate use of the additional degrees of freedom to consistently outperform the Kalman filter. Finally, we suggest an alternative to the standard pseudo-inverse approach to solving for the Wiener filter parameters. The resulting algorithm almost always outperforms both of the previous approaches independent of the data set size.