Robust models for neural signal processing

In recording high-frequency biological signals such as neural activity, noise may be difficult to distinguish from useful signal content. Conventional noise-removal techniques such as autoregressive modeling do not perform well on impulsive signals such as neural activity. Impulsive data can be more effectively extracted using p -normed error models, where p=2 corresponds to the least squares model, and p=1 corresponds to the least absolute value case. The least absolute value model is best when the model error is Laplace distributed. Thus, a judicious choice of p-normed model will allow outliers, such as the spikes from neural activity, to be passed through the algorithm while other types of noise are suppressed