Clustering and observation on neuron tuning property for brain machine interfaces

Neurons' tuning describes how the neural activity responses to the stimulus. As the prior knowledge, understanding more about the neuron tuning helps better decode the movement information from the neural firings for brain machine interfaces. We are interested in qualifying the neural tuning and observe whether there are similar tunings among the ensemble recordings and how they change over time. We propose to implement a linear-nonlinear-Poisson model to describe the neural tuning function. And the function parameters build a feature space, where the neuron tuning characters can be visually observed. We use k-means algorithm to cluster neuron tuning characters and find that there are three types of neurons with different tuning curve shapes. The nonlinear-shaping neurons are not majority in number but have important contribution (evaluated by mutual information) relative to the movement task than the linear ones. Furthermore, we find some neuron tunings shows clear time-varying properties in the feature space, which can be predicted by a random walk model. And we prove it through two kinds of way: Kernel size-CC estimation and Kolmogorov-Smirnov plot (KS plot). The predictable time-varying tuning suggests a better understanding of neuron property and potentially contributes to decode the non-stationary neuron activities.