Compressed and Distributed Sensing of Population Activity for Multiscale Decoding of Motor Cortical Response Properties

Estimating time varying response properties in the motor cortex is an essential task to decode motor neural activity for adequate and reliable neuroprosthetic control. Response properties are typically governed by the precision in neuronal firing and changes in the mean firing rate of individual neurons in relation to the preferred movement direction. This paper proposes a new approach for simultaneously estimating both types of information directly from a multiscale representation of neural data. The approach is based on exploiting the sparsity introduced in the data through a wavelet transformation to first derive distinctive features of neuronal spike waveforms in addition to precision in neuronal firing. Then, changes in mean firing rates are captured along an extended path of the representation to co-localize both response properties. We use spatiotemporal point process models to describe these properties and demonstrate the validity of the approach in decoding a 2D movement trajectory synthesized with 24 neurons directly from the sparsely represented data.