On the use of compressive sensing (CS) exploiting block sparsity for neural spike recording

Abstract

This paper presents a novel compressive sensing (CS) algorithm for neural spike recording that exploits the concept of block sparsity in both dictionary training and signal reconstruction. Initially, the block K-SVD (BK-SVD) algorithm is employed to train a block-sparsifying dictionary for neural spikes, followed by the block sparse Bayesian learning (BSBL) algorithm to reconstruct the compressed data using the trained dictionary. Using various neural spike datasets for validation and across compression ratio, CR, values in the range of 2 to 16, the combined BK-SVD + BSBL algorithm is shown to improve the reconstruction signal-to-noise and distortion ratio (SNDR) by ∼ 0.5 to 6.5dB over K-SVD + basis pursuit (BP) algorithm that does not exploit block sparsity. Next, the combined BK-SVD + BSBL algorithm is augmented with spike detection for a more practical implementation. The resulting block sparse CS with spike detection (BSCS + SD) algorithm achieves an overall CR of ∼11.6 and SNDR that is up to 8dB higher than that obtained without spike detection.