Low-rank representation of neural activity and detection of submovements

In this study, Robust Principal Component Analysis (RPCA) is applied to neural spike datasets to extract neural signatures that signify the onset of submovements, a type of motor primitive. Given neural activity recorded from rhesus macaques during a set of reaches between targets in a horizontal plane, we aim to identify common event-related neural features and validate submovement-based motor primitives inferred from the hand velocity profiles. Such features represent common dynamic patterns across many experimental trials and may be used as a signature to detect discrete events such as submovement onset. We present RPCA, a method well suited for extracting data matrices´ low-rank component and this method allows (1) removal of task-irrelevant signal from data, (2) identification of task-related dynamic patterns, and (3) detection of submovements. We also explored using the Random Projection (RP) technique and applying RP to data prior to applying RPCA improved the submovement prediction performance by de-sparsifying neural data while preserving certain statistical characteristics of aggregate neural activity.