Memory efficient on-line streaming for multichannel spike train analysis.

Rapid advances in multichannel neural signal recording technologies in recent years have spawned broad applications in neuro-prostheses and neuro-rehabilitation. The dramatic increase in data bandwidth and volume associated with multichannel recording requires a significant computational effort which presents major design challenges for brain-machine interface (BMI) system in terms of power dissipation and hardware area. In this paper, we present a streaming method for implementing real-time memory efficient neural signal processing hardware. This method exploits the pseudo-stationary property of neural signals and, thus, eliminates the need of temporal storage in batch-based processing. The proposed technique can significantly reduce memory size and dynamic power while effectively maintaining the accuracy of algorithms. The streaming kernel is robust when compared to the batch processing over a range of BMI benchmark algorithms. The advantages of the streaming kernel when implemented on field-programmable gate array (FPGA) devices are also demonstrated.