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

Area of Science:

• Neuroscience and Biomedical Engineering
• Hardware and Signal Processing

Background:

• Advances in multichannel neural recording generate large data volumes, posing computational challenges for brain-machine interface (BMI) systems.
• High data bandwidth necessitates efficient processing hardware, particularly concerning power dissipation and physical size.

Purpose of the Study:

• To present a real-time, memory-efficient neural signal processing hardware implementation method.
• To address the computational and hardware design challenges in BMI systems.

Main Methods:

• Developed a streaming method for neural signal processing that leverages the pseudo-stationary property of neural signals.
• Eliminated the need for temporal storage typically required in batch-based processing techniques.
• Implemented the streaming kernel on field-programmable gate array (FPGA) devices.

Main Results:

• Significantly reduced memory size and dynamic power consumption compared to traditional methods.
• Effectively maintained the accuracy of signal processing algorithms.
• Demonstrated robustness of the streaming kernel across various BMI benchmark algorithms.

Conclusions:

• The proposed streaming method offers a viable solution for memory-efficient, low-power neural signal processing hardware.
• This technique is well-suited for real-time applications in neuro-prostheses and neuro-rehabilitation.
• FPGA implementation validates the practical advantages of the streaming kernel for BMI systems.