Neural Spike Sorting Using Binarized Neural Networks.

Area of Science:

• Neuroscience
• Computer Engineering
• Artificial Intelligence

Background:

• Conventional artificial neural networks (ANNs) require substantial memory and computational resources for neural spike sorting.
• Real-time processing of neural data from brain implants is challenging due to hardware limitations.

Purpose of the Study:

• To design and implement an efficient hardware system using binarized neural networks (BNNs) for brain-implantable neural spike sorting.
• To evaluate the performance and resource efficiency of BNNs compared to traditional ANNs in this application.

Main Methods:

• Developed a BNN architecture with binarized weights and activation functions.
• Trained the BNN using realistic neural datasets for spike sorting accuracy verification.
• Implemented the BNN on an application-specific integrated circuit (ASIC) and a field-programmable gate array (FPGA).

Main Results:

• The ASIC implementation occupied 0.33 mm² silicon area and consumed minimal power at 24 kHz.
• The FPGA implementation reduced on-chip memory requirements by 89% compared to state-of-the-art systems.
• Achieved accurate neural spike sorting with significantly reduced computational complexity.

Conclusions:

• BNNs offer a highly efficient solution for real-time neural spike sorting in brain-implantable devices.
• This work represents the first known application of BNNs for real-time in vivo neural spike sorting, demonstrating significant hardware advantages.