Background Spike sorting, which isolates single neuron activities from multi-unit neural recordings, plays a pivotal role in neuroscience research. Recently, advanced recording technologies have enabled large-scale neural signal recording involving thousands of channels synchronously. This advancement has led to a growing demand for computationally efficient spike sorting approaches, particularly for brain-implantable devices where power and memory resources are highly constrained.Methods We propose QuanSort, a spiking neural network (SNN)-based spike sorter designed for energy-efficient on-chip computation with neuromorphic chips. It employs a novel "reset-to-mod" membrane potential reset mechanism. This strategy utilizes suprathreshold states during computation, which preserves higher precision in the value quantization process during on-chip deployment while theoretically ensuring the bounded accumulation of quantization error under the Weyl discrepancy norm.Results Experimental evaluations conducted on both synthetic and real-world datasets demonstrate that QuanSort can be effectively quantized to 16-bit integers. Despite this aggressive quantization, the model maintains high spike sorting performance, validating its suitability for resource-constrained environments.Conclusion QuanSort offers a robust and energy-efficient solution for large-scale spike sorting on neuromorphic hardware. By combining the reset-to-mod mechanism with SNN architecture, it successfully balances computational precision with the strict power and memory limitations inherent to implantable brain-computer interfaces.
Yu et al. (Tue,) studied this question.