Hybrid LIF-ABRF spiking neural networks for fully neuromorphic spike detection in brain-machine interfaces

Abstract

To make implantable Brain Machine Interfaces (iBMI) systems wireless, it is necessary to use data compression given the limited bandwidth and power budget of implants. A recently popular method that tries to exploit the spatiotemporal sparsity of biological action potentials (AP) is the neuro-inspired conversion of analog signals directly to events or spikes such that datarate is proportional to neural activity. However, due to the noisy nature of the intracortical signal, the generated events do not directly correlate with biological APs and are often triggered by noise. Hence, there is a need to develop algorithms and architectures to perform efficient spike detection (SPD) for event-based neural recording frontends (EBNR). The use of neural networks, specifically Spiking Neural Network (SNN), which is a low power solution naturally interfacing with events, is a promising direction. In this work, we propose a lightweight hybrid SNN architecture that leverage Leaky Integrate-and-Fire (LIF) and Absolute Threshold Balanced Resonate-and-Fire Neuron (ABRF) spiking neurons to detect biological action potentials for EBNR. We show that LIF and ABRF extract different information from the input spike trains with LIF more sensitive to average firing rates while RAF more sensitive to detailed inter-spike intervals. The proposed SNN acts as both a feature extractor and classifier while operating in a sliding window fashion. We show this architecture is scalable and easy to realize on chip due to lower memory requirements than a streaming SNN. Our experiments show that the proposed model is able to achieve comparable result with much denser SNN models, using 4.3× less memory footprint which can be further compressed by up to 80% using sparsification training technique with accuracy loss of only ≈ 2%.