Benchmarking Neural Decoding Backbones towards Enhanced On-edge iBCI Applications

TL;DR

This work addresses the challenge of running neural decoding on edge devices for everyday iBCI usage by systematically benchmarking four backbones—GRU, Transformer, RWKV, and Mamba—on a nonhuman primate random-reach task. The study demonstrates that RWKV and Mamba provide faster inference and calibration with favorable scaling, while GRU remains competitive in some scenarios but suffers from slower processing and limited scalability, and Transformer tends to be less tractable on edge hardware. Scaling analyses show RWKV and Mamba improve with increased model size up to several million parameters, suggesting strong potential for future edge deployments as data availability grows. The findings offer practical guidance for selecting edge-friendly neural decoding backbones and highlight avenues for further research in online learning and cross-task/adaptive decoding.

Abstract

Traditional invasive Brain-Computer Interfaces (iBCIs) typically depend on neural decoding processes conducted on workstations within laboratory settings, which prevents their everyday usage. Implementing these decoding processes on edge devices, such as the wearables, introduces considerable challenges related to computational demands, processing speed, and maintaining accuracy. This study seeks to identify an optimal neural decoding backbone that boasts robust performance and swift inference capabilities suitable for edge deployment. We executed a series of neural decoding experiments involving nonhuman primates engaged in random reaching tasks, evaluating four prospective models, Gated Recurrent Unit (GRU), Transformer, Receptance Weighted Key Value (RWKV), and Selective State Space model (Mamba), across several metrics: single-session decoding, multi-session decoding, new session fine-tuning, inference speed, calibration speed, and scalability. The findings indicate that although the GRU model delivers sufficient accuracy, the RWKV and Mamba models are preferable due to their superior inference and calibration speeds. Additionally, RWKV and Mamba comply with the scaling law, demonstrating improved performance with larger data sets and increased model sizes, whereas GRU shows less pronounced scalability, and the Transformer model requires computational resources that scale prohibitively. This paper presents a thorough comparative analysis of the four models in various scenarios. The results are pivotal in pinpointing an optimal backbone that can handle increasing data volumes and is viable for edge implementation. This analysis provides essential insights for ongoing research and practical applications in the field.

Figures (2)

• Figure 1: Raw neural signals were recorded from the primary motor cortex (M1) area of a monkey using a 96-channel Utah microelectrode array during random reach tasks. Spike activity detected from these neural signals was binned temporally across the 96 channels. The resulting matrix of spike counts served as inputs for various methods after normalization and smoothing, and the outputs were the predicted finger velocities along the x and y axes. Experiments conducted under different scenarios facilitated comparisons of predictive accuracy, inference speed, and scalability among the four types of backbone models.
• Figure 2: Scaling parameter counts for the models range from 300k to 3.8M with error