An energy-efficient spiking neural network with continuous learning for self-adaptive brain-machine interface

TL;DR

The paper tackles non-stationarity in implantable brain–machine interfaces by proposing energy-efficient self-adaptive decoders based on Deep Spiking Neural Networks (DSNNs) trained with low-cost reinforcement-learning algorithms Banditron and AGREL. It uses transfer learning to enable deep DSNNs with minimal update burden, demonstrating open-loop stability and robust closed-loop performance under perturbations, while achieving substantial reductions in memory access and MAC operations during training (notably with DSNN_Banditron). The work provides a comprehensive open- and closed-loop evaluation on neural data and brain-model simulations, along with a rigorous resource-performance analysis that highlights its suitability for wireless iBMI deployments. Overall, the DSNN_Banditron and DSNN_AGREL approaches offer a promising path toward self-adaptive, low-power neural decoding in dynamic, real-world iBMI use cases.

Abstract

The number of simultaneously recorded neurons follows an exponentially increasing trend in implantable brain-machine interfaces (iBMIs). Integrating the neural decoder in the implant is an effective data compression method for future wireless iBMIs. However, the non-stationarity of the system makes the performance of the decoder unreliable. To avoid frequent retraining of the decoder and to ensure the safety and comfort of the iBMI user, continuous learning is essential for real-life applications. Since Deep Spiking Neural Networks (DSNNs) are being recognized as a promising approach for developing resource-efficient neural decoder, we propose continuous learning approaches with Reinforcement Learning (RL) algorithms adapted for DSNNs. Banditron and AGREL are chosen as the two candidate RL algorithms since they can be trained with limited computational resources, effectively addressing the non-stationary problem and fitting the energy constraints of implantable devices. To assess the effectiveness of the proposed methods, we conducted both open-loop and closed-loop experiments. The accuracy of open-loop experiments conducted with DSNN Banditron and DSNN AGREL remains stable over extended periods. Meanwhile, the time-to-target in the closed-loop experiment with perturbations, DSNN Banditron performed comparably to that of DSNN AGREL while achieving reductions of 98% in memory access usage and 99% in the requirements for multiply- and-accumulate (MAC) operations during training. Compared to previous continuous learning SNN decoders, DSNN Banditron requires 98% less computes making it a prime candidate for future wireless iBMI systems.

Figures (8)

• Figure 1: (a) Schematic illustration of an open-loop experiment where decoders are trained using previously obtained brain signals and tested. (b) Flow diagram of a closed-loop experiment that includes an environment for participants to operate, a brain model simulator (OPS), and a trainable decoder.
• Figure 2: Transforming a regression task into a classification task. (a) Continuous target values are converted to discrete labels by grouping target values into bins. (b) Reconstruction of continuous output from discrete labels.
• Figure 3: Illustration of the network architecture. Hidden neurons are represented in blue, while input and output neurons are shown in purple. The activity of neurons is indicated in the upper right circle of each neuron (a firing neuron is marked as 1, otherwise 0). (a) Forward pass of SNN. (b) Weight update pass for DSNN_Banditron. (c) Weight update pass for DSNN_AGREL.
• Figure 4: The comparison of continuous learning performance in Open-loop experiments across several days. The accuracy of DSNN_AGREL and DSNN_Banditron is the most stable across days and retains high accuracy across all days.
• Figure 5: (a) The trajectory of the closed-loop experiment right after loss of neuron with a perturbation ratio of 0.6. (b) The trajectory of the closed-loop experiment with loss of neuron after continuous learning for 50 trials has reduced time to target.