MPA method adapts spiking neural nets for brain-computer interfaces with 4ms resolution

Brain-computer interfaces (BCIs) that decode intracortical signals face a persistent challenge: day-to-day neural signal shifts degrade the performance of pretrained decoders. Existing unsupervised adaptation methods rely on deep recurrent or adversarial networks, which are too computationally expensive for implantable hardware. In a new paper accepted at ICANN 2026, researcher Guangzhi Tang introduces Membrane Potential Alignment (MPA), a test-time adaptation method tailored for spiking neural networks (SNNs). MPA realigns a pretrained decoder to shifted neural recordings by simply matching membrane potential distributions using KL divergence, without requiring any labeled data. By restricting updates to low-rank (LoRA) weights, MPA adapts fewer than 9% of the network's parameters, making it highly efficient for on-device deployment.

Tested on a non-human primate reaching task spanning over a month, MPA achieves performance comparable to the state-of-the-art NoMAD method, while using a simpler architecture and offering finer temporal resolution (4 ms vs. 20 ms). This finer resolution allows for more precise decoding of neural activity. The results demonstrate that efficient SNN-based test-time adaptation is a practical path toward long-term, recalibration-free BCIs, potentially enabling more reliable and user-friendly neural prosthetics that maintain high performance without frequent manual recalibration.