Saved in:
Bibliographic Details
Main Authors: Nallani, Sriram V. C., Shah, Sahil
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2509.14447
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914460015263744
author Nallani, Sriram V. C.
Shah, Sahil
author_facet Nallani, Sriram V. C.
Shah, Sahil
contents Intracortical brain-machine interfaces require decoders that adapt continuously to neural signal instability while operating within strict memory budgets. We introduce a dual-timescale Hebbian accumulator learning rule for spiking neural networks that enables per-timestep online supervised updates with training memory constant in sequence length, avoiding backpropagation through time. The rule combines synapse-specific fast and slow eligibility traces, error-modulated three-factor updates, and integer-friendly RMS homeostasis, operating without adaptive gradient optimizers (Adam, RMSProp) or replay buffers. On two primate intracortical datasets, the method achieves Pearson correlations of $R \geq 0.81$ on MC~Maze and $R \geq 0.63$ on Zenodo~Indy, with 63--86\% measured memory reduction versus BPTT at sequence length $T = 1000$. Closed-loop simulations demonstrate online adaptation to neural disruptions and learning from scratch without offline calibration.
format Preprint
id arxiv_https___arxiv_org_abs_2509_14447
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Dual-Timescale Hebbian Accumulators for Online Spiking Neural Network Decoding in Intracortical Brain Machine Interfaces
Nallani, Sriram V. C.
Shah, Sahil
Signal Processing
Intracortical brain-machine interfaces require decoders that adapt continuously to neural signal instability while operating within strict memory budgets. We introduce a dual-timescale Hebbian accumulator learning rule for spiking neural networks that enables per-timestep online supervised updates with training memory constant in sequence length, avoiding backpropagation through time. The rule combines synapse-specific fast and slow eligibility traces, error-modulated three-factor updates, and integer-friendly RMS homeostasis, operating without adaptive gradient optimizers (Adam, RMSProp) or replay buffers. On two primate intracortical datasets, the method achieves Pearson correlations of $R \geq 0.81$ on MC~Maze and $R \geq 0.63$ on Zenodo~Indy, with 63--86\% measured memory reduction versus BPTT at sequence length $T = 1000$. Closed-loop simulations demonstrate online adaptation to neural disruptions and learning from scratch without offline calibration.
title Dual-Timescale Hebbian Accumulators for Online Spiking Neural Network Decoding in Intracortical Brain Machine Interfaces
topic Signal Processing
url https://arxiv.org/abs/2509.14447