Biologically Plausible Online Hebbian Meta‐Learning: Two‐Timescale Local Rules for Spiking Neural Brain Interfaces

Sriram V.C. Nallani, Gautham Ramachandran, Sahil Shah

Keywords: spiking neural networks, online learning, eligibility traces, meta-plasticity, surrogate gradients, biologically inspired learning, recurrent SNN, reward-modulated learning, memory efficiency, Hebbian plasticity, brain-computer interface, dual-timescale learning

TL;DR: We propose an online spiking neural network with dual-timescale plasticity and meta-adaptive learning that achieves stable, memory-efficient sequence learning without backpropagation through time.

Abstract: Brain-Computer Interfaces face neural signal instability and tight memory budgets in real-time implantable settings. We introduce an online intracortical SNN decoder trained with a temporally local, layer-local but spatially non-local three-factor rule with dual-timescale eligibility traces, avoiding backpropagation through time and requiring memory that is constant in sequence length. On two primate datasets, this Online SNN attains Pearson correlations of at least $R \geq 0.63$ on Zenodo Indy and $R \geq 0.81$ on MC Maze while converging faster in early training than a BPTT-trained SNN and reducing measured training memory by 63-86\% at sequence length $T=1000$. The learning rule combines synapse-specific dual-timescale Hebbian accumulators, error-modulated updates, and integer-friendly RMS homeostasis, and operates without unrolled computational graphs or adaptive optimizer state. Closed-loop simulations with synthetic neural populations demonstrate online supervised adaptation to neural disruptions and learning from scratch without offline calibration. Overall, the method provides a memory-efficient, continuously adaptive decoder that is temporally local and Hebbian but still relies on spatial backpropagation across layers, yielding a partially biologically plausible algorithm that is competitive with BPTT-trained SNN and Kalman baselines while trading some offline accuracy relative to LSTM/GRU decoders for online learning and deployment-oriented properties.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 8103