Personalized Spiking Neural Networks with Ferroelectric Synapses for EEG Signal Processing

Electroencephalography (EEG)-based brain-computer interfaces (BCIs) suffer from non-stationary neural signals that vary across sessions and individuals, making subject-agnostic models impractical. To enable post-deployment adaptation on resource-constrained platforms, researchers turned to programmable memristive hardware. In a new paper (arXiv:2601.00020), Nikhil Garg and colleagues fabricate and characterize ferroelectric synapses—devices that emulate synaptic weight updates using polarization switching. They deploy convolutional-recurrent spiking neural networks (SNNs) on these synapses and evaluate two adaptation strategies: a mixed-precision scheme that accumulates gradient updates digitally and triggers discrete programming events only when a threshold is exceeded, and a transfer learning approach that retrains only the final layers on-device. Both strategies account for the nonlinear, state-dependent programming dynamics of ferroelectric devices, mitigating endurance and energy constraints.

The results show that subject-specific transfer learning achieves classification accuracy comparable to software-based SNNs, despite the limited weight resolution and variability of physical synapses. The mixed-precision scheme further reduces programming events, extending device lifespan. This work demonstrates that ferroelectric neuromorphic hardware can support robust, low-overhead personalization for neural signal processing. By decoupling training from deployment and enabling on-chip fine-tuning, the approach opens a practical path toward adaptive, energy-efficient BCIs that learn individual brain patterns without cloud dependency.