Personalized Spiking Neural Networks with Ferroelectric Synapses for EEG Signal Processing

TL;DR

This work tackles non-stationary EEG decoding for brain-computer interfaces by deploying spiking neural networks on ferroelectric memristive synapses. It introduces a device-aware learning framework with a Beta-shaped weight-update model and accumulation-based updates, enabling on-device adaptation under realistic programming constraints, plus a subject-specific transfer strategy that fine-tunes only the final layers. The results show that hardware-constrained SNNs can achieve software-like accuracy, remain robust to weight quantization and programming variability, and benefit from low-overhead on-device re-tuning to recover performance. The findings offer a practical path to personalized, energy-efficient neuromorphic EEG processing at the edge, with broader implications for adaptive biosignal systems.

Abstract

Electroencephalography (EEG)-based brain-computer interfaces (BCIs) are strongly affected by non-stationary neural signals that vary across sessions and individuals, limiting the generalization of subject-agnostic models and motivating adaptive and personalized learning on resource-constrained platforms. Programmable memristive hardware offers a promising substrate for such post-deployment adaptation; however, practical realization is challenged by limited weight resolution, device variability, nonlinear programming dynamics, and finite device endurance. In this work, we show that spiking neural networks (SNNs) can be deployed on ferroelectric memristive synaptic devices for adaptive EEG-based motor imagery decoding under realistic device constraints. We fabricate, characterize, and model ferroelectric synapses. We evaluate a convolutional-recurrent SNN architecture under two complementary deployment strategies: (i) device-aware training using a ferroelectric synapse model, and (ii) transfer of software-trained weights followed by low-overhead on-device re-tuning. To enable efficient adaptation, we introduce a device-aware weight-update strategy in which gradient-based updates are accumulated digitally and converted into discrete programming events only when a threshold is exceeded, emulating nonlinear, state-dependent programming dynamics while reducing programming frequency. Both deployment strategies achieve classification performance comparable to state-of-the-art software-based SNNs. Furthermore, subject-specific transfer learning achieved by retraining only the final network layers improves classification accuracy. These results demonstrate that programmable ferroelectric hardware can support robust, low-overhead adaptation in spiking neural networks, opening a practical path toward personalized neuromorphic processing of neural signals.

Figures (9)

• Figure 1: Spiking neural network architecture. The time series signals from all the 64 electrodes were used where the participants imagine limb movements. The electrode potential at each time step is converted to a 2D map, which is passed to the network from ref. kumar2022decoding comprising four convolutional layers followed by a recurrent and two fully connected layers.
• Figure 2: Ferroelectric synaptic device programming. a Schematic of the ferroelectric synaptic device stack and the programming scheme, consisting of write pulses with fixed pulse width (50 µ s) and increasing amplitudes ($V_{\mathrm{prog}}$). Positive-polarity pulses induce long-term depression (LTD), followed by negative-polarity pulses that induce long-term potentiation (LTP). The device conductance is read after each programming pulse. b Characterization results showing the applied programming-pulse amplitudes (top) and the corresponding evolution of device conductance over time (bottom) for ten programming cycles. Programmed conductance values grouped by pulse amplitude and fitted with Gaussian distributions for LTD (c) and LTP (d) pulse sequences. e Standard deviation of the fitted conductance distributions plotted as a function of the programmed conductance level.
• Figure 3: Model a The change in normalized weight ($\Delta W$) is plotted with respect to the initial weight for long-term potentiation (LTP) and depression (LTD) including the device characterization data and the fitted model. b The evolution of weight with programming pulse number is plotted for the device data and model.
• Figure 4: Simulation framework: The measurements from device characterization was used to fit the model. Thereafter the model is used to compute the weight updates of the network. The weight updates computed from the gradient descent is accumulated until a predefined threshold is surpassed, after which the weight update is computed through the fitted model and applied to the respective weights.
• Figure 5: Software baseline: Network's classification performance for 2-class problems: Left/Right hand classification on the validation set plotted against the number of epochs. The error bars depicts the standard deviation across five folds.