| name | ferroelectric-snn-eeg |
| description | Personalized Spiking Neural Networks with Ferroelectric Synapses for EEG Signal Processing. Covers deployment of SNNs on ferroelectric memristive hardware for adaptive EEG-based motor imagery decoding, mixed-precision training with device-aware updates, and subject-specific transfer learning on neuromorphic platforms. Use when working with: ferroelectric synapses, memristive SNN deployment, EEG-based BCI personalization, neuromorphic hardware constraints, mixed-precision spiking training, or device-aware weight updates. Activation: ferroelectric SNN, memristive EEG, neuromorphic BCI, ferroelectric synapse, personalized SNN, device-aware training, EEG motor imagery, hardware-constrained SNN
|
Ferroelectric SNN for EEG Signal Processing
Based on arXiv:2601.00020 (Garg et al., May 2026).
Paper Overview
Title: Personalized Spiking Neural Networks with Ferroelectric Synapses for EEG Signal Processing
Authors: Nikhil Garg, Anxiong Song, Niklas Plessnig, Nathan Savoia, Laura Bégon-Lours
Published: Submitted Dec 2025, revised May 6, 2026 (v3)
DOI: 10.1063/5.0319912
Categories: cs.NE, cs.AI, cs.ET, cs.LG, eess.SY
Core Problem
EEG-based BCIs suffer from non-stationary neural signals varying across sessions and individuals,
limiting subject-agnostic model generalization. Programmable memristive hardware enables
post-deployment adaptation but faces challenges: limited weight resolution, device variability,
nonlinear programming dynamics, and finite device endurance.
Key Contributions
1. Ferroelectric Memristive Synapses for SNN
- Fabricated and characterized ferroelectric synapses for spiking neural network deployment
- Modeled weight update dynamics under realistic device constraints
- Demonstrated SNN deployment on ferroelectric memristive synaptic arrays
2. Mixed-Precision Training Strategy
- Gradient-based updates accumulated digitally
- Converted to discrete programming events only when threshold exceeded
- Device-aware weight updates accounting for nonlinear, state-dependent programming dynamics
- Mitigates endurance and energy constraints during learning/adaptation
3. Subject-Specific Transfer Learning
- Software-trained weights transferred to hardware
- Low-overhead on-device re-tuning of final network layers only
- Achieves classification performance comparable to software-based SNNs
- Enables personalized neuromorphic processing of neural signals
Architecture
Convolutional-Recurrent SNN
├── Convolutional layers (feature extraction from EEG)
├── Recurrent layers (temporal dynamics)
└── Ferroelectric memristive synapses (hardware deployment)
├── Mixed-precision accumulation (digital)
└── Threshold-based programming events (analog)
Device-Aware Training Algorithm
- Compute gradients digitally (high precision)
- Accumulate weight updates in digital memory
- When accumulated update exceeds threshold:
- Map to discrete programming events
- Account for nonlinear, state-dependent device dynamics
- Apply to ferroelectric synapses
- Repeat for adaptation phase
Transfer Learning Pipeline
- Train SNN in software on population data
- Transfer weights to ferroelectric hardware
- Retrain only final layers with subject-specific data
- Achieve personalized model with minimal hardware overhead
Key Insights
- Ferroelectric hardware supports robust low-overhead adaptation
- Mixed-precision strategy bridges software-hardware performance gap
- Subject-specific transfer learning outperforms subject-agnostic models
- Device-aware training mitigates endurance constraints
- Classification accuracy comparable to software SNNs despite hardware constraints
Activation Keywords
- ferroelectric SNN
- memristive EEG processing
- neuromorphic BCI
- personalized spiking networks
- device-aware training
- mixed-precision SNN
- ferroelectric synapse
- EEG motor imagery decoding
- hardware-constrained neural networks
- neuromorphic adaptation
Related Skills
- spiking-neural-network-analysis
- eeg-brain-connectivity-bci
- snn-learning-survey
- snn-performance-analysis
