Research map / Continual learning / Online and replay-based
SPICED: A Synaptic Homeostasis-Inspired Framework for Unsupervised Continual EEG Decoding
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SPICED, a synaptic-homeostasis-inspired synaptic network that performs unsupervised continual EEG decoding across new individuals via critical memory reactivation, synaptic consolidation, and renormalization.
Problem
Unsupervised continual EEG decoding where new individuals with inter-individual variability arrive sequentially, requiring adaptation (plasticity) without catastrophic forgetting (stability).
Approach
A synaptic network of individual-specific nodes linked by feature similarity; critical nodes are reactivated by an importance score, their models fused for initialization, their connection strengths consolidated (capped at 3), and all strengths globally downscaled with Ebbinghaus-style time-dependent decay.
Key results
- On ISRUC at 30% source proportion, SPICED reaches 75.6% ACC vs 74.8% for BrainUICL (the strongest baseline).
- On FACED at 50% source proportion, SPICED achieves 48.9% ACC vs 20.2% for BrainUICL in the comparison table.
- On Physionet-MI at 10% source proportion, SPICED achieves 43.4% ACC vs 43.1% for BrainUICL, a near-tie.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SPICED instead of the strongest listed baseline | Unsupervised continual EEG decoding · ISRUC, 10%/30%/50% source | ACC: 60.6±0.71 -> 62.6±1.35; 74.8±0.11 -> 75.6±0.13; 73.9±0.29 -> 74.5±0.39 | Helped |
| use SPICED instead of the strongest listed baseline | Unsupervised continual EEG decoding · FACED, 10%/30%/50% source | ACC: 25.0±0.43 -> 25.7±0.10; 41.4±1.50 -> 43.7±0.51; 48.7±0.66 -> 48.9±0.42 | Helped |
| use SPICED instead of the strongest listed baseline | Unsupervised continual EEG decoding · Physionet-MI, 10%/30%/50% source | ACC: 43.1±0.33 -> 43.4±0.44; 48.3±0.38 -> 48.7±0.17; 52.5±0.36 -> 52.4±0.11 | Mixed |
| remove synaptic renormalization (only consolidation remains) (ablation) | SPICED · three EEG datasets, 30% source | performance | Hurt |
| remove synaptic consolidation (only renormalization remains) (ablation) | SPICED · three EEG datasets, 30% source | performance | Hurt |
| vary connection threshold ξ (ablation) | SPICED · ISRUC, 30% source | ACC: 75.5 (ξ=0.1); 74.9 (ξ=0.2); 75.0 (ξ=0.3); 74.1 (ξ=0.4); 73.4 (ξ=0.5); 73.9 (ξ=0.6) | Mixed |
| vary connection threshold ξ (ablation) | SPICED · FACED, 30% source | ACC: 43.7 (ξ=0.1); 43.9 (ξ=0.2); 43.6 (ξ=0.3); 44.2 (ξ=0.4); 43.7 (ξ=0.5); 43.5 (ξ=0.6) | Mixed |
| vary connection threshold ξ (ablation) | SPICED · Physionet-MI, 30% source | ACC: 48.7 (ξ=0.1); 48.8 (ξ=0.2); 48.6 (ξ=0.3); 48.6 (ξ=0.4); 48.9 (ξ=0.5); 48.7 (ξ=0.6) | Mixed |
| increase Gaussian noise up to 10% of signal standard deviation (ablation) | SPICED · ISRUC, 30% source | ACC: 75.6±0.13 -> 73.8±0.26 | Hurt |
Surprising findings
- On Physionet-MI at 50% source proportion, BrainUICL (52.5%) slightly outperforms SPICED (52.4%).
Limitations
- The framework only approximates synaptic homeostasis and does not model deeper neurobiological mechanisms.
- A unified handcrafted feature extraction is used across tasks without accounting for task-specific variability.
- Connection thresholds are dataset-specific and must be tuned.
Experimental setup
Datasets: ISRUC FACED Physionet-MI
Benchmarks: ISRUC FACED Physionet-MI
Metrics: accuracy macro-F1
Baselines: MMD EWC UCL-GV ReSNT CoUDA BrainUICL
Builds on: BrainUICL Contrastive Predictive Coding Elastic Weight Consolidation
Closest papers by meaning
- Learning the Plasticity: Plasticity-Driven Learning Framework in Spiking Neural Networks (NeurIPS 2025 · similarity 0.88)
- A Scalable, Causal, and Energy Efficient Framework for Neural Decoding with Spiking Neural Networks (NeurIPS 2025 · similarity 0.85)
- Discovering plasticity rules that organize and maintain neural circuits (NeurIPS 2024 · similarity 0.85)
- The Expressive Leaky Memory Neuron: an Efficient and Expressive Phenomenological Neuron Model Can Solve Long-Horizon Tasks. (ICLR 2024 · similarity 0.85)
- From Cortical Synchronous Rhythm to Brain Inspired Learning Mechanism: An Oscillatory Spiking Neural Network with Time-Delayed Coordination (NeurIPS 2026 · similarity 0.85)
- Hebbian Learning based Orthogonal Projection for Continual Learning of Spiking Neural Networks (ICLR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.