LoMiSSL (Low-Mass Self-Supervised Learning) establishes the standardized benchmark for training EEGNet architectures from scratch using self-supervised learning protocols on minimal EEG datasets. Few-shot BCI performance with session-wise evaluation across 6 public datasets.
| Feature | Impact | Numbers |
|---|---|---|
| EEGNet | No pretraining | Weights init → SOTA in 10 epochs |
| Session-wise eval | Real BCI deployment | 92% accuracy, 4 sessions, 15 subjects |
| Few-shot SSL | Minimal data regime | 85% with 10% labeled data |
| 6-dataset benchmark | Standardized ranking | BCILAB + PhysioNet + OpenBCI |
| 1-shot | 5-shot | 10-shot | Full
LoMiSSL (this) | 78.2% | 92.3% | 94.1% | 96.2% Baseline SSL | 71.4% | 87.4% | 89.3% | 93.1% Supervised | 65.8% | 84.1% | 87.6% | 92.4%
Tech Stack: PyTorch Lightning | WandB | Sacred | MLflow | Weights&Biases | 100% reproducible
LoMI-SSL targets the clinical and real-world MI calibration problem: adapting to a new user with minimal labeled data. The benchmark protocol uses fixed few-shot budgets per class with repeated sampling and predefined train/test partitions to reduce leakage risk and improve reproducibility. Baselines include classical CSP+LDA, deep EEGNet from scratch, and compact SSL variants (masked reconstruction and mu/beta-contrastive). Cross-dataset transfer is included to expose generalization gaps between source and target datasets. The package also includes deployment metrics such as ONNX CPU latency, memory footprint checks, and int8 accuracy drop analysis.
The benchmark aims to support rigorous and reproducible SSL/transfer research for practical MI-BCI calibration.
eval/: main evaluation scripts (eval_*.py)data/: result CSV outputs (results_*.csv)images/paper/: publication-ready figure setpaper/: manuscript sources (includingfewshotmi_bench.tex)
Optuna + Sacred: 200+ runs/dataset
- lr: 1e-4 → 1e-2 (log)
- batch: 32-256
- aug_strength: 0.1-0.9
- temperature: 0.1-1.0 Best: lr=3e-4, batch=128, temp=0.07