Research map / Label-efficient and robust learning / Recognition and long-tail
The Case for Cleaner Biosignals: High-fidelity Neural Compressor Enables Transfer from Cleaner iEEG to Noisier EEG
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 BrainCodec, an RVQ-based neural compressor for EEG and iEEG, and shows that training on cleaner iEEG transfers to noisier EEG at high compression ratios.
Problem
Lossy neural compression of EEG and iEEG for storage and transmission, where existing compressors either lack high fidelity or have not been validated on these signals, and the effect of signal quality (SNR) on transfer between modalities is unclear.
Approach
Adapts a SoundStream/EnCodec-style quantized autoencoder with residual vector quantization, a multi-scale STFT discriminator and a new relative line-length loss to EEG/iEEG; finds that iEEG-trained models generalize better to EEG at high compression than EEG-trained models.
Key results
- BrainCodec maintains PRD below 30 at 64x compression on iEEG, with the seizure detection F1 at 0.78 vs 0.79 for the original signal.
- On BCI IV-2a motor imagery, BrainCodec Base at 16x gives 72% accuracy vs 77% for the original EEG, with PRD 29.0.
- A BrainCodec trained on iEEG (SWEC) compresses EEG better than an EEG-trained model at moderate-to-high compression ratios.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train BrainCodec on cleaner iEEG instead of EEG, then compress EEG at moderate-to-high compression ratios | BrainCodec · EEG across tested datasets | PRD | Helped |
| Train BrainCodec on EEG instead of iEEG, then compress iEEG | BrainCodec · SWEC iEEG | PRD | Hurt |
| Train on both iEEG and EEG instead of EEG only | BrainCodec · EEG reconstruction at high compression ratios | PRD | Helped |
| Train on both iEEG and EEG instead of iEEG only | BrainCodec · iEEG reconstruction | PRD | No clear effect |
| Use initial encoder kernel size 3 instead of 5 or 7 (ablation) | BrainCodec Base · SWEC iEEG, CR 64 | PRD: 15.94 / 15.91 -> 15.74 | Helped |
| Increase residual count from 1 to 4 (ablation) | BrainCodec Base · SWEC iEEG | PRD: 38.36 -> 15.74 | Helped |
| Increase residual count from 4 to 16 (ablation) | BrainCodec Base · SWEC iEEG | PRD: 15.74 -> 8.74 | Helped |
| Increase codebook size from 64 to 256 (ablation) | BrainCodec Base · SWEC iEEG | PRD: 21.09 -> 15.74 | Helped |
| Increase codebook size from 256 to 1024 (ablation) | BrainCodec Base · SWEC iEEG | PRD: 15.74 -> 12.74 | Helped |
| Add line length loss (ablation) | BrainCodec Base · SWEC iEEG, CR 64 | PRD: 15.95 -> 15.74 | Helped |
| Add line length loss (ablation) | BrainCodec GAN · iEEG, high-frequency reconstruction | reconstruction fidelity | Helped |
| Increase training subjects from 1 to 4 (ablation) | BrainCodec Base · SWEC iEEG, CR 64 | PRD: 15.56 -> 13.96 | Helped |
Surprising findings
- Training on higher-SNR iEEG yields better reconstruction of noisier EEG than training on EEG directly at high compression ratios.
- The GAN variant degrades more on the motor imagery task (57% accuracy at 64x) than the Base model despite better high-frequency reconstruction.
Limitations
- The EEG motor imagery task supports only up to about 16x compression before useful accuracy drops, well below iEEG.
- Models are trained on a single subject per compression model, and the GAN variant is hard to train stably.
Experimental setup
Datasets: SWEC iEEG Multi-center iEEG Brain Treebank iEEG CHB-MIT BONN BCI Competition IV-2a
Benchmarks: SWEC iEEG MC iEEG CHB-MIT BONN BCI IV-2a
Metrics: PRD PRD-spec RMSE SNR PSNR F1-score accuracy
Baselines: NLSPIHT AAC ANN CS lz4 gzip
Builds on: SoundStream EnCodec VQ-VAE MelGAN
Closest papers by meaning
- Visual Decoding and Reconstruction via EEG Embeddings with Guided Diffusion (NeurIPS 2024 · similarity 0.85)
- CORTEG: Foundation Models Enable Cross-Modality Representation Transfer from Scalp to Intracranial Brain Recordings (NeurIPS 2026 · similarity 0.84)
- Decoding Natural Images from EEG for Object Recognition (ICLR 2024 · similarity 0.84)
- MindAlign: Bridging EEG, Vision, and Language for Zero-Shot Visual Decoding (NeurIPS 2026 · similarity 0.82)
- EEG2Video: Towards Decoding Dynamic Visual Perception from EEG Signals (NeurIPS 2024 · similarity 0.82)
- EEGPT: Pretrained Transformer for Universal and Reliable Representation of EEG Signals (NeurIPS 2024 · similarity 0.82)
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.