Research map / Architectures and efficiency / Pruning and efficient tuning
Are Large Brainwave Foundation Models Capable Yet ? Insights from Fine-Tuning
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.
Evaluates existing EEG foundation models (LaBraM, NeuroGPT) via full fine-tuning, frozen probing, and LoRA ablations on five BCI benchmarks, finding marginal gains over small CNN baselines at far higher parameter cost.
Problem
Whether current Large Brainwave Foundation Models (LBMs) offer real advantages over traditional deep learning models for EEG-based BCI tasks, and how parameter-efficient fine-tuning (LoRA) behaves on them.
Approach
Systematic benchmarking of pre-trained LaBraM and NeuroGPT against EEGNet and EEGInception under full fine-tuning, frozen-backbone probing, and LoRA with rank and layer-combination ablations, plus a dropout study within LoRA; first application of LoRA to LBMs.
Key results
- Fully fine-tuned NeuroGPT encoder reaches mean accuracy 0.745 vs 0.731-0.742 for EEGNet, EEG-Inception, and LaBraM, a margin of roughly 0.9%-1.2%.
- LoRA on LaBraM with rank 2 reaches mean accuracy 0.777 with about 67.7K trainable parameters versus 5.85M for full fine-tuning.
- Frozen-backbone probing lags traditional deep baselines by roughly 8-10% on average, showing full fine-tuning is needed.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fine-tune LaBraM instead of EEG-Inception | LaBraM · Sleep-EDF | accuracy: 0.688 (.057) -> 0.704 (.025) | Helped |
| Fine-tune NeuroGPT encoder instead of EEG-Inception | NeuroGPT encoder · Motor | accuracy: 0.590 (.087) -> 0.695 (.085) | Helped |
| Use foundation-model fine-tuning rather than EEG-Inception | Large brainwave foundation models · Memory | accuracy | Hurt |
| Freeze the pre-trained model and train only the classification head (ablation) | LaBraM and NeuroGPT · Motor | accuracy | Hurt |
| Vary LoRA rank from 1 to 16 (ablation) | LaBraM · Mean across four LoRA tasks | accuracy: 0.774; 0.777; 0.770; 0.769; 0.770 | Mixed |
| Vary LoRA rank from 1 to 16 (ablation) | NeuroGPT full model · Mean across four LoRA tasks | accuracy: 0.741; 0.743; 0.745; 0.742; 0.741 | Mixed |
| Vary LoRA rank from 1 to 16 (ablation) | NeuroGPT encoder · Mean across four LoRA tasks | accuracy: 0.744; 0.746; 0.746; 0.747; 0.745 | Mixed |
| Apply LoRA to a combination of layer types rather than one layer type (ablation) | LaBraM · Mean across four LoRA tasks | accuracy: 0.745 -> 0.780 | Helped |
| Apply LoRA to attention and convolution layers rather than attention only (ablation) | NeuroGPT full model · Mean across four LoRA tasks | accuracy: 0.716 -> 0.739 | Helped |
| Apply LoRA to attention and convolution layers rather than attention only (ablation) | NeuroGPT encoder · Mean across four LoRA tasks | accuracy: 0.724 -> 0.745 | Helped |
| Add dropout (0.5) to LoRA adapters (ablation) | LaBraM · Mean across four LoRA tasks | accuracy: +0.011; +0.011; +0.017; +0.019; +0.016 | Helped |
| Add dropout (0.5) to LoRA adapters (ablation) | LaBraM · KU ERP | accuracy: -0.002; +0.003; +0.004; +0.006; +0.006 | Mixed |
Surprising findings
- Frozen LBM backbones with only classification heads trained underperform small CNNs by about 8-10%.
- LoRA gains appear only when adapting combinations of layer types (e.g., attention plus convolution), not single layers alone.
Limitations
- Only two LBMs (LaBraM and NeuroGPT) evaluated, so conclusions may not generalize to other foundation models.
- Small margins between models; differences often not statistically significant per the paired t-tests.
- Hardware and compute details are not reported.
Experimental setup
Datasets: High Gamma Dataset (motor) KU ERP dataset Working Memory dataset Sleep-EDF PhysioNet Motor eyes open/closed TUH EEG Corpus (NeuroGPT pre-training, not used by authors)
Benchmarks: High Gamma motor KU ERP Working Memory Sleep-EDF PhysioNet eyes open/closed
Metrics: classification accuracy trainable parameter count paired t-test p-values
Baselines: EEGNet EEG-Inception LaBraM NeuroGPT
Builds on: LaBraM NeuroGPT LoRA EEGNet EEGConformer
Closest papers by meaning
- NeuroAtlas: Benchmarking Foundation Models for Clinical EEG and Brain-Computer Interfaces (NeurIPS 2026 · similarity 0.86)
- NeurIPT: Foundation Model for Neural Interfaces (NeurIPS 2025 · similarity 0.86)
- Standing on the Shoulders of Giants: Rethinking EEG Foundation Model Pretraining via Multi-Teacher Distillation (NeurIPS 2026 · similarity 0.86)
- What do EEG Foundation Models Capture from Human Brain Signals? (NeurIPS 2026 · similarity 0.85)
- CBraMod: A Criss-Cross Brain Foundation Model for EEG Decoding (ICLR 2025 · similarity 0.85)
- NeuroLM: A Universal Multi-task Foundation Model for Bridging the Gap between Language and EEG Signals (ICLR 2025 · 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.