eeg-channel-adaptation-benchmark

name: eeg-channel-adaptation-benchmark description: "Systematic benchmark of channel adaptation methods for EEG foundation models. Compares Conv1d, SSI, source-space decomposition, and Riemannian re-centering across 5 FMs (5M-157M params), 5 tasks, revealing architecture-dependent optimal methods and probe-SFT asymmetry."

EEG Channel Adaptation Benchmark

Paper: Channel Adaptation for EEG Foundation Models: A Systematic Benchmark Across Architectures, Tasks, and Training Regimes
arXiv: 2604.23091 (April 2026)
Authors: Kuntal Kokate, Bruno Aristimunha, Dung Truong, Arnaud Delorme
Categories: cs.LG

Core Contribution

First systematic comparison of channel adaptation methods for EEG foundation models, addressing the challenge of heterogeneous electrode montages that prevent scaling EEG FMs across datasets.

Problem

EEG data comes from different electrode configurations (montages):

• 10-20 system (19-21 channels)
• 10-10 system (64+ channels)
• Custom clinical setups
• Consumer headsets (few channels)

Foundation models need to handle all these for pretraining and deployment.

Four Adaptation Methods Compared

1. Conv1d Projection

• Learnable 1D convolution maps input channels to model's expected channel count
• Simple, flexible
• Optimal for: BENDR architecture

2. Spherical Spline Interpolation (SSI)

• Interpolates electrode signals on spherical surface
• Biophysically motivated
• Optimal for: Neuro-GPT architecture

3. Source-Space Decomposition

• Projects sensor-space data to source space using inverse modeling
• Montages become irrelevant in source space
• Optimal for: Depression detection tasks

4. Riemannian Re-centering

• Uses Riemannian geometry of covariance matrices
• Aligns data distributions across montages
• Optimal for: Neuro-GPT architecture

Five Foundation Models Tested

Key Findings

1. Rigid vs. Flexible Models

• Rigid-montage models (BENDR, Neuro-GPT) require external adaptation
• Flexible-montage models (EEGPT, CBraMod) match or exceed rigid models natively when fine-tuned
• Flexible models benefit from external methods under frozen-encoder deployment

2. Probe-SFT Asymmetry

• External adaptation can cause severe negative transfer during fine-tuning of flexible models
• Probing (linear readout) benefits from adaptation, but SFT (full fine-tuning) may not
• Recommendation: Don't apply external adaptation before fine-tuning flexible models

3. Architecture-Dependent Optimal Method

• No single best method for all architectures
• Conv1d for BENDR, SSI/Riemannian for Neuro-GPT, source-space for depression detection

4. Compact Models Can Outperform Large Models

• 5M-parameter CBraMod outperforms models up to 31x larger on 4/5 datasets
• Consistent with independent findings that compact EEG-specific architectures can match larger models

Evaluation Protocol

• 5 pretrained EEG foundation models (5M–157M parameters)
• 5 downstream tasks
• 2 training regimes (probe vs. SFT)
• 10–15 random seeds per configuration

Recommendations

For Practitioners

1. Choose architecture first: Flexible montage models reduce adaptation overhead
2. Match method to architecture: Conv1d for BENDR, SSI/Riemannian for Neuro-GPT
3. Avoid adaptation before SFT: Don't apply external adaptation before fine-tuning flexible models
4. Consider compact models: 5M CBraMod matches 31x larger models on most tasks

For Researchers

• External adaptation methods are complementary, not competitive
• Architecture choice determines adaptation strategy
• Probe-SFT asymmetry needs theoretical explanation

Trigger Keywords

• eeg channel adaptation, montage alignment, eeg foundation model, spherical spline interpolation, riemannian recentering, source-space decomposition, EEG通道适配

Related Skills

• eeg-foundation-model-adapters
• tta-eeg-foundation-models
• laya-eeg-foundation
• reve-eeg-foundation