PaperlayerView exact revision on arXiv

Source-linked AI summary

EEG-FM-Compass: Progress, Benchmarking, and Future Directions for EEG Foundation Models

Dingkun Liu, Yuheng Chen, Zhu Chen, Zhenyao Cui, Yaozhi Wen, Jiayu An, Jingwei Luo, Dongrui Wu

arXiv:2601.17883v3cs.LGcs.CV

TL;DR

Heterogeneous objectives and evaluation protocols make it difficult to draw consistent conclusions about EEG foundation-model generalization, so EEG-FM-Compass reviews and benchmarks models across BCI tasks. Its results show that fine-tuning offers advantages over linear probing, specialists remain competitive, and larger models do not guarantee better generalization.

  • Problem

    Existing EEG foundation models use heterogeneous pre-training objectives and downstream evaluation protocols, complicating consistent conclusions about factors driving generalization.

    Source: §2.1 · p. 6

  • Method

    EEG-FM-Compass reviews 55 studies into a unified framework and benchmarks 12 open-source EEG foundation models against specialist baselines across 13 datasets and nine BCI paradigms.

    Source: §5 · pp. 25–26

  • Results

    Current EEG foundation models lack universally transferable representations: full-parameter fine-tuning provides substantial advantages over linear probing, specialists remain competitive, and larger models do not guarantee improved generalization.

    Source: §5 · p. 26

  • Takeaways & Limitations

    The findings point to future research on pre-training strategies and robustness to noise and cross-task heterogeneity.

    Source: §5 · p. 26

    When pre-training targets the waveform, models may expend representational capacity reconstructing idiosyncratic noise that is not predictive for downstream tasks.

    Source: Model pre-training · p. 11

Abstract

from arXiv · show

Electroencephalography (EEG) foundation models (FMs) have recently emerged as a promising paradigm for brain-computer interfaces, aiming to learn transferable neural representations from large-scale heterogeneous recordings. Despite rapid progress, a fair and comprehensive comparison of existing EEG FMs is still lacking, owing to inconsistent pre-training objectives, preprocessing choices, and downstream evaluation protocols. To fill this gap, we present EEG-FM-Compass. We first review 55 representative models and organize their design choices into a unified taxonomic framework including data standardization, model architectures, and self-supervised pre-training strategies. We then evaluate 12 open source FMs and competitive specialist baselines across 13 EEG datasets spanning nine brain-computer interface paradigms. Emphasizing real-world deployments, we consider both cross-subject generalization under a leave-one-subject-out protocol and rapid calibration under a within-subject few-shot setting. We further compare full-parameter fine-tuning with linear probing to assess the transferability of pre-trained representations, and examine the relationship between model scale and downstream performance. Our results indicate that: 1) linear probing is frequently insufficient; 2) specialist models trained from scratch remain competitive across many tasks; and 3) larger FMs do not necessarily yield better generalization performance under current data regimes and training practices.

1. INTRODUCTION · Q3. Do EEG FMs exhibit behavior consistent with scaling laws? · 2. OVERVIEW OF EEG FMs

The paper addresses inconsistent evaluation of EEG foundation models by surveying their design space and proposing systematic comparisons of transfer, scratch-trained baselines, and model scale. It frames EEG FMs as a way to learn transferable representations from heterogeneous recordings, while emphasizing unresolved questions about generalization and scaling.

1. INTRODUCTION

  • EEG BCI spans motor, perceptual, cognitive, and clinical states, but costly labels, heterogeneous devices, limited transfer, and inconsistent evaluation motivate the paper’s questions about generalization and performance versus scratch-trained methods.The introduction also asks whether EEG FMs generalize across downstream tasks and whether they consistently outperform traditional and deep-learning methods trained from scratch.Source: §1 · pp. 1–2

Q3. Do EEG FMs exhibit behavior consistent with scaling laws?

  • The paper asks whether larger models and greater pre-training data volumes improve downstream generalization, without presupposing that scaling helps.The benchmark is designed to investigate this across EEG FMs of varying parameter sizes and pre-training datasets.Source: Q3. Do EEG FMs exhibit behavior… · p. 2, §2 · p. 3
  • EEG FMs are motivated by the premise that self-supervised pre-training on heterogeneous recordings can produce representations transferable across diverse tasks with minimal task-specific adaptation.The stated motivation is to leverage unlabeled EEG recordings that would otherwise remain unused and accommodate diverse electrode configurations.Source: Q3. Do EEG FMs exhibit behavior… · p. 2, §1 · p. 2, Figure 1 · p. 3
  • The survey covers 55 EEG FMs and organizes their data scale, preprocessing, pre-training strategies, and architectures within a unified taxonomy.The paper presents this as a comprehensive collection and structured comparison of existing EEG FMs.Source: Q3. Do EEG FMs exhibit behavior… · p. 2

2. OVERVIEW OF EEG FMs

  • The proposed benchmark compares full-parameter with classification-head fine-tuning, scratch-trained approaches with EEG FMs, and models of varying parameter sizes.It includes leave-one-subject-out evaluation and within-subject few-shot adaptation using approximately 1/20–1/100 of the fine-tuning data typically used in LOSO.Source: §2 · p. 3

2.1. Advances and Trends of EEG FMs

  • 16.4% of surveyed studies were published in 2024 and 67.3% in 2025 or 2026, indicating a surge in EEG FM research alongside greater design diversity.The overview tracks model scope, modalities, architectures, training methods, and dataset use across the surveyed models.Source: §2.1 · p. 3, Figure 3 · p. 6
  • Transformer backbones dominate surveyed EEG FMs, but capacity and training resources vary substantially; model sizes range from fewer than one million to several billion parameters.The overview notes that model capacity does not follow a monotonic scaling trend, while heterogeneous objectives and evaluation protocols complicate conclusions about drivers of generalization.Source: §2.1 · pp. 5–6

2.2. Definition of EEG FMs

  • EEG FMs aim to learn transferable, generalizable representations, unlike conventional task- and dataset-specific BCI pipelines; EEG’s resource-intensive collection also constrains current training scale.The overview says current EEG FMs primarily target classification across electrode configurations and BCI paradigms, with fewer parameters and lower computational budgets than text or image models.Source: §2.2 · pp. 6–7

Data collection and curation · Data preprocessing

EEG FM pipelines draw on public datasets and self-collected recordings, often pooling diverse unlabeled data for pre-training; preprocessing addresses signal and device variability through alignment, normalization, and other standardization. The review also finds augmentation strategies sparse and heterogeneous.

Data collection and curation

  • EEG FM pre-training commonly aggregates large volumes of unlabeled data across diverse tasks, sourced from public datasets and self-collected recordings.Fig. 3 presents dataset-use frequencies for pre-training and downstream evaluation across existing EEG FMs.Source: Data collection and curation · p. 7

2.3. Problem Definition

  • Screening noisy subjects and channels can improve dataset quality, stabilize optimization, and support better learned representations.Common curation excludes participants with abnormally low task performance or excessive noise, and removes channels with persistent artifacts or disconnections.Source: §2.3 · p. 7

Data preprocessing

  • Subject, device, and recording differences can shift EEG distributions, so preprocessing and normalization aim to reduce nuisance variability and stabilize optimization.Electrode placement, impedance, environmental noise, and physiological state are cited sources of variability.Source: Data preprocessing · p. 7
  • Channel unification ranges from fixed montages and template remapping to channel-wise processing, spatial embeddings, coordinate-aware encoding, and unified latent spaces.Fixed montages simplify alignment but discard unmatched channels and constrain transfer; templates normalize spatial layouts but remain tied to a chosen template or region definition. Channel-wise modeling accommodates variable channel counts but may need mechanisms to recover spatial interactions; fixed channel embeddings can limit generalization to unseen layouts, whereas coordinate-aware and latent-space approaches support heterogeneous configurations.Source: Data preprocessing · p. 8
  • 50.0% of surveyed studies resample signals to 200 Hz, while 34.1% resample to 250 Hz or 256 Hz.Bandpass filters commonly attenuate slow drifts and high-frequency noise, while 50 Hz or 60 Hz notch filters reduce power-line interference.Source: Data preprocessing · pp. 8–9
  • Common preprocessing includes z-score normalization and CAR, while Euclidean Alignment explicitly reduces covariance shifts and EMA normalization handles gradual drift in long recordings.Z-score rescales each channel to zero mean and unit variance; CAR subtracts the across-channel average, and EMA can adapt to non-stationarity in online or streaming settings.Source: Data preprocessing · p. 9
  • Only a few surveyed EEG FMs explicitly use augmentation, and their approaches are sparse and heterogeneous.Examples include contrastive perturbed views, overlapping temporal windows, Gaussian noise, channel dropout, temporal shifts, and noise added to electrode coordinates; physiologically grounded policies are proposed as a future direction.Source: Data preprocessing · p. 10

Model pre-training

EEG foundation models predominantly use masked prediction, but differ in target representation and inductive bias. The review contrasts waveform, token, frequency, codebook, and autoregressive objectives, emphasizing that their suitability depends on data quality and deployment needs.

Pre-training trends

  • About 83.6% of reviewed models use Transformer backbones, while random masking accounts for 71.7% of masking strategies and raw-signal reconstruction for 25.5% of targets.Masked reconstruction is the dominant pre-training paradigm; token reconstruction and hybrid raw-and-token targets together make up a comparable fraction to raw-signal reconstruction.Source: Model pre-training · pp. 10, 14

Token and latent-space objectives

  • Token-level reconstruction and contrastive learning can reduce waveform-noise sensitivity and stabilize optimization, but tokenizer quality matters and auxiliary objectives can add global semantic signal.A tokenizer that is too shallow can leave the target close to waveform regression, while one that is too coarse can discard transient or task-relevant patterns.Source: Model pre-training · pp. 11–12

Masked Reconstruction of Raw Signals

  • Raw-signal reconstruction preserves waveform and cross-channel structure and is a strong baseline with controlled data and challenging masks, but can memorize non-predictive EEG noise.The noise risk motivates robust losses and denoising targets that suppress artifacts rather than reproduce them.Source: Model pre-training · pp. 10–11

Frequency-Domain Reconstruction

  • Frequency-domain supervision emphasizes rhythmic EEG structure, while time-domain or auxiliary losses are often retained to preserve waveform fidelity and task-relevant information.The review describes frequency targets used alone or alongside raw-signal, codebook, or auxiliary objectives.Source: Model pre-training · p. 12

Codebook-Based Objectives

  • Codebook objectives provide compact symbolic EEG representations that can reduce low-level noise and support sequence modeling, but poorly controlled codebook design can cause collapse or imbalanced usage.Design choices include codebook size, commitment regularization, residual quantization depth, and update schedules.Source: Model pre-training · pp. 12–13

Hybrid Objectives and Practical Selection

  • Objective selection should match deployment: token or codebook targets can aid robustness, frequency supervision can suit rhythmic tasks, and autoregression supports generation but depends on careful tokenization and ordering.Causal objectives can be harder to optimize than bidirectional masked reconstruction, and hybrid losses are intended to constrain representations from complementary perspectives.Source: Model pre-training · p. 13

Downstream generalization

EEG foundation models are expected to support adaptation with minimal calibration, motivating evaluation across both data-rich cross-subject transfer and data-limited calibration. Existing downstream evaluations cover several paradigms but concentrate largely on clinical EEG, motor imagery, affective decoding, and sleep analysis.

Downstream generalization

  • A comprehensive evaluation suite should test both data-rich cross-subject transfer and data-limited calibration under consistent protocols.This addresses the expectation that foundation models reduce reliance on task-matched labeled data and enable adaptation with minimal calibration.Source: Downstream generalization · p. 14
  • Downstream evaluation assesses whether representations learned from large-scale EEG corpora transfer effectively to practical BCI tasks.The paper’s benchmark design is constructed to address both cross-subject transfer and data-limited calibration requirements.Source: Downstream generalization · p. 14
  • Frequently used downstream datasets span clinical EEG, motor imagery, emotion recognition, and sleep, with existing evaluation concentrated in these areas.Examples include TUAB, TUEV, and CHB-MIT for clinical EEG; BCIC-IV-2A and PhysioNetMI for motor imagery; FACED, SEED, and SEED-V for emotion; and Sleep-EDF for sleep.Source: Downstream generalization · p. 14

3. BENCHMARK OF EEG FMs

The benchmark evaluates open-source EEG foundation models and specialist baselines across diverse BCI datasets, using cross-subject and within-subject protocols to assess generalization and transferability. Its results include leakage-aware model rankings and qualitative comparisons of learned representations.

3.1. Datasets

  • The benchmark spans 13 datasets and 9 BCI paradigms, broadening evaluation beyond commonly studied clinical EEG, motor imagery, emotion recognition, and sleep staging.For widely used paradigms such as MI, P300, and clinical diagnosis, the benchmark included two to three datasets.Source: §3.1 · p. 14, §3 · p. 14

3.2. Evaluation Scenarios

  • LOSO tests cross-subject transfer without test-subject calibration, while within-subject few-shot tests personalization from limited target-user data; both are compared using full-parameter fine-tuning and linear probing.Few-shot calibration can reduce adaptation cost and avoid requiring a matching external training corpus, but requires labeled target-user data; LOSO requires labeled data from multiple subjects and a matching device-task corpus.Source: §3.2 · pp. 15–16

3.3. Evaluated Approaches

  • The benchmark compares 12 open-source EEG FMs, including 11 general-purpose models and MIRepNet for motor imagery, with specialist baselines across the evaluation suite.The broader benchmark description reports 8 specialist baselines, including conventional machine-learning and deep-learning methods.Source: §3 · p. 14, §3.3 · p. 16

3.4. Main Results

  • Masked reconstruction objectives may help models capture EEG temporal dynamics when target stimuli flicker at distinct frequencies and produce periodic signals.The passage presents this as a possible explanation for model capability.Source: §3.4 · p. 17
  • Leakage-aware rankings retained CBraMod and Neuro-GPT among the top FMs, suggesting pretraining–downstream overlap did not primarily drive the reported findings.The analysis excluded downstream datasets used in each FM’s pretraining corpus before computing average rank.Source: §3.4 · p. 17
  • Full-parameter fine-tuning generally produced clearer between-class separation and more compact within-class clusters than linear probing in the visualized embeddings.The visualizations covered SEED and TUAB for selected specialist models and FMs; the authors interpreted them as evidence that frozen FM representations are not always sufficiently separable for downstream decoding.Source: §3.4 · p. 17

Can FMs learn generalized representations?

Pre-trained EEG FMs show some transferability, but their representations are not robustly universal: most benefit from full-parameter task adaptation, and performance varies across tasks.

Can FMs learn generalized representations?

  • Linear probing was inferior to full-parameter fine-tuning for most FMs, indicating that downstream decoding generally required task-specific encoder adaptation, while neither strategy was consistently strong.Even the best-performing FMs degraded on specific tasks, leaving truly universal EEG representations an open challenge.Source: Can FMs learn generalized… · pp. 17–18
  • 9 out of 12 FMs had significant positive median Δ values favoring full fine-tuning; BENDR and CBraMod favored it in 23/25 scenarios, and TFM in 24 with one tie.EEGPT was the only model with a significant linear-probing advantage; BrainOmni-Tiny and SingLEM showed no significant difference. The results indicate that current FMs have not learned sufficiently universal representations for robust decoding.Source: Can FMs learn generalized… · p. 17
  • FM performance varied by task: CBraMod placed first and second on SEED in LOSO and few-shot settings but had the highest SEED-VIG LOSO RMSE; LUNA led on TUAB but generalized poorly beyond clinical tasks.LUNA's limitation was likely related to pre-training exclusively on TUEG and Siena.Source: Can FMs learn generalized… · p. 17
  • On Nakanishi2015, several FMs, including BENDR, EEGPT, and Neuro-GPT, performed strongly despite only 12 fine-tuning trials per class.Nakanishi2015 is an SSVEP dataset.Source: Can FMs learn generalized… · p. 17
  • The evaluation also compared existing FMs with eight specialist models trained from scratch, comprising one traditional machine-learning method, four CNN-based methods, and three Transformer-based methods.The supplied passage does not report the outcome of this comparison.Source: Can FMs learn generalized… · p. 19

Can FMs consistently outperform specialist models?

Across the benchmark, specialist models achieved higher average decoding accuracy overall and were especially strong in few-shot evaluation, while performing comparably to FMs in LOSO. Current EEG FMs did not significantly outperform top specialist models, and model size showed no clear relationship with downstream performance among top FMs.

Can FMs consistently outperform specialist models?

  • EEGNet led top-1 counts despite 2K parameters, ShallowConv led top-3 counts, and specialists achieved 16 first-place finishes and 47 top-3 placements versus selected FMs.For the group comparison, the FM set comprised the seven FMs with the highest top-3 counts.Source: Can FMs consistently outperform… · p. 20
  • EEGPT required approximately 361 GFLOPs per forward pass versus EEGNet’s 0.35 GFLOPs, with EEGPT fine-tuning throughput about 46 times lower in the 4s, 10-channel setting.The authors note that this overhead may limit deployment in resource-constrained or real-time BCI applications.Source: Can FMs consistently outperform… · p. 20
  • Specialists achieved higher average decoding accuracy overall, led few-shot top-1 placements, and performed comparably to FMs in LOSO; top specialists were not significantly outperformed by the best FM.These scenarios represent labeled multi-subject evaluation and rapid personalization with minimal calibration, respectively.Source: Can FMs consistently outperform… · pp. 20–21

3.5. Comparison of Different Fine-Tuning Ratios

  • Across fine-tuning data ratios from 10% to 90% in 20% increments, most selected specialist models and FMs improved as data increased, leaving minimal-calibration adaptation an open challenge.The comparison included EEGNet, ShallowConv, and LMDA, and CBraMod, Neuro-GPT, and LaBraM.Source: §3.5 · p. 21
  • 6.56 was CBraMod’s average rank with 4.0M parameters, while 2K-parameter EEGNet ranked second; the top six FMs showed no clear model-size and downstream-performance correlation.The rankings covered 13 datasets under both LOSO and few-shot scenarios.Source: §3.5 · p. 21

4. DISCUSSION

The discussion finds that neither EEG FM scale nor pre-training objective reliably predicts downstream generalization. It highlights paradigm-specific pretraining and Euclidean Alignment (EA) as practical avenues for adaptation, while noting open questions about auxiliary data.

4. DISCUSSION

  • EEG FM parameter growth does not guarantee better downstream generalization under current data and training regimes.The discussion attributes this partly to limited, noisy EEG data and a shortage of large, high-quality corpora [86], and partly to objectives that may not convert added capacity into richer representations.Source: §4 · p. 21

4.1. Pre-training Objectives vs. Downstream Performance

  • Pre-training objective categories show no consistent advantage for particular BCI paradigms, including SSVEP and clinical tasks.On Nakanishi2015, BENDR, EEGPT, and Neuro-GPT perform strongest despite using time-domain or token-level objectives; on TUAB and CHB-MIT, both raw-signal and embedded-token objectives are competitive.Source: §4.1 · p. 22
  • Objective choice alone is not decisive: raw-signal, embedded-token, and codebook models span ranks 1–15, 7–18, and 9–20, respectively.The top-ranked models in the three categories—CBraMod, Neuro-GPT, and LaBraM—are each competitive overall; codebook-model average ranks (9.36, 12.33, 12.76, and 16.20) do not improve monotonically with size. The discussion points instead to corpus scale, data curation, masking strategy, and architectural inductive bias as influential factors.Source: §4.1 · pp. 21–22

4.2. Paradigm-Specific FMs

  • When the target paradigm is known before data collection, paradigm-specific FMs are a viable adaptation strategy; whether other-paradigm data helps pretraining remains open.In MI comparisons, MIRepNet achieved state-of-the-art subject-averaged accuracy and Cohen’s Kappa across three datasets.Source: §4.2 · p. 23, §4.3 · p. 23

4.3. Effectiveness of EA

  • On BNCI2014001, adding EA during training or fine-tuning improved generalization for most models.EA aligns subjects’ average trial covariance to the identity, reducing inter-subject differences in second-order statistics.Source: §4.3 · p. 23, Figure 7 · p. 24

Paradigm-specific FMs … 8. AUTHOR CONTRIBUTIONS

EEG-FM-Compass reviews EEG foundation-model designs and benchmarks open-source models across BCI paradigms and evaluation settings. Its findings show limited universal transfer and motivate stronger pre-training strategies and higher-quality EEG data.

Paradigm-specific FMs

  • Paradigm-specific FMs—including MEET [36], MIRepNet [61], PSGFM [62], and EpilepsyFM [65]—are presented as practical; MIRepNet's MI advantage may reflect paradigm-specific pre-training and neurophysiological principles.These factors may help MIRepNet learn more effective representations for MI tasks.Source: Paradigm-specific FMs · p. 25
  • EA was used only for models whose original implementations adopted it to avoid alignment mismatch, but its whitening transformation may distort spatial structure and reduce interpretability.The trade-off matters for spatially aware models, including those using electrode-coordinate encoding or graph-based representations.Source: Paradigm-specific FMs · p. 25

Efficient pre-training strategies

  • Future work should address cross-device heterogeneity, learn more universal transferable representations, and enable efficient adaptation with less calibration data and faster distribution alignment.The paper reports a preliminary parameter-efficient fine-tuning experiment in the supplementary information.Source: Efficient pre-training strategies · p. 25

Large-scale high-quality data construction

  • Low EEG signal-to-noise ratios and raw-signal reconstruction may encourage fitting noise; current FMs lack scaling-law behavior, while cleaner, smaller MI data yielded better performance than uncleaned data.The authors call for systematic collection and rigorous cleaning of large, high-quality EEG corpora across paradigms.Source: Large-scale high-quality data… · p. 25

5. CONCLUSION

  • The benchmark evaluates 12 open-source EEG FMs alongside competitive specialist baselines across 13 datasets spanning 9 representative BCI paradigms, using cross-subject LOSO and within-subject few-shot protocols.The dataset profiles include P300, seizure and abnormal-EEG detection, sleep staging, emotion recognition, SSVEP, and workload tasks. Table 4 summarizes preprocessing, while Table 5 overviews pre-training strategies and backbone architectures.Source: §5 · pp. 25–26, A1. SUMMARY OF EEG FMs · p. 27
  • Full-parameter fine-tuning consistently outperforms linear probing; scratch-trained specialist models remain competitive, and increasing model size alone does not guarantee better generalization.The findings indicate that pre-trained encoders cannot yet serve as fixed feature extractors across diverse downstream tasks.Source: §5 · p. 26

8. AUTHOR CONTRIBUTIONS

  • D.L. conceived, analyzed, and wrote the study; other authors contributed experiments, analysis, literature review, data collection, revisions, and supervision, and no conflicts were declared.Y.C. and Z.C. performed experiments; Z.C. and Y.W. contributed to review and data collection; J.A. and J.L. revised the manuscript; D.W. supervised and revised it.Source: §8 · p. 26

A2. PRE-TRAINING AND DOWNSTREAM DATASETS

The section inventories datasets used for EEG FM pre-training and downstream evaluation, then describes the 13 benchmark datasets and their within-subject few-shot fine-tuning protocols.

A2. PRE-TRAINING AND DOWNSTREAM DATASETS

  • Tables 6–13 systematically document the pre-training and downstream evaluation datasets used by each existing EEG FM.Together, the tables provide an overview of data resources underlying current EEG FMs.Source: A2. PRE-TRAINING AND DOWNSTREAM DATASETS · p. 27

A3. DATASET DESCRIPTIONS

  • The benchmark covers 13 datasets, with their characteristics and corresponding within-subject few-shot fine-tuning protocols summarized separately.The characteristics are summarized in Table 14 and the protocols in Table 15.Source: A3. DATASET DESCRIPTIONS · p. 27
  • The first four datasets span motor imagery tasks in BNCI2014001, BNCI2015001, and BNCI2014004, plus P300 speller data in BNCI2014009.Their descriptions include subject counts, recording setups, and task or session details.Source: A3. DATASET DESCRIPTIONS · p. 27
  • Things-EEG2 supports visual-object decoding across 1,854 object classes, while SEED-VIG supports continuous vigilance estimation as a regression task.Things-EEG2 recordings come from 10 participants viewing natural object images; SEED-VIG uses 8-second EEG epochs and PERCLOS-derived vigilance levels.Source: A3. DATASET DESCRIPTIONS · p. 30

Main Results

The supplied passages identify the scope and conditions of the benchmark results, but do not provide the table values needed to report model winners or comparative performance.

Main Results

  • Detailed benchmark results for each subject are presented in Tables 16–51; cited table examples include accuracies on BNCI2014001, classification AUCs on BNCI2014009, RMSE in SEED-VIG, and correlation coefficient in SEED-VIG.Source: Main Results · p. 30, Table 16 · p. 36, Table 22 · p. 42
  • The few-shot protocol fine-tunes each seizure subject on the first seizure’s ictal segments and its 10-minute pre-ictal segment, then evaluates on remaining seizures and their corresponding pre-ictal segments.Epochs originally 0.8 seconds long are extended to 1 second by prepending the preceding 0.2-second segment; SEED-VIG is a continuous-PERCLOS regression task, so trials per class do not apply.Source: Main Results · p. 35, Table 15 · p. 35
  • Overall rankings cover 13 datasets under LOSO and few-shot scenarios, with separate summaries excluding overlapping pre-training datasets and grouping models by pre-training objective.The ranking tables are Tables 52–54; the leakage-aware ranking excludes rankings on downstream datasets that overlap with each model’s pre-training data before computing average rank.Source: Table 52 · p. 71, Table 54 · p. 71, Table 53 · p. 71
  • Reported implementation measurements use PyTorch 2.6.0, CUDA 12.4, one NVIDIA A100-SXM4-80GB GPU, and a fixed batch size of 32.The accompanying descriptors include trainable parameters, aggregated feature dimension, and single-sample forward-pass FLOPs.Source: Main Results · p. 71

Parameter-Efficient Fine-tuning

The section reports a preliminary evaluation of LoRA-based parameter-efficient fine-tuning on two EEG foundation models across two datasets and two evaluation settings. It also points to broader model-ranking summaries across 13 datasets.

Parameter-Efficient Fine-tuning

  • LoRA was evaluated on CBraMod and Neuro-GPT using BNCI2014001 and BNCI2014009 under LOSO and within-subject few-shot settings.The results are reported in Table 57, but the supplied passage does not provide the results themselves.Source: Parameter-Efficient Fine-tuning · p. 71
  • Tables 52–54 summarize overall model rankings, FM rankings by pre-training objective, model sizes, and a leakage-aware ranking across 13 datasets.The supplied passage does not give specific ranking values.Source: Parameter-Efficient Fine-tuning · p. 71

Computational Efficiency … Comparison of different fine-tuning ratios

Across fine-tuning data ratios, most models improved as more data became available, while relative rankings remained largely stable. The paper also compares fine-tuning costs, analyzes model rankings statistically, and situates its survey and benchmark against prior work.

Statistical Significance Analysis

  • For each EEG FM, the study paired full fine-tuning and linear probing over dataset–scenario pairs, defining positive Δ as favoring full fine-tuning.For SEED-VIG, where lower RMSE is better, the sign was reversed; results summarize full wins, linear wins, ties, and rank-biserial correlation as effect size in Table 58.Source: Statistical Significance Analysis · pp. 71–72

Computational Efficiency

  • Tables 55 and 56 compare fine-tuning computational costs for evaluated models under flatten and pooling feature aggregation, reporting full-parameter and head-only fine-tuning costs.Each cell reports Full / Linear, respectively.Source: Statistical Significance Analysis · p. 71

A5. Representation Visualization

  • Figures 15–17 visualize learned representations on SEED, TUAB, and BNCI2014008 for seven specialist models and nine EEG FMs.For each FM, the visualizations show both full-parameter fine-tuning and linear probing.Source: A5. Representation Visualization · p. 72

A6. EXISTING SURVEYS AND BENCHMARKS

  • The survey covers 55 EEG FMs, and the benchmark compares 12 open-source FMs with 7 specialist baselines across 13 datasets and 9 BCI paradigms.Evaluations use LOSO and within-subject few-shot scenarios with full-parameter fine-tuning and linear probing; the authors describe this as the largest survey and most comprehensive benchmark to their knowledge.Source: A6. EXISTING SURVEYS AND BENCHMARKS · p. 72
  • The reported statistic was F = 98.19 (p = 7.56×10−14), and the Iman-Davenport corrected statistic was also significant (FF = 7.81, p = 4.34 × 10−16).The Nemenyi post-hoc test yielded a critical difference of 4.94 at α = 0.05; models connected by a horizontal bar in Fig. 8 are not significantly different at the 0.05 level.Source: A6. EXISTING SURVEYS AND BENCHMARKS · p. 72

Comparison of different fine-tuning ratios

  • Most models improved as fine-tuning data rose from 10% to 90%, while relative rankings remained largely stable across ratios.TRCA led on Nakanishi2015 at every ratio, and EEGNet remained competitive on BNCI2014008; rapid or calibration-free adaptation remains an open challenge for practical BCI deployment.Source: Comparison of different fine-tuning… · p. 72

A7. SYMBOLS AND NOTATION

  • Tables 60 and 61 summarize the paper’s principal mathematical symbols and loss functions, while model-specific auxiliary losses are listed in Table 5.The symbols and losses tables do not exhaustively enumerate model-specific auxiliary losses.Source: A7. SYMBOLS AND NOTATION · p. 72

Symbol Meaning Usage

This section defines notation for EEG inputs, preprocessing, patching, pre-training objectives, and downstream evaluation. It also distinguishes common loss symbols from model-specific losses, which are summarized in Table 5.

Inputs and patching

  • X denotes a C×T raw EEG trial, while G(·), N_p, S(·), and P define preprocessing and temporal patch construction.P is the patch tensor with shape N_p×C×M.Source: Symbol Meaning Usage · p. 87

Tokenization and reconstruction

  • M_x(·) and M_z(·) mask raw-signal patches and embedded tokens; D_φ(·), T_ψ(·), z_k,c, and barred reconstructions denote reconstruction and tokenization components.The barred symbols represent reconstructions of X, P, and Z.Source: Symbol Meaning Usage · p. 87

Pre-training objectives

  • sim(·,·) denotes similarity, F_r(·) a spectral transform with target type r, Q(·) a quantizer, I a codebook-index sequence, and u_1:L an ordered autoregressive token sequence.The listed spectral target types are spec, sa, bp, and freq.Source: Symbol Meaning Usage · p. 87

Optimization and evaluation

  • The notation defines pre-training and task-j fine-tuning optima, downstream tasks and splits, trial counts, ground-truth labels and predicted distributions, and an indicator function for evaluation metrics.N_j and n_j denote total and fine-tuning trial counts for task τ_j.Source: Symbol Meaning Usage · p. 87

Loss notation

Related papers

Loading 2601.17883v3…