PaperlayerView exact revision on arXiv

Source-linked AI summary

Separating personal from population gains when calibrating EEG foundation models for new users

Xilin Tao, Kani Chen

arXiv:2609.34801v1cs.LGeess.SPq-bio.NC

TL;DR

Brain–computer interfaces require calibration for new users, and apparent personalization gains may reflect a stronger population model rather than user-specific information. The paper evaluates personal adaptation of frozen EEG foundation models using population and exchanged-adapter comparisons. Personal fitting improves performance and favors its owner across the tested combinations, while the size and affordability of that benefit vary.

  • Problem

    In brain–computer interfaces, calibration for each new user is costly, and an apparent personal gain may instead reflect a stronger population model.

    Source: §1 · pp. 1–2

  • Method

    The study evaluates personal fitting in same-session motor imagery, comparing it with a population decoder and adapters fitted to other subjects.

    Source: §1 · p. 2

  • Results

    Across all nine model–dataset combinations, personal fitting improved mean BA over the population reference by 1.5–5.4 percentage points and outperformed exchanged adapters by 2.3–7.3 pp.

    Source: §2.1 · p. 3, §3 · p. 7

  • Takeaways & Limitations

    Personalization should be reported against both a named population reference and exchanged parameters across population-training budgets.

    Source: §3 · pp. 7–8

    The recorded checkpoints do not establish the behavior of a fully trained population model.

    Source: S8.5 New-model convergence records · p. 45

Abstract

from arXiv · show

Foundation models are increasingly adapted to individual users, but an apparent personalization gain can simply reflect a stronger population model. This distinction matters for brain-computer interfaces, where every new user must be calibrated. We evaluated personal adaptation of three frozen EEG foundation models (CBraMod, REVE and LaBraM) in 235 held-out subjects from three motor-imagery datasets, comparing each subject's adapter with the population model and with adapters fitted to other subjects. Using all first-half session labels, personal adapters improved mean balanced accuracy over the population model by 1.5-5.4 percentage points and outperformed exchanged adapters by 2.3-7.3 points in all nine model-dataset combinations. The size of this benefit depended on population training: with four times the original budget, median gains remained positive (1.0-2.0 points) but were smaller for every model, and no population model reached a confirmed plateau. Acquiring the benefit cheaply was unreliable: few-label calibration was consistently non-negative on only one dataset, and in CBraMod neither unlabeled context nor meta-learned initialization outperformed matched controls. Personalization should therefore be evaluated against both a population reference and exchanged parameters, across population-training budgets.

1 Introduction

Across the tested model–dataset combinations, personal fitting provides benefit and an owner-matching advantage, while its size depends on population training and cheaper acquisition is unreliable. The introduction argues for separating personal information from population-model improvements using paired comparisons.

1 Introduction

  • For brain–computer interfaces, calibration of each new user costs time and sustained participation, motivating reuse of information across people or sessions.EEG foundation models such as CBraMod, REVE, and LaBraM learn reusable representations from large recording collections.Source: §1 · p. 1
  • Personal adaptation methods include feature-wise modulation, low-rank updates, learned initializations, personal adapter branches, subject-specific encoders, and transfer from resting-state recordings.An apparent personal gain can instead reflect a stronger head, greater trainable capacity, or more population training.Source: §1 · pp. 1–2
  • Net benefit over a population decoder and the owner-matching advantage answer distinct questions, so either comparison alone is incomplete.An adapter fitted to another person may transfer, while parameter exchange can also damage predictions even if the owner's adapter does not improve on the population model.Source: §1 · p. 2
  • Personal fitting provides benefit and an owner-matching advantage across tested model–dataset combinations, while its magnitude depends on population training and cheaper acquisition is unreliable.The introduction recommends reporting both contrasts against a named population reference across training budgets.Source: §1 · p. 2
  • Full earlier-label fitting characterizes achievable behavior within a fixed adapter family and optimization budget, whereas few-label, unlabeled-context, and meta-learning routes seek the benefit with less target information.The study evaluates same-session motor imagery in subjects excluded from downstream population training and validation.Source: §1 · p. 2

2 Results

Across 235 held-out subjects, full-support personal fitting improved balanced accuracy over population decoders and favored each adapter’s owner. The benefit depended on population-training budget, while cheaper calibration methods recovered it inconsistently.

2.1 Personal adapters improve performance and favor their owners across models

  • Personal LoRA improved mean BA over the population reference by 1.5–5.4 pp and outperformed exchanged adapters by 2.3–7.3 pp across all nine model–dataset combinations.Every corresponding median was also positive.Source: §2.1 · p. 3, Table 2 · p. 4, Figure 2 · p. 4

2.2 Personal gains depend on population training and remain positive at the largest budget

  • At four times the initial population-update budget, pooled median own-minus-population gains were lower for every model but remained 1.0–2.0 pp; no population model reached a confirmed plateau.REVE had the largest observed reduction; CBraMod and LaBraM followed nonmonotonic paths.Source: §2.2 · pp. 4–5, Figure 3 · p. 5

2.3 Cheaper calibration recovers the benefit inconsistently

  • Few-label calibration had nonnegative mean gains at every tested label budget for all three models only on PhysioNet.Calibration used earliest chronological label prefixes, which could contain only one class.Source: §2.3 · p. 5, Figure 4 · p. 6
  • In CBraMod, unlabeled context and meta-learned initialization showed no dependable advantage over matched controls or ordinary continuation.CBraMod was the sole backbone tested for unlabeled context and meta-learning.Source: §2.3 · pp. 5, 7, Figure 5 · p. 6, Figure 6 · p. 7

3 Discussion

Across the nine core model–dataset combinations, personal fitting improved on the population decoder and favored its owner, but gains depended on population training and cheaper calibration did not recover them reliably. The authors recommend reporting both contrasts against a named population reference across training budgets, while noting the evidence is limited to EEG.

3 Discussion

  • Across nine core model–dataset combinations, personal fitting improved on the population decoder and favored its owner, but gains depended on population training and cheaper calibration did not recover them reliably.The authors recommend reporting net personal benefit and owner-matching advantage against a named population reference across training budgets.Source: §3 · p. 7
  • The two contrasts distinguish net benefit from owner matching, but either alone is incomplete because donor exchange can harm predictions and useful personal fitting can transfer to others.Interpretation depends on donor support sizes, class balance and same-session recording conditions.Source: §3 · p. 7
  • Lopes et al. report no accuracy advantage for personal encoders over a fine-tuned shared baseline when both use target labels [9, p. 20], addressing a different question from this paper’s population-reference comparison.Yang et al.’s ICLR 2026 benchmark compares foundation models with traditional decoders across tasks and evaluation protocols.Source: §3 · p. 7
  • Training-budget curves make the population reference part of the result, while behavior beyond the tested budgets remains unresolved.Additional updates can change validation and query performance differently; acquisition methods should distinguish initial-decoder quality from the adaptation increment.Source: §3 · p. 8
  • The evidence is limited to EEG, specifically binary motor imagery with within-session calibration, a small BNCI cohort and a channel-restricted Lee readout.Pointwise subject-bootstrap intervals are conditional on selected checkpoints, omit dependence from shared population models and reused donors, and do not establish benefit for every user; the negative meta-learning result reflects one procedure and does not establish equivalence.Source: §3 · p. 8

4 Methods

The evaluation compares each held-out subject’s personal adapter with a population decoder and exchanged adapters, while varying the population-training budget. It also tests whether cheaper calibration and control procedures capture personal benefit.

4.1 Evaluation unit and estimands

  • Held-out subjects’ earlier same-session trials form support and later trials form query; Δ measures own-fit benefit over population, while U measures specificity against exchanged parameters.Both contrasts are needed: a harmful swap can make U positive even when Δ is negative.Source: §4.1 · p. 9

4.2 Population-budget curves as part of attribution

  • Population-budget curves compare both contrasts while holding subjects, donor procedure, and validation-selected personal-fitting grid constant.Population test performance and validation trajectories accompany the contrasts to show how reference quality changes with update count.Source: §4.2 · p. 9

4.3 Three control families

  • Controls separate shared capacity from target information using contextual interventions, donor exchanges, and additional population training.Context is shuffled within dataset; personal adapters are exchanged among same-dataset, same-fold subjects. Ordinary continuation and meta-learning add population training, with meta-learning also using inner-loop computation and sample exposure.Source: §4.3 · pp. 9–10

4.5.1 Signal processing and subject selection

  • The evaluation uses motor-imagery datasets with subject-disjoint population training, earlier support trials, and later two-class query trials for balanced-accuracy scoring.Preprocessing, selected sessions, channel readouts, exclusions, and donor eligibility vary by dataset; BNCI’s own-versus-swapped contrast covers only donor-eligible subjects.Source: §4.5.1 · pp. 10–11, §4.5.2 · p. 11, Table 3 · p. 11

4.5.3 Pretraining overlap

  • Reported pretraining-source lists overlap with some evaluated datasets for REVE and LaBraM, but none are listed for CBraMod.These are source-list comparisons without record-level deduplication; “unseen subject” means exclusion from downstream population training and validation.Source: §4.5.3 · p. 11

4.6 CBraMod implementation and selection

  • CBraMod uses contextual modulation or mixing over shared LoRA bases, while REVE and LaBraM use shared attention LoRA; personal fitting keeps shared parameters and backbones fixed.Exchange averages weight updates for multi-donor LoRA transfer; validation subjects select personal-fitting settings.Source: §4.6 · p. 12, §4.7 · p. 12

4.9 Statistical analysis and fixed decision rules

  • Bootstrap intervals are conditional on selected fits and omit shared-model, reused-donor, model-selection, and training-seed uncertainty.Positive group summaries do not establish benefit for every subject; directional tests use Holm correction, and the staged decision rules were fixed before each stage’s results.Source: §4.4 · p. 10, §4.9 · pp. 12–13

4.10 Population-training budget extension

  • Population continuation extends each run’s initial update count to twice and four times that count, without intermediate checkpoint selection.Personal-fitting settings are reselected at new endpoints; few-shot results use only the initial population-training budget.Source: §4.10 · p. 13

Data availability

The paper identifies source repositories for the original EEG recordings and states that access and reuse follow those repositories’ terms.

Data availability

  • Original recordings are available from PhysioNet, Dreyer, GigaDB, and BNCI2014-001, subject to source-repository terms.Supplementary Data 1 contains 904 comparison records, and saved subject-level results, split definitions, and numerical provenance records accompany the analysis-code release.Source: Data availability · p. 14

Code availability

The analysis code, configurations, dependencies, reporting scripts, and saved results are publicly available in a versioned release.

Code availability

  • Code and saved results are publicly available at the project’s GitHub release and Zenodo archive.Reporting scripts regenerate summaries and figures from saved results without training or model inference.Source: Code availability · p. 14

Funding

Computational resources and AI-tool access were provided through Kani Chen’s research group at HKUST.

Funding

  • Kani Chen’s research group at The Hong Kong University of Science and Technology provided computational resources and access to AI tools.Source: Funding · p. 17

Competing interests

This material records competing interests, notation, bootstrap uncertainty procedures, and intervals for the paper’s main comparisons. The pooled results show positive median personal benefit at the original training budget for all three models.

Competing interests

S1 Subject-bootstrap uncertainty and mean figures

  • Bootstrap intervals resample subjects while retaining paired observations, but remain conditional on recorded fits and omit shared-model dependence, donor reuse, selection, and seed uncertainty.Pooled resampling retains the empirical subject mixture; per-dataset resampling uses that dataset alone.Source: S1 Subject-bootstrap uncertainty and… · p. 21

S1.1 Complete intervals for the main comparisons

S2 CBraMod Supplementary Results

The supplementary results include alignment experiments and comparisons with related work. They report a transductive Euclidean-alignment improvement in one configuration, while emphasizing that the comparison does not isolate spatial mixing or support a general claim that alignment is harmful.

S2.1 Euclidean alignment and positive counterexamples

  • All-unlabeled-task Euclidean alignment with a multilayer head improved Dreyer by 1.43 pp on average, but used unlabeled query-task inputs and was transductive.The authors report substantial PhysioNet degradation in the same experiments.Source: S2.1 Euclidean alignment and positive… · p. 29
  • Because whitening changes amplitude and spatial structure and precision paths differ, the alignment comparison does not isolate spatial mixing alone.The authors caution that favorable results from other alignment approaches limit general claims that alignment is harmful.Source: S2.1 Euclidean alignment and positive… · p. 29

S2.2 Rest-based transfer and few-shot disagreements

  • The supplement notes that positive results from personal encoders and EEG-Reptile can coexist with its negative small-budget outcomes, which bound generalization rather than invalidate those studies.Label budget, target parameterization, source-model quality, and validation size may contribute to the differing outcomes.Source: S2.2 Rest-based transfer and few-shot… · p. 31

S3.1 Control-by-control comparison with Lopes et al.

  • The literature comparison concerns study designs and inspected controls, not an accuracy ranking; this study asks about joint classification benefit in frozen EEG foundation models as population training varies.Its control scope includes matched-capacity true, shuffled, and default context comparisons and a named population reference.Source: Table S4 · p. 31, S3.1 Control-by-control comparison… · p. 32

S3.2 Distinguishing model benchmarking from personal attribution

  • Model-ranking benchmarks address a different question from the paired adaptation contrasts used to attribute personal benefit and owner specificity.The cited foundation-model benchmark compares models with traditional decoders across evaluation protocols.Source: S3.2 Distinguishing model benchmarking… · p. 32

S3 Literature Comparison and Search Scope

S4 Complete Prespecified Checks and Comparisons

The appendix preserves prespecified comparison decisions alongside unavailable or unexecuted items and continuation records. Several favorable individual components do not pass joint decision rules, and none of the reported runs confirms joint convergence.

S4 Complete Prespecified Checks and Comparisons

  • The appendix inventory includes negative and nonsignificant results; repeated pooled, dataset, and subgroup summaries are not independent replications.Unavailable or inapplicable entries remain blank rather than being treated as zero.Source: S4 Complete Prespecified Checks and… · p. 32

S4.2 Original joint gates and specified secondary endpoints

  • Contextual FiLM and LoRA both failed their joint gates because the required p-value comparisons did not pass, despite reported recovery values above 0.30.Their reported drops also exceeded the rule’s 0.10 threshold.Source: S4.2 Original joint gates and… · p. 34
  • Personal-gain gates passed for FiLM offset, mixture offset, and LoRA, while nearest-versus-random donor gates failed for all three variants.The personal-gain rule used a pooled median threshold of 2 pp; the donor comparison followed the personal-gain gate.Source: S4.2 Original joint gates and… · p. 34
  • Both meta-learning initialization variants failed the prespecified main gate, although M1 passed the separate adaptation-increment comparison and M2 failed it.The main-gate observed medians were 0.517 pp for M1 and 0.576 pp for M2; their drops were 0.123 and 0.132.Source: S4.2 Original joint gates and… · p. 34

S4.4 Cross-model scope

  • The completed cross-model core covers LoRA for REVE and LaBraM, while their FiLM, neighbor, external-dataset, and meta-learning extensions remain unexecuted.The three-model continuation curves are reported separately, including CBraMod’s documented precision exception and initial failure.Source: S4.4 Cross-model scope · p. 35

S4.5 Run-level convergence records

  • No run confirms the joint training-and-validation plateau criterion, and the continuation table reports additional updates rather than evidence of convergence.“Both flat” requires the original training and validation plateau criteria together.Source: S4.5 Run-level convergence records · p. 36, Table S9 · p. 36

S5 All statistical comparison rows

Supplementary Data 1 is the complete inventory of the paper’s statistical comparison rows, with identifiers and supporting documentation. It includes negative and nonsignificant results as well as unavailable entries.

S5 All statistical comparison rows

  • Supplementary Data 1 contains all 904 comparison rows, including negative and nonsignificant results.Source: S5 All statistical comparison rows · p. 39
  • Each row retains original identifiers, values, missing entries, source-row links, decisions, units, contrast directions, and correction-family descriptions.A Holm-adjusted value belongs only to its identified source family; no correction across the combined inventory is applied.Source: S5 All statistical comparison rows · p. 39
  • The inventory is available as CSV and XLSX, with a README describing CSV columns and XLSX sheets for columns, families, and notation.A blank adjusted value indicates unavailable information rather than zero.Source: S5 All statistical comparison rows · p. 39

S6 Recorded training resources

The supplement records the study’s resource use, controls, configuration grids, and interpretation rules. Its supplementary comparisons include a query-informed meta-learning control, descriptive cross-model contrasts, and criteria for few-shot reliability and population-model convergence.

S6 Recorded training resources

  • The study records 21.14 GPU card-hours for population and personal adaptation and 7.30 for ordinary continuation/BNCI follow-up, not a complete exploratory-work total.The initial budget batch used 17.81, 51.59, and 18.70 GPU card-hours for CBraMod, REVE, and LaBraM, respectively, including a failed CBraMod run.Source: S6 Recorded training resources · p. 39

S7 Query-informed prior controls and configuration grids

  • At ten labels, first-order MAML and learned-rate variants gained 0.52/0.58 pp over R1, but both failed the original joint criterion under a query-informed control.R1 selected its prior neighborhood and shrinkage using test-pool query scores, so these decisions are not independent confirmation under test-label isolation.Source: S7 Query-informed prior controls and… · p. 40

S7.1 Configuration grids and execution identifiers

S8.1 Effect summaries and source strata

S8.2 Original diagnostics and descriptive comparisons

S8.3 Few-shot reliability and original descriptive endpoints

  • Few-shot stability requires nonnegative mean gains at every prescribed label budget and no decreases as the label budget rises.This criterion does not require every person to improve or function as an inferential success threshold.Source: S8.3 Few-shot reliability and original… · p. 44

S8.4 Prespecified family inventory

  • The supplementary family inventory specifies the cross-model scope of the evaluation families; its caption does not report their outcomes.The REVE and LaBraM few-label tables identify gains in percentage points and subject decline fractions, without values in the supplied passages.Source: Table S16 · p. 44, Table S17 · p. 44, Table S18 · p. 45

S8.5 New-model convergence records

  • Every recorded plateau field is false, so the checkpoints do not establish how a fully trained population model would behave.The continuation-record caption distinguishes selected from executed updates and states that the stopping rule does not establish a joint plateau.Source: S8.5 New-model convergence records · p. 45, Table S19 · p. 45

S9 Completed population-budget extension results

The completed extension compares personal-adaptation contrasts at the original, doubled, and quadrupled population-training budgets. Pooled personal gains remain positive but are smaller at four times the initial budget, and the records do not establish population-model convergence.

S9.1 CBraMod

  • The completed CBraMod budget table reports population balanced accuracy and personal benefit and specificity in percentage points.The supplied caption identifies its scope but does not provide cell values.Source: Table S20 · p. 47

S9.2 REVE

  • The completed REVE budget table reports population balanced accuracy and personal benefit and specificity in percentage points.The supplied caption identifies its scope but does not provide cell values.Source: Table S21 · p. 47

S9.3 LaBraM

  • The completed LaBraM budget table reports population balanced accuracy and personal benefit and specificity in percentage points.The supplied caption identifies its scope but does not provide cell values.Source: Table S22 · p. 47

S9 Completed population-budget extension results

S9.4 Prespecified interpretations and complete comparison inventory

  • The interpretation rules do not establish convergence or a pure causal effect of compute; the operational criterion is not a statistical demonstration of curve stability.Dataset-level comparisons remain descriptive, and the budget-extension family reserves correction slots for unexecuted variants.Source: S9.4 Prespecified interpretations and… · p. 48

S9.5 Validation behavior alongside the test curves

  • REVE’s population mean BA fell from 81.13 [79.51, 82.70]% to 79.81 [78.32, 81.32]% across the reported endpoints, alongside its remaining personal contrast.The completed-run summaries weight fold/seed runs equally, and the runs ended at their budget caps without a joint plateau.Source: S9.5 Validation behavior alongside the… · p. 49, Table S25 · p. 49

S9.6 Initial CBraMod failure, missingness and run-level record

S9.7 Recorded numerical events in the meta-learning runs

S10 Supplementary Methods

The supplementary methods document decision rules, model parameterization and training procedures, and the scope of the evaluation. They specify population-budget endpoints and limits on interpreting cross-model contrasts, donor-based comparisons, and data checks.

S10.1 Decision rules and interpretation

  • Stage-specific rules were fixed before their respective results, but the study was not a single independently preregistered confirmatory design.The supplement preserves original decisions, including stopped extensions and failed gates.Source: S10.1 Decision rules and interpretation · p. 54

S10.2.1 Shared and personal parameterization

  • Personal FiLM adds zero-initialized offsets to generated modulation, while personal LoRA adds a zero-initialized update with shared bases, mixture weights, head, and backbone fixed.Full-support exchange draws ten other subjects uniformly with replacement from the same dataset and held-out fold.Source: S10.2.1 Shared and personal… · p. 54

S10.2.2 Continuation and meta-learning

  • Meta-learning and ordinary continuation share an initial population LoRA model and scheduled outer-update budget, but meta-learning additionally computes support-set updates.The inner loop updates personal LoRA; query-loss backpropagation updates shared bases and the task head without differentiating through support updates.Source: S10.2.2 Continuation and meta-learning · p. 55

S10.3 REVE and LaBraM implementation

  • REVE and LaBraM use native EEG channels and distinct recorded input pipelines, and their between-model performance differences cannot be attributed to backbone alone.REVE uses overlapping patches without tail padding; LaBraM uses nonoverlapping patches and reads EEG patch tokens.Source: S10.3 REVE and LaBraM implementation · pp. 55–56

S10.4 Correction families and descriptive rules

  • The full-support fit is an empirical reference within the tested adaptation family and population budgets, not an optimized upper bound over all possible adaptations.Cross-model consistency and few-shot stability are descriptive rules, not claims that every subject improves or new significance thresholds.Source: S10.4 Correction families and… · p. 56

S10.5 Euclidean alignment implementation

  • The all-task alignment variant can access unlabeled target queries and is transductive; the alignment and unaligned activation-precision paths also differ.Thus spatial mixing is not isolated by these comparisons.Source: S10.5 Euclidean alignment implementation · p. 56

S10.6 Population-budget endpoints and interpretation rules

  • Budget endpoints use each run’s original checkpoint, then the states at twice and four times its cumulative selected update count, without selecting a better intermediate checkpoint.Personal settings are selected again at added endpoints, while the reference endpoint reuses its original personal fit and selection.Source: S10.6 Population-budget endpoints and… · pp. 56–57

S10.7 Screening procedures and scope of Cho timeline checks

  • Cho trial ordering uses available official metadata, but verification does not reconstruct every subject’s continuous recording or establish the absolute rest-to-task interval.Independent signal-level and packaged-event checks covered subject s01.Source: S10.7 Screening procedures and scope… · p. 58

Related papers

Loading 2609.34801v1…