acceptodds
Under review as a conference paper at ICLR 2027

Spectral Accessibility: Probing What EEG Foundation Models Preserve

Abstract

Downstream accuracy measures task utility, but does not reveal how broadly a frozen EEG representation exposes signal structure. We introduce spectral accessibility, a label-independent audit of a declared second-order state built from multiresolution complex cross-spectral matrices. A subject-disjoint fit selects and whitens 64 reliable state directions. Their recoverability defines an accessibility spectrum; SPA-20 tests broad coverage by averaging its weakest 20% under the worst declared input environment. We distinguish the unrestricted population operator, its best-linear counterpart, and the finite cross-fitted ridge estimate. Population theory gives data processing inequalities and task-error bounds, including bounds restricted to accessible subspaces. Across five masked-reconstruction or masked prediction EEG foundation models, seven datasets, and four input conditions, all 140 estimated linear lower tails reach zero. Strong split-taper controls on four datasets show that the same pipeline can recover the target’s weak directions, although pooling, subject transfer, and probe estimation remain possible bottlenecks. Non-zero access is concentrated in upper subspaces. We report their exploratory environmental profiles and conditional uncertainty. In this regime, SPA-20 is a diagnostic stress test; the full spectrum, null floor, and split-taper reference carry the comparative information.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Representation Learning, Foundation Models, Representation Evaluation, EEG, Spectral Analysis

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_hYGZOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes "Spectral Accessibility", a label-independent audit framework to evaluate what signal structures are preserved by frozen EEG foundation models (EEG-FMs). The authors construct a target state from multiresolution complex cross-spectral density (CSD) matrices, select and whiten 64 reliable PCA directions on a subject-disjoint set, and use nested cross-subject ridge regression to predict these directions from pooled model embeddings. They define SPA-20 to measure the average recoverability of the weakest 20% spectral directions under the worst input environment. Across 5 masked-reconstruction/prediction models, 7 datasets, and 4 environments (140 cells in total), all estimated linear SPA-20 scores collapse to zero, while non-zero accessibility is concentrated in upper subspaces (analyzed via a post-hoc U-SPA-20 metric).

Critical strengths
  1. Complementary evaluation perspective: Moving beyond downstream task accuracy to audit what intrinsic second-order spectral properties (power and cross-channel phase) are preserved in frozen representations is a well-motivated direction.
  2. Rigorous statistical controls and honesty: The use of split-taper reliability references and subject-block derangement nulls is solid. The authors also honestly report the complete collapse of their primary pre-specified metric (SPA-20) without disguising post-hoc upper-tail analyses as primary results.
Critical weaknesses
  1. Confusing feature extraction with lossless compression: The goal of EEG-FM is to extract task-relevant low-dimensional manifolds and suppress noise, not lossless signal reconstruction. The paper treats all 64 whitened PCA directions equally and penalizes models on the weakest 20% directions. Mathematically repeatable weak directions across split tapers may simply be stable physiological artifacts (e.g., EMG) or background noise. Without downstream criterion-validity experiments, treating the removal of weak spectral directions as a failure is unconvincing.
  2. Collapse of the main metric and degenerate theoretical bounds: L-SPA-20 is exactly 0 in all 140 test cells. At this boundary (s=0s=0), the derived global task-error bound degenerates into a trivial zero-state baseline, while the localized bound admits that tasks aligned with upper subspaces can still succeed. Thus, SPA-20 loses all discriminative power across models and environments, forcing the paper to rely on a post-hoc descriptor (U-SPA-20).
  3. Confounded pipeline due to global mean pooling: Up to 96.7% of raw target coordinates come from fine 2-second local time windows, yet the adapter pipeline crudely applies global mean pooling over all time and channel tokens before probing. Predicting local dynamic spectral blocks from a single time-averaged static vector naturally destroys degrees of freedom. Without token-level or window-aligned ablations, the audit cannot separate backbone information loss from trivial pooling loss.
  4. Mismatch between pretraining objectives and the linear probe: The evaluated models use very different pretraining paradigms. Counter-intuitively, LaBraM—the only model explicitly pretrained to reconstruct Fourier amplitude and phase—achieves the worst spectral accessibility score. The paper ignores that discrete vector-quantization bottlenecks may encode spectral features in non-linear manifolds (making a shared-penalty linear ridge probe unfair) and fails to explain this obvious contradiction.
Questions for authors
  1. Value of the lower tail: Can you provide any empirical evidence (e.g., correlation with downstream tasks) showing that preserving the weakest 20% whitened CSD directions is actually beneficial for BCI tasks, rather than just retaining task-irrelevant background variance?
  2. Utility of SPA-20: Given that L-SPA-20 is 0 across all 140 cells and its theoretical task bound becomes vacuous at s=0s=0, what is the practical utility of keeping SPA-20 as the core framework instead of directly formulating the audit around subspace alignment or effective rank?
  3. Unpooled ablation: Since cached embeddings are absent, can the authors run a small-scale ablation on at least one dataset (e.g., HMC) using unpooled token-level or window-aligned representations to clarify how much of the zero lower-tail is purely an artifact of global mean pooling?
  4. Why LaBraM fails: How do you explain that LaBraM explicitly reconstructs Fourier spectra during pretraining but ranks last in your spectral audit? Would a simple non-linear probe (e.g., 2-layer MLP) or per-coordinate ridge regularization significantly change the ordering of LaBraM?
Reviewer_ChAqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Spectral Accessibility, a label-independent framework to audit structural signal preservation in frozen EEG foundation models. Using a cross-fitted multi-output Ridge regression probe across 5 models, 7 datasets, and 4 perturbations, it measures lower-tail (SPA-20) and upper-tail (U-SPA-20) spectral coverage. The key finding is that SPA-20 collapses to zero across all 140 cases, showing linear recovery is heavily restricted to a narrow upper subspace.

Critical strengths
  • The paper addresses a critical gap by evaluating representations beyond downstream task accuracy, which can hide structural signal loss.
  • The study evaluates 5 major foundation models across 7 biomedical datasets and 4 input perturbations.
Critical weaknesses
  • The main finding that SPA-20 reaches zero everywhere is confounded by using a single scalar Ridge penalty selected via global MSE across all 64 target coordinates. Because high-variance upper components dominate MSE, weak lower-tail directions are severely under-regularized, causing their eigenvalues to become negative and forcing the clipping mechanism to artificially push the lower-tail score to zero.
  • The framework's use of a linear Ridge probe assumes that un-decodable signals are lost signals. In reality, deep Transformers map EEG data into non-linear manifolds where critical features are preserved yet hidden from linear probes. Therefore, mistaking a linear probing failure for information destruction overlooks the nature of non-linear geometric encoding
  • The framework relies on several fixed, heuristic hyperparameters without clear justification, such as a 2-second CSD window and feature pooling across epochs without sensitivity analysis. Because this temporal pooling may destroy fine local dynamics, window-aligned or token-level probes are needed to rule out an operational bottleneck.
Questions for authors
  • Would fitting the Ridge probe with coordinate-specific penalties restore non-zero lower-tail accessibility on high-sample datasets?
  • If input adapters and embedding dimensions were fully standardized, how much of the U-SPA-20\text{U-SPA-20} gap between models would remain?
Reviewer_QTt7OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces Spectral Accessibility, a label-independent framework for evaluating which second-order spectral properties of EEG signals are recoverable from frozen foundation-model representations. The authors construct a multiresolution cross-spectral state, apply reliability-based dimensionality reduction and whitening, and define an accessibility operator whose eigenvalue spectrum characterizes the recoverability of different state directions. A lower-tail summary, SPA-20, is proposed to measure broad spectral accessibility under multiple input perturbations. The framework is evaluated on five EEG foundation models across seven datasets and four input conditions. Interestingly, all 140 estimated linear SPA-20 scores are zero, while the upper portions of the accessibility spectra remain nonzero. The authors subsequently introduce an exploratory upper-tail measure to characterize this remaining accessibility.

Critical strengths
  1. A label-free audit of what frozen EEG representations keep is a useful complement to task benchmarks.
  2. The experimental setup is careful: subject-disjoint splits, nested grouped CV, and derangement nulls.
  3. The paper is honest about its limits. It clearly separates population theory from the finite estimator and flags its post-hoc analyses.
  4. Five foundation models across seven datasets and four input conditions provides useful empirical breadth
Critical weaknesses
  1. The paper is difficult to follow, and the writing needs significant improvement. There is a lack of flow between sections, with ideas introduced without clear connections, making it hard to understand how the different parts of the paper fit together. Several sections are also overly technical and repetitive, and the authors should focus on presenting their motivation, methodology, and results in a clearer and more organized manner. It feels like a lot of jargon is being used without clear motivation or connection.

  2. The main metric never moves. SPA-20 is zero everywhere, so it can't distinguish models or environments, and the worst-case-over-environments design is never exercised. The score that does vary (U-SPA-20) was chosen after the primary metric collapsed, which makes it exploratory at best.

  3. I'm not convinced the zeros reflect the models rather than the probe. The split-taper control predicts one near-copy of the target from another. That is a much easier problem than predicting from a 200–512-d embedding with as few as 5–7 score subjects. With small samples, the smallest eigenvalues get pushed down and then clipped to zero. A planted-signal power analysis would show whether this pipeline could ever detect a non-zero lower tail.

  4. The model ranking is confounded. The two top models (REVE, EEGPT) both have 512-d embeddings, and the bottom three are 200-d. The intervals are also wide. A width-matched comparison is needed before reading anything into the ordering.

  5. Other issues:

a. The paper is hard to read because nearly every sentence is hedged.

b. Raw spectra are not released.

c. Cohorts are capped at 30 subjects even where far more exist (e.g., HMC has 151).

Questions for authors

See weaknesses

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.