{"id":"8187d120-c588-4322-95fc-37a7fb674633","arxiv_id":"2608.01898","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"FAME, a frequency-balanced masked autoencoder that standardizes and equally weights per-band reconstruction targets, reduces the low-frequency bias of EEG representations and beats prior models on 24 of 41 downstream tasks.","lead":"This paper finds that EEG foundation models consistently favor low-frequency brain signals and proposes FAME, a pretraining method that weights every frequency band equally during reconstruction. The method improves accuracy on 24 of 41 downstream EEG tasks, suggesting that balanced spectral supervision yields more transferable representations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"FAME-vs-MAE comparison conflates frequency balancing with dense supervision, so the central causal attribution of the gains to band-wise weighting is not yet isolated.","rationale":"The reader identified the spectral-diagnostic confound in Eqs. 3-5 as the weakest assumption, and I agree that the Wasserstein and recoverability metrics are computed on 1/f^alpha signals and may partly reflect input power rather than encoder deficiency. However, the controlled MAE/FAME comparison was designed to avoid that issue by keeping the input distribution fixed, so the more decisive weakness is that this comparison itself is confounded. Because FAME's loss includes dense supervision over unmasked positions while MAE reconstructs only masked positions, the observed differences in Fig. 4 and the downstream tables cannot be uniquely attributed to band-wise standardization and equal weighting. This is a resolvable internal-validity issue: adding a masked-only FAME arm would settle it. The empirical package, including 24/41 SOTA and the NeuralBench comparison, may still stand, so this is not a rejection; it strengthens the conditionality of the paper. The reader's verdict of CONDITIONAL remains appropriate, with the additional condition that an ablation isolating dense supervision from frequency balancing be reported before the causal claim is treated as established.","tokens_in":26179,"tokens_out":7732,"duration_ms":74788,"concrete_test":"Run the controlled TUAB experiment with a third arm: FAME with LFAME evaluated only on masked channel-time positions, i.e., restrict Omega_b in Eq. 17 to positions with m_i=1, keeping the same band-wise standardized log-power targets and equal band weights. Compare Fig. 4(A-B) band-wise loss trajectories and W1 distances, plus downstream TUEV/HFO linear-probing, against both standard MAE and dense FAME. If masked-only FAME retains the balanced trajectories and downstream gains, dense supervision is not the driver; if it collapses to MAE-like low-frequency bias, the central attribution to frequency balancing fails. A fourth arm with dense raw-signal MAE over all positions would further separate target representation from supervision density.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that equalized band-wise supervision corrects low-frequency bias and drives state-of-the-art transfer. The evidence for this causal role rests on the controlled comparison (Fig. 4) and the supplementary FAME-vs-MAE results. But FAME's objective, Eq. 17, is defined over Omega_b containing 'both masked and unmasked positions' — dense supervision over the full sequence — whereas the conventional MAE control reconstructs only masked positions, as is standard and as the paper's MAE description implies. The 'Controlled Comparison' section says the two models 'differ in their reconstruction heads and targets' but does not mention this third difference: supervision density. The treatment therefore changes three things at once: targets (raw samples vs band-averaged log-power), weighting (power-weighted vs band-equalized), and supervision density (masked-only vs dense). Any one of these could flatten the band-wise loss trajectories in Fig. 4(B), reduce the W1 differences in Fig. 4(A), or improve linear probing, for example by supplying more gradient signal at high frequencies. Without an ablation that fixes density while toggling band standardization and equal weighting, the paper has not shown that 'frequency-balanced supervision' is the active ingredient; it has shown that the FAME objective as a package works. The reader's input-spectrum confound in Eqs. 3-5 is real but secondary; this confound directly undermines the explanatory claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that existing EEG foundation models exhibit a persistent low-frequency representational bias, explains this bias through the interaction between EEG's 1/f^alpha spectral structure and neural networks' low-frequency learning preference, and proposes FAME, a masked autoencoder that reconstructs independently standardized log-power targets within predefined EEG bands with equal band-wise weighting. The authors evaluate FAME on 41 OmniEEG-Bench tasks with linear probing and report state-of-the-art results on 24 tasks, and further report an advantage under full fine-tuning on a subset of NeuralBench tasks. They also propose diagnostic metrics for frequency bias and correlate them with downstream performance.","tokens_in":26578,"tokens_out":6269,"duration_ms":57192,"significance":"If the causal claim is established, FAME is a simple and broadly applicable pretraining objective, and the proposed frequency-bias diagnostics would be a useful tool for EEG representation analysis. The paper has clear strengths: Eq. (1) is a correct Parseval-based identity that motivates the objective-level imbalance; the controlled small-scale comparison in Fig. 4 is a sensible experimental design; the evaluation covers many datasets; and the Limitations section is honest about the exploratory nature of the bias-performance correlations. However, the central attribution of the gains to band-wise weighting is not yet isolated from supervision density, and several evaluation choices need additional controls before the state-of-the-art claim is fully supported.","major_comments":[{"comment":"The controlled comparison is confounded because the FAME loss in Eq. (17) is defined over Omega_b containing 'both masked and unmasked positions', i.e. dense supervision, while the traditional MAE baseline reconstructs only masked positions. The text states that the two models differ only in their reconstruction heads and targets, but the objectives differ in three ways: targets, band weighting, and supervision density. Any of these could flatten the band-wise loss trajectories in Fig. 4(B), reduce the Wasserstein distances in Fig. 4(A), and improve linear probing by providing more gradient signal at high frequencies. The causal claim that 'frequency-balanced supervision' is the active ingredient requires an ablation that fixes supervision density while toggling band standardization and equal weighting. The supplementary FAME-vs-MAE comparison inherits the same confound.","section":"Controlled Comparison between Traditional MAE and FAME; Eq. (17)"},{"comment":"The representation layer and checkpoint were selected using ADHD validation accuracy, and the main text states that the selected 80%-depth layer was applied consistently to all model variants, including the 50M and 1B models. However, the supplementary linear-probing protocol for the 1B model uses the 15th Transformer block of a 20-block network, which is 75% depth, not 80%. This inconsistency changes the representation being probed across model sizes and can affect the scaling comparison and the reported SOTA numbers. In addition, selecting a layer and checkpoint on ADHD and then reporting ADHD as one of the 24 SOTA tasks risks selection bias; the paper should report sensitivity to layer and checkpoint choice across tasks.","section":"Selection of checkpoints and representation layers; Details of Linear Probe of 1B Model"},{"comment":"The Wasserstein distance in Eq. (3) is computed between embeddings of broadband EEG and embeddings of 5-Hz band-limited versions of the same signals. Because raw EEG power decays as roughly 1/f^alpha, high-frequency band-limited inputs have much smaller variance than low-frequency ones, so a larger W1 at high frequencies may reflect input statistics rather than a deficiency of the learned encoder. The recoverability analysis in Eqs. (4)-(5) standardizes targets and is less affected, but the W1 metric in Figs. 3(A) and 4(A) needs a control that normalizes each band-limited input to equal variance, or a comparison against a trivial or linear encoder, before the result is interpreted as a learned representational bias. This is load-bearing for the cross-model persistence claim and for the explanatory narrative.","section":"Frequency Preference of Pretrained Representations; Eqs. (2)-(3)"},{"comment":"The state-of-the-art claim on 24 of 41 tasks rests on OmniEEG-Bench without significance tests, and several entries have overlapping standard deviations across models. The benchmark allows model-specific input adapters, which is reasonable but further complicates direct comparability. The paper should report paired significance tests across seeds for the highlighted SOTA differences, and should disclose the authors' role in building OmniEEG-Bench in the main text. As written, the SOTA claim is not statistically grounded.","section":"Generalization Performance; Table 1"}],"minor_comments":[{"comment":"The displayed equality is approximate because r_f uses epsilon in the denominator; the approximation should be stated explicitly rather than presented as an exact Parseval identity.","section":"Eq. (1)"},{"comment":"The supplementary text says the MAE comparison covers 42 tasks with FAME better on 27 tasks, while the main text and Table 1 report 41 tasks and 24 SOTA results; the counts should be unified.","section":"Supplementary Figure 6"},{"comment":"The entry '56.87±6.2654.50±2.02' is missing a separator between the FAME-50M and FAME-1B values.","section":"Table 1, BCIC4-1 row"},{"comment":"The 1B model is described with N=24 Transformer layers in the architecture section but with 20 Transformer blocks in the linear-probing supplement; these descriptions should be aligned.","section":"Architecture descriptions"},{"comment":"The correlation analyses use a small number of models and are not corrected for multiple comparisons; the Limitations paragraph acknowledges low power, but the main-text phrase 'consistent relationships' should be softened to match the exploratory nature of the analysis.","section":"Figure 7 and Limitations"}],"recommendation":"major_revision","confidential_remarks":"The paper is promising and likely of interest to the journal, but the causal attribution and evaluation protocol need substantial strengthening. In particular, the dense-supervision confound in the FAME-vs-MAE comparison and the inconsistency in the 1B probing layer are load-bearing for the central claims. The OmniEEG-Bench overlap should also be disclosed clearly, since the SOTA claim rests on that benchmark. The required fixes are feasible within the manuscript's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about arXiv:2608.01898. First, the core observation is real and worth taking seriously: across a range of EEG foundation models, representations are systematically biased toward low-frequency activity, and this bias persists as model size and data scale grow. Second, the proposed fix, FAME, appears to work, but the paper's headline explanation for why it works is not yet fully supported.\n\nWhat is actually new: the synthesis of EEG's 1/f-like spectrum with neural spectral bias, plus a concrete objective—per-band standardized log-power targets with equal band weighting—that directly addresses the power-weighted supervision problem. The Parseval identity in Eq. 1 is correct and cleanly motivates the method. The controlled MAE-vs-FAME comparison in Fig. 4 shows that FAME flattens band-wise loss trajectories and reduces the Wasserstein distance between broadband and band-limited embeddings. The external NeuralBench evaluation, with full fine-tuning, is a good check and FAME-50M does well there. The limitations section is honest, and the supplementary is unusually detailed on dataset harmonization and evaluation protocols.\n\nNow the soft spots, in order of importance. The stress-test concern lands: FAME's objective, Eq. 17, applies supervision to both masked and unmasked positions, while the conventional MAE control reconstructs only masked positions. The controlled comparison says the models \"differ in their reconstruction heads and targets\" but never mentions that they also differ in supervision density. That means the treatment changes three things at once—targets, weighting, and density—so the gains cannot be causally attributed to frequency balancing alone. This is the main thing to fix before the central claim is solid: add an ablation that fixes density and toggles band standardization and equal weighting. The reader's input-spectrum confound in Eqs. 3–5 is also real but secondary; those diagnostics may partly reflect the raw 1/f power distribution rather than encoder deficiency alone.\n\nOther issues are more minor but still worth addressing: the headline SOTA claim rests on OmniEEG-Bench, which is co-authored by this group; there are no significance tests across the 41 tasks; the representation layer and checkpoint were selected using ADHD validation accuracy; the downstream task count is inconsistent across the paper (41, 42, 44, 46); and no code or checkpoints are released. None of these are fatal, but together they mean the paper's strongest claims need tightening.\n\nBottom line: this paper deserves a serious referee and likely publication after major revision. The diagnostic framing is valuable, the method is plausible and well-motivated, and the external benchmark helps. What is missing is a clean ablation that isolates the active ingredient, plus a more cautious benchmark framing. I would bring it to a reading group and would cite it if I worked on EEG pretraining.","headline":"A genuinely useful diagnosis of low-frequency bias in EEG foundation models, plus a plausible fix—but the paper's central causal claim is not yet isolated because FAME changes supervision density along with band weighting.","tokens_in":27007,"tokens_out":1405,"would_cite":true,"duration_ms":15699,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that EEG foundation models are biased toward low frequencies and that FAME, a masked autoencoder that reconstructs per-band standardized log-power time-frequency targets, corrects the bias and reaches state-of-the-art on…","keywords":["EEG foundation models","low-frequency bias","spectral bias","masked autoencoding","time-frequency reconstruction","1/f spectrum","frequency-balanced pretraining","representation transfer"],"falsifier":"Pretrain a masked autoencoder and FAME on spectrally whitened EEG (per-band variance equalized), then recompute the embedding-distribution distances and per-band recoverability losses; if the low-frequency bias and the FAME downstream gap disappear, the input spectrum was the cause rather than the model's learning dynamics.","tokens_in":25967,"feed_emoji":"🧠","tokens_out":10530,"duration_ms":85751,"temperature":0.7,"pith_summary":"EEG foundation models transfer poorly and fail to improve consistently with more data or parameters. This paper identifies a persistent low-frequency bias in their representations and traces it to the interaction of EEG's $1/f^\\alpha$ power spectrum with neural networks' tendency to learn low-frequency functions first, amplified by the $\\ell_2$ reconstruction loss of masked autoencoders, which weights each frequency by its power. The paper's proposed remedy, FAME, reconstructs log-power time-frequency activity in predefined EEG bands, standardizes each band's target independently with equal band weights, and thereby decouples supervision from intrinsic band power. In controlled experiments FAME shows more balanced per-band optimization dynamics, and on 41 downstream tasks its 1B-parameter model reaches state-of-the-art linear-probing accuracy on 24 of them. If correct, the paper shows that spectral supervision, not scale alone, is what makes EEG representations transferable.","feed_headline":"Band-balanced pretraining beats low-frequency bias on 24 of 41 tasks","feed_subtitle":"Scaling data and parameters alone did not fix EEG models; rebalancing supervision across frequency bands did.","key_machinery":"FAME (Frequency-balanced Autoencoding for Masked EEG) is the central method: from masked channel-time patches it predicts dense time-frequency activity per predefined band, sets the targets as averaged log power within each band, standardizes each band with corpus-level mean and variance, and weights every band's loss equally. The load-bearing mathematical identity is the Parseval expansion of Eq. 1, which shows that pointwise mean-squared error in the time domain equals a power-weighted sum of squared relative spectral errors; this is what turns EEG's $1/f^\\alpha$ spectrum into a low-frequency supervision bias. Supporting diagnostics include the Wasserstein distance between broadband and band-limited embeddings and per-band recoverability losses from frozen embeddings, summarized by low-frequency bias, frequency imbalance, and signed frequency slope.","core_discovery":"Existing pretrained EEG encoders systematically retain more low-frequency than high-frequency information, a preference that persists across pretraining objectives, dataset scales, and model capacities. The paper explains this by a power-weighted reconstruction argument: by Parseval's theorem, $\\|x-\\hat{x}\\|_2^2 \\propto \\sum_f |X_f|^2 |r_f|^2$, so under EEG's $1/f^\\alpha$ spectrum, equal relative errors at low frequencies contribute far more to the loss than at high frequencies. FAME replaces raw-signal reconstruction with the prediction of band-averaged log power, where each band's target is standardized with corpus-level statistics and band losses are averaged equally, so no band dominates supervision. On the 41-task benchmark, FAME with 1B parameters achieves the best linear-probing balanced accuracy on 24 tasks, and its 50M version improves on the standard masked-autoencoder baseline in 27 of 42 tasks, supporting the claim that balanced spectral supervision, not scale alone, drives transfer.","pith_inferences":["Beyond the paper's claims: a band-normalized reconstruction target is a natural candidate for any 1/f-like neural signal, such as ECoG or local field potentials, where the same power-weighted-loss argument applies.","Beyond the paper's claims: the bias-performance correlation is based on a small number of pretrained models; a denser sweep over seeds, checkpoints, and architectures would reveal whether the correlation is robust when model size is controlled.","Beyond the paper's claims: whitening the input spectrum before pretraining is a direct test of the causal story — if low-frequency preference persists under whitened inputs, the network's spectral bias is doing the work; if not, the input spectrum alone explains it.","Beyond the paper's claims: an adaptive version that learns band weights or band boundaries per task could help tasks with concentrated discriminative bands, but the paper does not explore adaptive partitions."],"forward_implications":["If FAME's central claim holds, a 1B-parameter transformer trained with frequency-balanced targets converts additional capacity into downstream gains, whereas the same scaling in conventional masked autoencoders does not consistently help.","The paper's bias metrics (low-frequency bias, frequency imbalance, signed slope) can serve as model-level diagnostics: on tasks like TUEV, lower measured bias is associated with higher linear-probing accuracy, so reporting these numbers alongside task accuracy would make representation quality visible.","The benefit of frequency balancing is task-dependent; tasks requiring information across many frequency bands gain most, while tasks dominated by low-frequency activity may see little benefit, so equal band weighting should be viewed as a default to adjust per task.","FAME's advantage extends outside linear probing: under full fine-tuning on 13 independent tasks, the 50M model attains the best average balanced accuracy among the compared models, indicating the representation change is not an artifact of probe calibration."],"supporting_citations":[{"why":"Supplies the spectral-bias result that neural networks fit low-frequency functions first, the premise of the paper's explanation.","marker":"Rahaman et al. 2019"},{"why":"Extends spectral bias to practical training dynamics and generalization, supporting the claim that low-frequency components dominate optimization.","marker":"Fridovich-Keil, Gontijo Lopes, and Roelofs 2022"},{"why":"Establishes the 1/f-like power-spectrum scaling of human EEG, the input property that makes low-frequency supervision stronger.","marker":"Pritchard 1992"},{"why":"Provides the modern parameterization of neural power spectra into periodic and aperiodic parts, grounding the $1/f^\\alpha$ assumption.","marker":"Donoghue et al. 2020"},{"why":"Provides the standardized benchmark and evaluation protocol on which FAME's 41-task state-of-the-art claims are measured.","marker":"Lu et al. 2026"},{"why":"Documents that scaling EEG foundation models does not consistently improve downstream performance, the pattern FAME is designed to fix.","marker":"Yang et al. 2026"},{"why":"A frequency-domain masked-reconstruction baseline and an existing EEG foundation model that FAME is compared against.","marker":"Jiang, Zhao, and Lu 2024"},{"why":"Provides the independent 13-task protocol used for the full-fine-tuning evaluation of FAME.","marker":"Banville et al. 2026"}],"fun_headline_variants":["Frequency-balanced EEG pretraining wins on 24 of 41 tasks","Rebalancing EEG supervision unlocks transfer on 24 of 41 tasks","Low-frequency bias fixed by band-balanced EEG pretraining","Balanced band losses fix EEG low-freq bias on 24/41 tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The core assumption is that the measured low-frequency preference belongs to the learned encoder; if the metrics merely mirror the input power spectrum, the explanation and the FAME advantage are not as clean as claimed.","fun_headline_variants_meta":{"raw":{"variants":["Frequency-balanced EEG pretraining wins on 24 of 41 tasks","Rebalancing EEG supervision unlocks transfer on 24 of 41 tasks","Low-frequency bias fixed by band-balanced EEG pretraining","Balanced band losses fix EEG low-freq bias on 24/41 tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00081,"raw_usage":{"total_tokens":3559,"prompt_tokens":956,"completion_tokens":2603,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":2536}},"tokens_in":572,"tokens_out":2603,"duration_ms":16031,"temperature":1.0,"reasoning_tokens":2536,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:03:34.980034+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Pretrain a masked autoencoder and FAME on spectrally whitened EEG (per-band variance equalized), then recompute the embedding-distribution distances and per-band recoverability losses; if the low-frequency bias and the FAME downstream gap disappear, the input spectrum was the cause rather than the model's learning dynamics.","supporting_citations":[],"review_version":2}