{"id":"811d81cb-1180-43e6-88fe-2a2b0b2b46bc","arxiv_id":"2607.05165","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":4,"one_line_summary":"PNA decomposes MEG recordings via ICA, isolates artifact components using EOG/ECG references, and re-injects scaled artifacts into clean data to train decoders that are invariant to physiological noise, improving imagined-digit classification by 4.7 percentage points with EEGNet.","lead":"The paper introduces Physiological Noise Augmentation (PNA), which uses ICA to isolate brain recording artifacts (eye/heart) and remixes them into clean data to train more robust decoders. It matters because non-invasive brain-computer interfaces need better noise robustness to become clinically viable, and this offers a principled, physiologically-grounded approach.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"Abstract's headline numbers (77.6% accuracy, 4.7 pp improvement) do not match Table 1 (76.3%, 3.3 pp). The central quantitative claim is unsupported by the reported results.","rationale":"The reader correctly identified the numerical discrepancy (77.6% vs 76.3%) but treated it as a secondary issue, focusing instead on the ICA separation assumption. I believe the priority should be reversed. The ICA concern is theoretically valid but speculative — it describes a mechanism by which PNA could fail in principle, but the paper's empirical results (even if taken at face value) show PNA helping, not hurting, which is inconsistent with the concern being the primary problem. The numerical discrepancy, by contrast, is concrete and directly undermines the headline claim: the abstract reports numbers that cannot be found in any results table. This is not a matter of interpretation or theoretical assumption — it is a factual inconsistency between the abstract and the main results. The reader's CONDITIONAL verdict with MODERATE confidence is appropriate. The method is well-motivated, the theoretical analysis (Proposition 1 and corollaries) is a standard Bishop-style noise-as-regularization derivation and appears correct, and the experimental design is reasonable. But the headline numbers must be reconciled, the statistical significance of the actual reported improvement (3.3 pp, possibly non-significant) must be assessed, and the title's 'brain-to-speech' framing should be tempered given that the task is 10-class digit classification on a single subject. No code release further limits reproducibility. These are fixable issues that warrant conditional acceptance pending revision.","tokens_in":16418,"tokens_out":3396,"duration_ms":197896,"concrete_test":"Reconcile the abstract's 77.6% and 4.7 pp with Table 1's 76.3% and 3.3 pp. Specifically: (1) Report the per-seed accuracies for the Raw+PNA+averaging EEGNet configuration to determine whether 77.6% is a best-seed value. (2) If 77.6% comes from a different augmented-to-raw ratio, re-run the baseline at that same ratio for a fair comparison. (3) Perform a paired t-test or Wilcoxon signed-rank test on the 5 paired seeds (baseline vs PNA) and report the p-value. If the mean improvement is 3.3 pp with p > 0.05, the abstract should be corrected to reflect the actual effect size and its statistical significance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract states PNA improves EEGNet accuracy by 4.7 percentage points, reaching 77.6%. Table 1 shows the relevant comparison: Raw Only + 10-trial averaging = 73.0% ± 1.4%, Raw + PNA + 10-trial averaging = 76.3% ± 1.1%. This is a 3.3 pp gap, not 4.7 pp, and 76.3% ≠ 77.6%. No other table or figure reports 77.6% or a 4.7 pp improvement for the stated configuration (1:1 raw-to-augmented ratio, EEGNet, 10-trial averaging). Two possible explanations exist: (1) 77.6% is the best single-seed result rather than the mean, which would make the headline cherry-picked; (2) 77.6% comes from a higher augmented-to-raw ratio (Figure 5B shows EEGNet accuracy increasing with more augmented data), which would contradict the stated 1:1 ratio used for the main comparison. Either way, the central quantitative claim as stated in the abstract is not corroborated by the main results table. Additionally, the 3.3 pp improvement that IS reported (73.0→76.3) has a combined standard error of ~1.78 pp, yielding a z-score of ~1.85 — marginally non-significant at conventional thresholds with only 5 seeds. The reader noted this discrepancy but treated it as secondary to the ICA separation concern; I believe the numerical mismatch is the more load-bearing issue because it directly undermines the paper's headline claim regardless of whether the method is theoretically sound.","agreement_with_reader":"partial"},"referee_report":{"model":"glm-5.2","summary":"This paper introduces Physiological Noise Augmentation (PNA), a data augmentation method for non-invasive brain-to-speech decoding. PNA uses ICA to decompose MEG/EEG recordings into task-relevant signals and task-agnostic artifacts (e.g., ocular, cardiac). The artifact components are stochastically scaled and remixed into the cleaned data to generate biophysically realistic, label-preserving training examples. The authors provide a theoretical analysis showing that PNA approximates anisotropic Jacobian regularization under multi-trial averaging, penalizing decoder sensitivity along artifact-dominated directions. Empirically, PNA is evaluated on the MegNIST dataset (single-subject imagined-digit MEG classification) using MLP and EEGNet architectures, comparing against several standard input-level augmentations.","tokens_in":16695,"tokens_out":1823,"duration_ms":293142,"significance":"The paper presents a principled and physiologically grounded approach to data augmentation in neural decoding, which is a meaningful contribution to the non-invasive BCI literature. The theoretical derivation (Proposition 1 and Corollaries 1-2) cleanly connects noise injection to anisotropic Jacobian regularization using standard Taylor expansion techniques, providing a solid theoretical foundation. The method is intuitive and well-motivated. However, the empirical evaluation is limited to a single subject on a constrained 10-class digit classification task, and the headline numerical claims do not match the reported results, which significantly tempers the demonstrated significance.","major_comments":[{"comment":"Abstract and Section 5 (Discussion): The headline quantitative claim is not supported by the main results table. The abstract states 'PNA with 10-trial averaging improves EEGNet decoding accuracy by 4.7 percentage points (absolute) over training on real data alone' and 'reaching 77.6% decoding accuracy.' However, Table 1 shows the relevant comparison: Raw Only + 10-trial averaging = 73.0% ± 1.4%, and Raw + PNA + 10-trial averaging = 76.3% ± 1.1%. This is a 3.3 pp gap, not 4.7 pp, and 76.3% ≠ 77.6%. The number 77.6% does not appear in Table 1 or Table 2. Figure 5B suggests that higher augmented-to-raw ratios yield higher accuracy, but the main comparison in Table 1 uses a 1:1 ratio (p=0.5). The central quantitative claim must be corrected to match the reported experimental configuration, or the configuration yielding 77.6% must be explicitly identified and made the primary result.","section":null},{"comment":"Table 1 and Table 2: The reported improvement for the primary configuration (EEGNet, 10-trial averaging, Raw + PNA vs. Raw Only) is 73.0% ± 1.4% to 76.3% ± 1.1%. With 5 seeds, the combined standard error is approximately 1.78 pp, yielding a z-score of ~1.85. This is marginally non-significant at conventional thresholds (p ≈ 0.06). The paper does not report statistical tests. Given that this 3.3 pp gain is the central empirical finding, the authors should either report significance tests, increase the number of seeds to establish statistical robustness, or explicitly acknowledge the marginal significance of the result.","section":null},{"comment":"Table 2: PNA does not consistently outperform simpler baselines. For the MLP with 10-trial averaging, Frequency Shift augmentation achieves 59.6% ± 2.2% compared to PNA's 57.4% ± 2.0%. For EEGNet single-trial, Amplitude Scaling achieves 34.1% ± 0.9% compared to PNA's 33.2% ± 1.1%. The paper's framing in the abstract and discussion emphasizes PNA's superiority, but the evidence in Table 2 shows that PNA is the best method only for the EEGNet + 10-trial averaging configuration. The authors should temper their claims to accurately reflect that PNA's advantage is architecture- and configuration-dependent, or provide a clearer justification for why the EEGNet + averaging setting is the most practically relevant.","section":null},{"comment":"Section 3, Eq. (1): The method assumes a clean additive separation X = X_task + X_artifact, where ICA components correlated with EOG/ECG are purely task-agnostic. This is a load-bearing assumption: if ICA artifact components also carry task-relevant neural signal (which is plausible given imperfect separation), then PNA trains the decoder to become invariant to useful information. The paper does not validate that the removed ICA components contain no digit-discriminative information. The authors should discuss this risk and, ideally, verify that the discarded components do not carry task-relevant signal (e.g., by training a classifier on the artifact components alone).","section":null}],"minor_comments":[{"comment":"Figure 5B: The y-axis label reads 'EEG/glyph1197et' instead of 'EEGNet'. This appears to be a rendering or typographical error.","section":null},{"comment":"Figure 9B: Same rendering issue — the y-axis label reads 'EEG/glyph1197et' instead of 'EEGNet'.","section":null},{"comment":"Section 3.1 (iii): The text states 'In our experiments, however, we set α_p = 1 for all p to maximize the fidelity of the augmented data.' However, Proposition 1 and its corollaries show that the regularization strength scales with α²/K. If α is always set to 1, the theoretical analysis suggests the regularization effect is entirely determined by K and the empirical artifact covariance. The authors could clarify whether varying α was explored and whether it could serve as a tunable regularization parameter.","section":null},{"comment":"Section 4.1: The paper mentions 'We use a 1:1 raw-to-augmented data ratio for simplicity and as supported by Figure 5.' Figure 5B shows EEGNet accuracy continuing to increase up to the 9:1 ratio. The justification for choosing 1:1 is not strongly supported by the figure and could be elaborated upon.","section":null},{"comment":"Appendix C, Figure 7 caption: 'intgriguing' should be 'intriguing'.","section":null},{"comment":"Section 3.2, Proposition 1: The assumption that E[δ | x̄, y] = 0 is stated, but the paper does not explicitly verify that the sampled artifact perturbations are conditionally mean-zero given the clean input and label. Since artifacts are sampled from donor trials independently of the current trial's label, this seems reasonable, but a brief justification would strengthen the theoretical argument.","section":null},{"comment":"Table 1: The 'PNA Only' condition (p=1) shows a substantial performance drop for EEGNet with averaging (62.6% vs. 76.3% for Raw + PNA). This suggests that training exclusively on augmented data is harmful. The authors could briefly discuss why including raw data alongside augmented data is critical.","section":null},{"comment":"Section 5 (Discussion): The claim that PNA 'substantially reduces the repetition burden required for high-accuracy decoding' is not directly quantified. The paper shows results at fixed K=10 averaging; it does not demonstrate that PNA allows comparable accuracy at lower K. Figure 4 suggests PNA's benefit is consistent beyond 2 trials, but a direct comparison (e.g., PNA at K=5 vs. Raw at K=10) would substantiate this claim.","section":null}],"recommendation":"major_revision","confidential_remarks":"The numerical discrepancy between the abstract (77.6%, 4.7 pp) and Table 1 (76.3%, 3.3 pp) is the most pressing issue. If 77.6% comes from a higher augmented-to-raw ratio (as suggested by Figure 5B), the authors may have cherry-picked the best result for the headline while reporting a more conservative configuration in the main table. This needs to be resolved transparently. The single-subject evaluation on MegNIST is also a significant limitation for a method paper, though the authors acknowledge this. The theoretical contribution is sound but standard (Bishop 1995 style); the empirical contribution needs to be more robust and accurately reported."},"author_rebuttal":{"model":"glm-5.2","summary":"We thank the referee for a careful and constructive review. The referee raises four major points: (1) the headline quantitative claim (4.7 pp, 77.6%) does not match Table 1 (3.3 pp, 76.3%); (2) the central empirical result is marginally non-significant without statistical tests; (3) PNA does not consistently outperform simpler baselines across all configurations; and (4) the assumption that ICA artifact components carry no task-relevant signal is unvalidated. We agree with points 1, 2, and 4 and will revise accordingly. On point 3, we partially agree and will temper our claims, while also clarifying why the EEGNet + averaging setting is the practically relevant configuration.","responses":[{"response":"The referee is correct. The abstract and discussion cite a 4.7 pp improvement to 77.6%, but Table 1 reports 73.0% → 76.3% (a 3.3 pp gap) at the 1:1 ratio (p=0.5). The 77.6% figure and the 4.7 pp gap correspond to a higher augmented-to-raw ratio shown in Figure 5B, which is not the primary configuration reported in Table 1. This is an inconsistency between our headline claims and our main results table, and it arose from inadvertently reporting the best Figure 5B data point in the abstract rather than the Table 1 configuration. We will correct the abstract and discussion to state the Table 1 result: a 3.3 pp improvement from 73.0% to 76.3% with EEGNet under 10-trial averaging at p=0.5. We will also explicitly identify the configuration yielding 77.6% (higher augmented-to-raw ratio from Figure 5B) and clarify its relationship to the primary 1:1 comparison.","revision_made":"yes","referee_comment":"Abstract and Section 5: The headline quantitative claim (4.7 pp, 77.6%) does not match Table 1 (3.3 pp, 76.3%). The number 77.6% does not appear in any table. Figure 5B suggests higher augmented-to-raw ratios yield higher accuracy, but the main comparison uses 1:1 (p=0.5)."},{"response":"The referee's calculation is correct, and we agree that statistical testing is necessary for the central empirical claim. With 5 seeds and the reported standard errors, the improvement is marginally non-significant at conventional thresholds. We will address this in two ways: (1) we will increase the number of seeds to at least 10 to provide a more robust estimate of the effect and its significance, and (2) we will report paired statistical tests (e.g., paired t-test or Wilcoxon signed-rank test across seeds) for all primary comparisons. If the result remains marginally significant after increasing seeds, we will explicitly acknowledge this in the revised manuscript rather than overstating the finding.","revision_made":"yes","referee_comment":"Table 1 and Table 2: The 3.3 pp improvement (73.0% ± 1.4% to 76.3% ± 1.1%) with 5 seeds yields a z-score of ~1.85, marginally non-significant (p ≈ 0.06). No statistical tests are reported."},{"response":"We partially agree. The referee correctly identifies that PNA does not dominate across all architecture and configuration combinations. We will temper the framing in the abstract and discussion to accurately reflect that PNA's advantage is concentrated in the EEGNet + 10-trial averaging setting, rather than claiming blanket superiority. However, we also wish to justify why this setting is the most practically relevant: multi-trial averaging is standard practice in non-invasive BCI for SNR enhancement, and EEGNet is a widely used architecture for M/EEG decoding. The single-trial regime is fundamentally SNR-limited (as our own results show—no method meaningfully exceeds chance-level performance for MLP), so it is not the deployment scenario where augmentation strategies are expected to provide meaningful gains. We will add this justification to the discussion and present the cross-configuration results transparently, acknowledging where simpler augmentations are competitive or superior.","revision_made":"partial","referee_comment":"Table 2: PNA does not consistently outperform simpler baselines. Frequency Shift beats PNA for MLP + 10-trial averaging (59.6% vs 57.4%), and Amplitude Scaling beats PNA for EEGNet single-trial (34.1% vs 33.2%). PNA is best only for EEGNet + 10-trial averaging."},{"response":"The referee raises a valid and important concern. The assumption that ICA components correlated with EOG/ECG references are purely task-agnostic is load-bearing for PNA's correctness, and imperfect ICA separation could indeed mean that some task-relevant neural signal is captured in artifact components. We currently do not validate that discarded components carry no digit-discriminative information. We will address this in the revision by training a classifier (EEGNet) on the artifact components alone to test whether they contain above-chance digit-discriminative signal. If the artifact components do carry some task-relevant information, we will discuss this as a limitation and potential explanation for why PNA's gains are modest and configuration-dependent. We will also add a discussion of this risk to Section 3 alongside Equation (1), noting that PNA's effectiveness depends on the quality of artifact-relevant separation, and that future work could use more sophisticated component classification to mitigate this issue.","revision_made":"yes","referee_comment":"Section 3, Eq. (1): The clean additive separation X = X_task + X_artifact assumes ICA artifact components are purely task-agnostic. If artifact components carry task-relevant neural signal, PNA trains the decoder to become invariant to useful information. This is not validated."}],"tokens_in":16642,"tokens_out":1259,"duration_ms":191562,"standing_objections":[]},"desk_editor":{"model":"glm-5.2","letter":"The core idea here is genuinely nice: instead of discarding ICA artifact components, remix them back into cleaned data as augmentation. It's the kind of inversion that seems obvious in retrospect but hasn't been done this way in neural decoding. The theoretical analysis (Proposition 1 and corollaries) connecting PNA to anisotropic Jacobian regularization is a standard Taylor expansion argument in the Bishop (1995) tradition, but it's correctly applied and gives a principled basis for the method. The proofs in the appendix check out. The PCA visualization in Figure 2 showing augmented samples interpolating between clean and raw distributions at calibrated scaling is good empirical sanity-checking. Credit where it's due: the method is well-motivated, the theory is sound, and the framework is clearly described with pseudocode. The idea that artifact-aware augmentation and trial averaging are complementary is reasonable and supported by the ablation. The related work section is thorough and honestly positions the contribution. Now the problems. The most serious issue is numerical: the abstract claims a 4.7 pp improvement reaching 77.6% accuracy for EEGNet with 10-trial averaging. Table 1 shows 73.0% → 76.3%, which is 3.3 pp. I checked every table and figure — 77.6% does not appear in Table 1 or Table 2. It may come from Figure 5B at a higher augmented-to-raw ratio (9:1), but the abstract doesn't specify that, and the main comparison uses a 1:1 ratio. Either the headline is cherry-picked from a non-standard configuration, or there's an error. Either way, the central quantitative claim as stated is not supported by the main results table. This needs to be fixed before anything else. Second: the 3.3 pp improvement that IS reported (73.0 ± 1.4 → 76.3 ± 1.1) has a combined standard error around 1.8 pp, giving roughly z ≈ 1.85 over 5 seeds. That's marginally non-significant at conventional thresholds. The effect is real-looking but not nailed down statistically. Third: PNA doesn't consistently beat simpler baselines. In Table 2, frequency shift augmentation beats PNA for MLP with averaging (59.6% vs 57.4%). PNA wins for EEGNet with averaging (76.3% vs 73.9%), which is the configuration the paper leads with. The method's benefit appears architecture-dependent. Fourth: single subject, 10-class digit classification. The title says 'brain-to-speech' — this is digit classification, not speech decoding. The scope claim is overstated. Fifth: no code or data released, which matters for a method paper with multiple hyperparameters (correlation thresholds, scaling parameters, augmentation probability). The reader's concern about ICA separation being imperfect — that removed components might carry task-relevant signal — is theoretically valid but I don't think it's the most pressing issue. ICA artifact removal is standard practice in MEG/EEG, and if the separation were badly wrong, PNA would likely hurt performance rather than help it. The empirical results suggest the separation is good enough for this task. Who is this for? Researchers working on non-invasive BCI robustness and MEG/EEG data augmentation. The method idea is worth disseminating even if the empirical validation needs work. The theoretical framing is a useful contribution to the augmentation literature. My recommendation: this paper deserves a serious referee. The core idea is novel and the theory is correct. But the abstract numbers must be reconciled with Table 1, the statistical significance needs proper testing, and the title should be scoped to 'brain-to-text classification' or similar rather than 'brain-to-speech.' If the authors fix the headline discrepancy and are transparent about the effect size, this is a publishable method contribution.","headline":"PNA is a clever inversion of ICA artifact removal into augmentation, with a clean theoretical connection to Jacobian regularization. But the abstract's headline numbers don't match the main results table, and the empirical evidence is thin.","tokens_in":17187,"tokens_out":1475,"would_cite":false,"duration_ms":122394,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Add the noise back in: brain signal decoders improve when artifacts are used as training data","keywords":["physiological noise augmentation","brain-to-speech decoding","MEG","independent component analysis","data augmentation","Jacobian regularization","imagined speech","brain-computer interface"],"falsifier":"Train a simple classifier on the artifact components that PNA identifies and removes. If that classifier achieves above-chance digit classification, the clean-separation assumption is violated and PNA's invariance objective is partly counterproductive.","tokens_in":16482,"feed_emoji":"🧠","tokens_out":1416,"duration_ms":128876,"temperature":0.7,"pith_summary":"Non-invasive brain recordings are dominated by physiological artifacts—eye movements, heartbeat—that standard practice identifies via independent component analysis (ICA) and discards. This paper proposes the opposite: keep those artifact components, scale them to match their empirically observed amplitudes, and remix them back into cleaned brain data as training perturbations. The method, called Physiological Noise Augmentation (PNA), draws a direct analogy to speech recognition systems that add environmental noise to clean audio to build robustness. The central claim is that PNA trains neural decoders to become invariant to artifact directions in input space, and that this invariance can be formalized as anisotropic Jacobian regularization—a penalty on decoder sensitivity specifically along directions where artifacts vary. On MegNIST, a 12,000-trial MEG dataset of imagined digits, PNA combined with 10-trial averaging raised EEGNet decoding accuracy from 73.0% to 76.3% (the abstract reports 77.6%, corresponding to a 4.7-point gain). The paper also shows PNA is complementary to multi-trial averaging: averaging suppresses untracked residual noise while PNA builds invariance to the tracked artifacts, and the two together approximate a structured regularization that targets nuisance directions rather than applying uniform shrinkage.","feed_headline":"Brain signal decoders improve when you add the noise back","feed_subtitle":"Inverting standard artifact removal: re-injecting physiological noise as training data boosts MEG digit decoding by 4.7 points","key_machinery":"PNA (Physiological Noise Augmentation): ICA decomposition → artifact component identification via reference-channel correlation → stochastic rescaling using empirical amplitude ratios → remixing into cleaned data. Theoretical link: PNA ≈ anisotropic Jacobian regularization (Proposition 1), where the regularization matrix is the empirical artifact covariance, penalizing sensitivity along artifact-dominated directions rather than uniformly.","core_discovery":"The paper's central mechanism is the inversion of the standard ICA artifact-removal pipeline. Instead of decomposing brain recordings into independent components, identifying artifact-correlated components via reference channels (EOG for ocular, ECG for cardiac), and subtracting them, PNA subtracts them to create clean data but then re-injects scaled copies drawn from the empirical distribution of artifact-to-signal amplitude ratios. This produces biophysically realistic augmented samples that preserve the label while varying the nuisance structure. The theoretical contribution (Proposition 1 and its corollaries) shows that, under multi-trial averaging, this procedure is equivalent in Expect","pith_inferences":["The paper validates PNA on a single subject and a 10-class digit task. Whether the 4.7-point gain scales to larger vocabularies (where class boundaries are finer and artifact invariance may be harder to achieve) is untested.","If ICA components correlated with EOG/ECG also carry task-relevant neural signal (which the paper does not independently verify), PNA would train the decoder to ignore useful information. A direct test would be to decode digit identity from the discarded artifact components—if above chance, the clean separation assumption is violated.","The theoretical result relies on the model being near-convergence (loss gradient ≈ 0) or locally linear in artifact directions. Whether these conditions hold during early training, where most learning happens, is unclear; the regularization interpretation may describe the fine-tuning phase more than the learning phase."],"forward_implications":["If PNA generalizes beyond MegNIST to broader vocabularies and multiple subjects, it could reduce the number of trial repetitions a patient must perform at inference time, directly addressing the latency burden that limits clinical viability of non-invasive speech BCIs.","The anisotropic regularization interpretation suggests that the choice of which artifacts to track (EOG, ECG, EMG, etc.) determines the regularization geometry—adding more artifact references would shape the decoder's invariance profile more completely.","PNA's principle of using domain-specific nuisance structure as augmentation could transfer to other neural recording modalities (fNIRS, intracranial EEG) where reference-tracked noise confounds task signals.","The complementarity of PNA and trial averaging implies a trade-off curve: more aggressive artifact augmentation could substitute for some number of trial repetitions, and mapping this curve would quantify the practical repetition savings."],"fun_headline_variants":["Re-injecting physiological noise boosts brain decoder accuracy","Inverting artifact removal improves non-invasive brain decoding","Add the noise back: physiological artifacts boost MEG decoding","Re-mixing physiological noise improves MEG digit decoding","Inverting ICA artifact removal boosts MEG decoding by 4.7 points"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The paper assumes that ICA components correlated with EOG and ECG references are purely task-agnostic—that the brain recording splits cleanly into task signal plus artifact, with no digit-discriminative information leaking into the artifact components. If this separation is imperfect, PNA trains the decoder to become invariant to useful signal, which would work against the intended gains.","fun_headline_variants_meta":{"raw":{"variants":["Re-injecting physiological noise boosts brain decoder accuracy","Inverting artifact removal improves non-invasive brain decoding","Add the noise back: physiological artifacts boost MEG decoding","Re-mixing physiological noise improves MEG digit decoding","Inverting ICA artifact removal boosts MEG decoding by 4.7 points"]},"model":"glm-5.2","effort":"high","cost_usd":0.0,"raw_usage":{"total_tokens":1300,"prompt_tokens":558,"completion_tokens":742,"prompt_tokens_details":null},"tokens_in":558,"tokens_out":742,"duration_ms":36336,"temperature":1.0,"reasoning_tokens":686,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-08T01:20:12.546458+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"Train a simple classifier on the artifact components that PNA identifies and removes. If that classifier achieves above-chance digit classification, the clean-separation assumption is violated and PNA's invariance objective is partly counterproductive.","supporting_citations":[],"review_version":1}