{"id":"95fc66da-47ac-470a-9e7a-e9572bd45ef2","arxiv_id":"2504.17352","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"The Means Field classifier, using distances to several matrix power means in a linear discriminant, outperforms the Riemannian MDM and approaches tangent-space logistic regression on 20 EEG BCI datasets.","lead":"This paper proposes a new way to classify EEG brain-wave signals for brain-computer interfaces: instead of comparing each signal to one average pattern per class, it compares each signal to many average patterns and feeds those comparisons into a simple linear classifier. On 20 public datasets covering 587 people, the new classifier beats the standard Riemannian baseline and comes close to the current state of the art, with no fine-tuning of typical hyperparameters.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claim that MF is virtually hyperparameter-free rests on an unverified assertion in Section 4; if results are sensitive to the means field or RPME threshold, the reported advantage over MDM/MDMF may depend on specific settings.","rationale":"The reader's weakest assumption is exactly the same as the concern I identify: the paper's assertion of hyperparameter insensitivity is unsupported by any sensitivity analysis. I find no more fundamental flaw in the core empirical comparison. The benchmark uses MOABB with fixed folds, paired statistical tests, and large across-database aggregation, which gives reasonable support to the claim that MF outperforms MDM and MDMF under the chosen configurations. The provenance of several P300 databases from the authors' group is a legitimate external-validity concern, but it is not a correctness risk and does not change the verdict. Since the reader already issued CONDITIONAL, my stress-test does not alter the verdict; the concrete test would either confirm the tuning-free claim or require tightening it.","tokens_in":14124,"tokens_out":2965,"duration_ms":31221,"concrete_test":"Re-run the MI and P300 MOABB benchmarks with a reduced means field (e.g., 5 power means: h={-1,-0.5,0,0.5,1}) and with RPME zth set to 2.0 and 3.0, keeping all other pipeline choices identical; then recompute the per-database SMD and Liptak meta-effect against MDM and MDMF. If either variant shifts the MI meta-effect SMD by more than about 20% or causes the p<0.001 significance to disappear, the Section 4 insensitivity claim fails and the reported headlining gains are tied to the specific hyperparameter settings.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central value proposition is a deterministic, tuning-free classifier that beats MDM and MDMF. Section 4 acknowledges that MF introduces hyperparameters (number and position of power means, ADCSP filter counts, RPME threshold zth, tolerance and iteration limits) and asserts they are 'not critical' and that 'reasonable a priori choices give results that are statistically indistinguishable,' but no sensitivity analysis or supporting evidence is supplied. This matters because RPME itself already shows a significant effect in the paper (Figure 10: SMD=-0.297, p<0.001 for P300; Figure 6: SMD=-0.11, p<0.01 for MI), so at least one introduced choice materially changes performance. Without perturbing the means-field grid, the number of means, the ADCSP thresholds, and zth, it is unknown whether the headline meta-effects (SMD=1.122 and 1.057 versus MDM and MDMF) reflect a robust margin or a favorable configuration. The paper's own call for 'further research' on a universal choice of these hyperparameters further undercuts the 'virtually hyperparameter-free' framing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Riemannian Means Field (MF) classifier that represents each trial by the squared affine-invariant distances to a field of power means of the class covariance matrices and feeds these features to LDA. The method is compared against the standard Riemannian MDM, the authors' earlier MDMF, and tangent-space logistic regression (TS+LR) on 20 public MOABB databases (10 motor imagery, 10 P300; 587 subjects total), using within-session 5-fold cross-validation with identical folds, per-database permutation/Wilcoxon tests, and a Liptak meta-analysis. The paper also introduces an adaptive double CSP (ADCSP) spatial filter for MI data and a robust power-mean estimation (RPME) procedure. The headline results are that MF outperforms MDM and MDMF with large meta-effect sizes (SMD = 1.122 and 1.057 for MI; 0.522 and 0.681 for P300), while being slightly but significantly worse than TS+LR on both paradigms. The authors characterize MF as deterministic and 'virtually hyperparameter-free'.","tokens_in":14390,"tokens_out":3406,"duration_ms":34490,"significance":"If the empirical results hold, the MF classifier is a practically valuable contribution: it is simple, deterministic, fast when combined with spatial filtering, and substantially closes the gap between manifold-based Riemannian classifiers and tangent-space state-of-the-art methods. The study has notable strengths: it uses 20 public databases, keeps folds identical across pipelines, applies exact permutation tests where possible, combines evidence with a weighted Liptak meta-analysis, and reports effect sizes. The code is made available. The main weakness is that the central 'virtually hyperparameter-free' claim is not supported by sensitivity analysis, and at least one of the introduced choices (RPME) demonstrably changes performance.","major_comments":[{"comment":"The claim that the newly introduced hyperparameters (the 11 power means and their positions, the ADCSP dimension thresholds 28 and 10, the RPME outlier threshold zth=2.5, I=4, tolerance, and iteration limits) are 'not critical' and that 'reasonable a priori choices give results that are statistically indistinguishable' is unsupported. No sensitivity analysis is provided. This is load-bearing because the paper's value proposition is a tuning-free deterministic classifier. Moreover, RPME itself is shown to change results significantly (Figure 10: SMD = -0.297, p < 0.001 for P300; Figure 6: SMD = -0.11, p < 0.01 for MI), so at least one introduced choice materially affects performance. The paper should either provide a sensitivity analysis over the means-field grid, the number of means, the ADCSP thresholds, and zth, or substantially soften the 'virtually hyperparameter-free' framing. The sentence 'Further research should establish a convenient choice of these hyperparameters to be used universally' also concedes that the choice is not yet settled.","section":"Section 4 and Section 2.1"},{"comment":"The statement that MF 'significantly outperforms' MDM and MDMF 'for all databases taken individually' cannot be correct for databases with very few subjects. For example, Zhou2016 has only 4 subjects (Table 1), so the exact paired permutation test's minimum one-sided p-value is 1/16 = 0.0625, which exceeds 0.05. The authors should clarify what 'individually' means here: if they mean the effect direction is positive in every database, that wording should be used; if they mean per-database statistical significance, the claim is logically impossible for the small-n databases in this study.","section":"Section 3.1, Table 4, Figures 2-3"}],"minor_comments":[{"comment":"The abstract says 'our code will be released as open source,' but Section 5 says 'Our code is published at ...' — please align these statements.","section":"Abstract and Section 5"},{"comment":"The caption contains a typo: 'ACSTP' should be 'ADCSP'.","section":"Figure 4 caption"},{"comment":"The entry '1/8' for the number of sessions in Brain Invaders 2013a is unclear; please clarify whether this means one session for some subjects and eight for others, or something else.","section":"Table 2"},{"comment":"The statistical analysis would benefit from a statement about whether any correction for multiple comparisons was applied across the many pipeline pairs tested; as written, the reader cannot tell whether the reported p-values are corrected.","section":"Section 2.3"},{"comment":"The sentence 'This suggests that the P300 data included in MOABB needs to be curated' is presented without evidence; consider rephrasing to avoid implying a data-quality judgment solely from improved outlier rejection.","section":"Section 4, Discussion of RPME"}],"recommendation":"major_revision","confidential_remarks":"The empirical core is sound and the direction of the results is credible, but the unsupported 'hyperparameter-free' claim and the overstatement about per-database significance are central to the paper's messaging. Both can be fixed with additional analysis and careful rewording, so I see this as a major revision rather than a rejection. I also note that the authors are comparing against their own prior MDMF baseline; this is legitimate given the external MDM and TS+LR comparisons, but the framing of 'improvement over MDMF' should be read in that light."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one. The Means Field classifier is a real new piece: instead of taking the minimum distance to a family of power means (MDMF), it feeds the squared distances into an LDA. On 20 public EEG databases with identical folds and exact permutation/Wilcoxon tests, it beats MDM and MDMF with meta-effect SMDs around 1.0 for motor imagery and 0.5–0.7 for P300, and it lands close to tangent-space logistic regression. That is a genuinely useful result for BCI benchmarking, and the empirical work is careful: same folds, per-database tests, Liptak meta-analysis, external pyRiemann baseline, and public data. The code is on GitHub and they fixed a real initialization bug in pyRiemann. Credit where due.\n\nThe soft spot is the \"virtually hyperparameter-free\" claim. Section 4 says the new hyperparameters — number and position of the power means, ADCSP filter counts, RPME outlier threshold — are \"not critical\" and reasonable a priori choices give statistically indistinguishable results. No sensitivity analysis is provided. That matters because RPME itself shows significant effects in their own figures (P300 SMD -0.297, MI -0.11), so at least one introduced choice changes performance materially. Without perturbing the means grid or zth, you cannot know if the headline advantage over MDM/MDMF is a robust margin or a favorable configuration. The paper's own call for \"further research\" to choose these hyperparameters universally undercuts the tuning-free framing. A sensitivity analysis over a plausible range, or a damped claim, would fix it.\n\nThe six P300 Brain Invaders datasets come from the authors' own group; the comparison is still against public data and an external baseline, so I don't treat that as circularity, just a note for the referee to look at dataset provenance.\n\nBottom line: the central claim — MF beats MDM and MDMF and approaches TS+LR — holds up in the data presented. The main unanswered question is how much of the margin depends on the particular means field and outlier threshold. That's a modify-and-resubmit issue, not a reject. Send it to a serious referee; the BCI community will want this baseline either way.","headline":"A genuinely new Riemannian classifier with a careful 20-database benchmark and honest reporting; the main flaw is an unverified 'hyperparameter-free' claim that should be fixed by sensitivity analysis.","tokens_in":14926,"tokens_out":2589,"would_cite":true,"duration_ms":25028,"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":"A means field of power means, fed as squared distances to LDA, beats the standard MDM across 20 EEG-BCI databases while staying deterministic.","keywords":["brain-computer interface","EEG classification","Riemannian geometry","symmetric positive-definite matrices","power means","minimum distance to mean classifier","motor imagery","P300"],"falsifier":"Re-run the 20-database benchmark while varying the means field (e.g., 5, 11, or 21 power means, and shifted values of $h$) and the RPME threshold $z_{th}$ from 1.5 to 4, and check whether MF still outperforms MDM and MDMF by similar margins. If the advantage over MDM shrinks or reverses for some plausible settings, the central claim of a robust, tuning-free classifier is contradicted.","tokens_in":1757,"feed_emoji":"🧠","tokens_out":8735,"duration_ms":116210,"temperature":0.7,"pith_summary":"The paper introduces the Means Field (MF) classifier for EEG-based brain–computer interfaces. Instead of classifying a trial by its distance to one geometric mean per class, as the standard MDM does, MF computes the squared Riemannian distance from the trial to each of a family of power means spanning the harmonic-to-arithmetic range, and feeds this vector of distances to a linear discriminant analysis. The authors claim that on a benchmark of 20 public databases (587 subjects, motor imagery and P300), MF clearly outperforms both MDM and the previous MDMF and comes close to the tangent-space logistic regression state of the art, while remaining deterministic and effectively free of hyperparameters. If right, this gives BCI practitioners a simple, reproducible classifier that needs no tuning and no tangent-space projection.","feed_headline":"Means field classifier beats the standard MDM on 20 BCI datasets","feed_subtitle":"By feeding squared distances to many power means into a plain LDA, it rivals tangent-space methods with no tuning.","key_machinery":"The central object is the means field: a family of power means $\\mathcal{P}_h$ of SPD covariance matrices for exponents $h \\in [-1,1]$, interpolating between the harmonic mean ($h=-1$) and arithmetic mean ($h=1$), with the geometric mean as the $h\\to 0$ limit. The MF classifier takes the squared Riemannian distance from each trial to every mean in the field and passes that distance vector to LDA, which learns a linear combination of distances; this is what distinguishes it from MDM and MDMF, which each use only one or the minimum distance. Two supporting mechanisms carry the result: RPME iteratively removes trials whose standardized geometric distance exceeds a threshold, making mean estimates robust, and ADCSP reduces high-dimensional motor-imagery covariance matrices in two stages with a CSP-like joint diagonalization.","core_discovery":"On the paper's own terms, the central discovery is that the soft information contained in a whole field of distances carries more classification signal than the single nearest-mean distance. The MF classifier estimates eleven power means per class, for exponents $h \\in \\{\\pm 1, \\pm 0.75, \\pm 0.5, \\pm 0.25, \\pm 0.1, 0\\}$, builds a feature vector from the squared affine-invariant distances of each trial to all of those means, and classifies that vector with LDA. Across 20 public EEG-BCI databases the authors report that this pipeline, combined with an adaptive double common spatial pattern filter for motor imagery and robust power-mean estimation for both paradigms, outperforms MDM and MDMF significantly on both paradigms, with meta-analytic effect sizes SMD = 1.122 and 1.057 for motor imagery and 0.522 and 0.681 for P300, all with p < 0.001. It remains slightly but significantly behind tangent-space logistic regression, with meta-effect sizes of -0.28 and -0.498, but the gap is small. The paper presents the result as evidence that manifold-native, deterministic classifiers can rival tangent-space methods.","pith_inferences":["Editorial inference: because LDA has a closed-form solution and the means are fixed once the exponents are chosen, the same means-field idea could replace the geometric mean in any MDM-based pipeline, including transfer-learning and calibration-free settings, at low computational cost.","Editorial inference: the means field need not be restricted to power means; other SPD mean families such as log-Euclidean or alpha-divergence means could enrich the distance vector, and the clustering variant suggested in the discussion could improve granularity.","Editorial inference: the larger RPME gains on P300 suggest that outlier handling matters most in less controlled recording conditions, so applying MF+RPME to out-of-the-lab data is a natural testable extension."],"forward_implications":["For motor imagery, MF with ADCSP outperforms MDM and MDMF on every database individually and overall, with meta-effect SMD = 1.122 and 1.057, both with p < 0.001.","For P300, MF outperforms MDM and MDMF overall, with meta-effect SMD = 0.522 and 0.681, both with p < 0.001, though TS+LR remains ahead in seven of ten databases.","MF approaches but does not surpass tangent-space logistic regression: TS+LR wins six of ten motor-imagery databases and seven of ten P300 databases.","RPME improves MF significantly on both paradigms, more so for P300, and ADCSP both improves accuracy and reduces runtime relative to standard CSP.","A fully deterministic, essentially tuning-free classifier working directly on the manifold can compete with state-of-the-art tangent-space methods."],"supporting_citations":[{"why":"Defines the benchmark and preprocessing used here and establishes that tangent-space classifiers are the state of the art the paper measures against.","marker":"[1]"},{"why":"Introduces Riemannian geometry for BCI classification, including the MDM and tangent-space logistic regression pipelines used as baselines.","marker":"[5]"},{"why":"Provides the mathematical definition of matrix power means that the means field is built on.","marker":"[15]"},{"why":"Supplies the fixed-point algorithm used to compute the power means and the earlier evidence that the optimal power-mean exponent varies per subject.","marker":"[19]"},{"why":"Describes the previous MDMF classifier, the main baseline the paper improves upon.","marker":"[20]"},{"why":"Defines the benchmarking and statistical framework, including cross-validation, permutation tests, and meta-analytic combination, used for the comparisons.","marker":"[22]"},{"why":"Provides the software implementation of covariance estimation, spatial filtering, and Riemannian operations used in the experiments.","marker":"[23]"},{"why":"Introduces the CSP spatial filter that the proposed ADCSP adapts for motor-imagery data.","marker":"[24]"},{"why":"Introduces the spatial filter used for P300 data in all compared pipelines.","marker":"[25]"},{"why":"Supplies the outlier-rejection strategy adapted for the robust power-mean estimation.","marker":"[27]"}],"fun_headline_variants":["Means field classifier beats MDM on 20 EEG-BCI datasets","Power means field rivals state-of-the-art BCI decoders","Riemannian means field: better than MDM, close to SOTA","New classifier: many means beat one for EEG-BCI","Means field: 20 datasets show gain over MDM for BCI"],"cache_read_input_tokens":17024,"weakest_assumption_plain":"The paper's claim that the new settings (the 11 power means and their positions, the ADCSP dimension thresholds, and the RPME outlier threshold) are not critical is asserted without sensitivity analysis, so the load-bearing premise is that these reasonable a priori choices do not drive the reported gains.","fun_headline_variants_meta":{"raw":{"variants":["Means field classifier beats MDM on 20 EEG-BCI datasets","Power means field rivals state-of-the-art BCI decoders","Riemannian means field: better than MDM, close to SOTA","New classifier: many means beat one for EEG-BCI","Means field: 20 datasets show gain over MDM for BCI"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000238,"raw_usage":{"total_tokens":1525,"prompt_tokens":973,"completion_tokens":552,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":589,"completion_tokens_details":{"reasoning_tokens":459}},"tokens_in":589,"tokens_out":552,"duration_ms":5734,"temperature":1.0,"reasoning_tokens":459,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:42:52.674815+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the 20-database benchmark while varying the means field (e.g., 5, 11, or 21 power means, and shifted values of $h$) and the RPME threshold $z_{th}$ from 1.5 to 4, and check whether MF still outperforms MDM and MDMF by similar margins. If the advantage over MDM shrinks or reverses for some plausible settings, the central claim of a robust, tuning-free classifier is contradicted.","supporting_citations":[{"cited_title":"Multiclass brain-computer interface classification by Riemannian geometry","cited_arxiv_id":null,"evidence_quote":"Introduces Riemannian geometry for BCI classification, including the MDM and tangent-space logistic regression pipelines used as baselines."},{"cited_title":"Fixed Point Algorithms for Estimating Power Means of Positive Definite Matrices","cited_arxiv_id":null,"evidence_quote":"Supplies the fixed-point algorithm used to compute the power means and the earlier evidence that the optimal power-mean exponent varies per subject."},{"cited_title":"The Riemannian Minimum Distance to Means Field Classifier","cited_arxiv_id":null,"evidence_quote":"Describes the previous MDMF classifier, the main baseline the paper improves upon."},{"cited_title":"Optimizing Spatial filters for Robust EEG Single-Trial Analysis","cited_arxiv_id":null,"evidence_quote":"Introduces the CSP spatial filter that the proposed ADCSP adapts for motor-imagery data."},{"cited_title":"Xdawn algorithm to enhance evoked potentials: Application to brain-computer interface","cited_arxiv_id":null,"evidence_quote":"Introduces the spatial filter used for P300 data in all compared pipelines."},{"cited_title":"The Riemannian Potato Field: A Tool for Online Signal Quality Index of EEG","cited_arxiv_id":null,"evidence_quote":"Supplies the outlier-rejection strategy adapted for the robust power-mean estimation."}],"review_version":1}