acceptodds
Withdrawn from ICLR 2027

SIFR: Does Session-Invariant Factorized Representation Improve Cross-Session EEG Identity Recognition?

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Cross-session EEG identity recognition faces a direct tension: a model has to keep the stable cues that distinguish one person from another, while still adapting to changes caused by headset re-wearing, acquisition conditions, and task state. We propose SIFR, a Session-Invariant Factorized Representation framework that turns this problem into controlled residual compensation inside the normalization layer. Its core module, FactorNorm, adds an input-conditioned session branch on top of shared affine parameters. A low-rank gate adjusts the compensation dimension by dimension, while zero-residual initialization keeps the starting point exactly aligned with LayerNorm. Session classification supervision gives the compensation branch an information target, and the Orthogonal Prototype Constraint (OPC) limits directional overlap with the current identity prototype. Under explicit assumptions, we derive a linear response bound for this geometric constraint and show how it limits perturbations along a specified direction. Across session-wise records on PhysioNet, MCD, and SEED-IV, SIFR reaches mean identification accuracies of 99.28%, 82.14%, and 72.34%, improving over the strongest external mean-ACC baseline on each dataset by 0.21, 4.85, and 3.06 percentage points, respectively. Ablation and identity-parameter freezing experiments further examine the roles of gating, geometric regularization, and restricted parameter updates. The results show that organizing cross-session representation learning around explicit compensation can improve identification accuracy while also making genuine and impostor identity claims more separable in score space; the component analysis further exposes the roles played by the gate and the geometric constraint.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG identity recognition, cross-session generalization, factorized normalization, adaptive gating, orthogonal prototype constraint

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_whgXOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates the problem of cross-session EEG identification, aiming to improve identification performance on EEG signals across different sessions. The authors propose a framework called the "Session-Invariant Factorized Representation Framework (SIFR)," which mainly replaces the standard LayerNorm with the proposed FactorNorm module in the last two blocks of the Transformer. The FactorNorm module encodes session-related information through a residual supervised by session classification, while an orthogonal prototype constraint (ORC) restricts the directional overlap between the session compensation and the current identity prototype. SIFR is evaluated on the cross-session identification task.

Critical strengths
  1. The motivation of this study is clear, and the problem is well positioned.
  2. The proposed method is specifically designed for the cross-session EEG identification problem, compensating for session variations while preserving identity information, and further restricting the overlap between the compensation direction and the current identity prototype direction through an orthogonal prototype constraint.
Critical weaknesses
  1. The logic linking the problem and the proposed solution is unclear. For example, why is session residual compensation a reasonable approach for cross-session EEG identification? What advantages does this scheme offer compared with mainstream methods such as global alignment?
  2. The manuscript is poorly readable. The method description is overly brief and lacks details; there are undefined symbols; and the experimental settings are unclear.
  3. The baselines are insufficiently current and representative. For example, data augmentation-based baselines are missing, and the Transformer-based baselines are not sufficiently up to date.
Questions for authors
  1. Please supplement the methodological details and the definitions of missing symbols; for example, γid\gamma_{id}, LfL_f, and whether the representation added to the residual has already been normalized.
  2. How sensitive is SIFR to different parameters, such as the representation dimension, the number of stacked layers, and the loss weights?
  3. The proposed method mainly replaces LayerNorm in the Transformer architecture with FactorNorm. Given the strong performance of the Transformer architecture itself on sequence structure, the authors should compare with state-of-the-art Transformer-based baselines.
  4. The reporting of the ablation experiments is somewhat coarse. For example, the radar chart in Figure 2(a) includes the metrics "robustness," "stability," and "generalization." How are these metrics defined and quantified? What is the relationship between Figure 2(c) and the proposed method? Why does Figure 2(g) include only 7 subjects rather than all of them? Selectively reporting experimental results raises concerns about the reliability of the method.
  5. The conclusions drawn on PhysioNet are almost unusable, as different [methods?] have approached saturation. In this case, treating PhysioNet as one of the three main datasets and conducting ablations based on it provides very little information.
  6. There is a lack of references to and introduction of the experimental results. For example, does "session" in Table 5 refer to the training set or the test set?
  7. There are spelling errors, for example, on line 377.
Reviewer_SyWEOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper aims to improve the cross-session performance of EEG identity recognition. They core idea is to produce a "Session-Invariant Factorized Representation (SIFR)", which involves trainable FactorNorm modules that replace the normalization in the final two blocks of a backbone transformer network. In addition to the identification task-loss, these modules are trained via two extra loss-terms for session classification and an "orthogonal prototype constraint" (which seems unclear). Both of these auxiliary losses are related to intermediate representations extracted from these FactorNorm modules. They conducted experiments that compared their method against 15 baselines over three evaluation datasets and performed ablation studies.

Critical strengths

Experimental results seem to suggest that SIFR is superior when compared to 15 other baseline methods across three evaluation datasets. Their method is the bolded best row (including one tie) across 27 columns of evaluation metrics over Tables 1, 2, and 3. They claim to have performed three independent trials per evaluation point, and have reported the mean and standard deviations.

Caveat: see the weaknesses, where I raise concerns about the statistical significance and even validity of the results, which seem to completely undermine these strengths.

Critical weaknesses

The presentation of the core methodology, throughout Section 3, is very poor, with many cases of undefined notation (across every equation) and generally unclear descriptions. The overall presentation is so unclear that any architectural design contributions of their proposal are completely obscured and reproducibility is impossible. Please see the itemized list further below.

Across all of the results tables (1-3), even though SIFR row is always bolded as the best, the advantage (difference in means) is often small compared to the reported standard deviations. For example, in the first column of Table 1, SIFR has an accuracy advantage over TSCeption of just 0.02, but the standard deviations of those cells are 1.26 and 0.31. In addition to providing these large tables of numbers (with 16*27 = 432 cells overall), the authors should also perform statistical analysis to check whether SIFR has statistically significant advantage. Given a rough eyeballing of the tables and seeing how many advantages that are within one std deviation, I would guess that significance is not established PhysioNet (table 1), due to that benchmark being saturated. However, there may be a significant advantage for the other two benchmarks (tables 2 and 3). Regardless, it should not have to be the job of the reviewer nor reader to parse and process 864 (= 432 * 2) numbers from these tables to check for statistical significance.

Some of the values in Table 1 appear to show impossible statistics. For DGCNN, the S2/Acc and S2/F1 are both reported as "99.36 ± 1.85" and for TSCeption, the S2/Acc is reported as "99.36 ± 1.35". Assuming (since Accuracy and F1 Score are calculated as percentages) that each sample must be between 0 and 100, the largest sample standard deviation (for N=3 trials, and dividing by (N-1)=2) possible (while being consistent with the mean of 99.36) is only approximately 1.11, which is realized when the three samples are 100, 100, and 98.08. Thus, these three cells appear to be mathematically impossible, which raises a serious concern about the accuracy and validity of their experimental results.

Table 1 also seems (frankly speaking) suspiciously unlikely. For this PhysioNet evaluation set, it appears that many methods nearly saturate the benchmark, with accuracy and F1 score often close to 100% and equal error rate close to 0%. Since several methods (including SIFR) perform similarly well (close to perfect), it should be harder to statistically distinguish them. The reported means between SIFR and several other methods are very small compared to the reported standard deviations. If these values are reasonably accurate, then it's quite remarkably lucky that SIFR just happened to take the best (slightly better) bolded value in each column. That is, one would expect that at least some of the closer sample mean advantages (across the many close pairwise comparisons) would have easily fallen the other way. At best, the authors may have had a bug in their code that led to inaccurately high standard deviations. The observation in the previous paragraph (of impossibly high std dev) would also be consistent with such a hypothesis.

Itemized list of notation and presentation issues

Note that this is not an exhaustive list. Instead, I'm just noting some issues to illustrate how prevalent that undefined notation and lack of clarity completely undermine the presentation of the paper.

In equation 1, μn\mu_n, vnv_n, and ϵ\epsilon are undefined notation. Of course, the structure of the equation can lead one to guess that these must be the feature-wise mean and variance of each token, and that ϵ>0\epsilon > 0 is small value for numerical stability, however, it's just sloppy to not define symbols that are used, and this is the mildest case of undefined notation.

In equation 2, W1W_1, W2W_2, b1b_1, and b2b_2 are undefined notation, and it is also sloppy to not clarify what stacking γsess\gamma_\text{sess} and βsess\beta_\text{sess} should mean in terms of how the dimensions of the right-hand-side of the equation are assigned to those two variables. Based on the overall context, one could guess that there must be 2D2D total dimensions and they must be split equally, but this should be made clear. Further, a critical issue with the undefined notation is that, even though one could guess that W1W_1 and W2W_2 are meant to be weight matrices and b1b_1 and b2b_2 are bias vectors, by not clarifying their dimensions, the intended low-rank nature of this operation is only indirectly inferred by the sentence following equation 2 that states "A second low-rank path" when referring to the next equation. Thus, one could guess that W1W_1 should have dimensions r×Dr \times D and W2W_2 should be 2D×r2D \times r, where rr is some (undefined and unclarified) rank hyper-parameter that should be smaller that DD? Is this correct? What is the setting for this hyper-parameter in the experiments?

In equation 3, bgb_g is undefined. Is this a bias vector, a scalar, or something else? What value of qq is used in the experiments? Does it coincide with the rank rr above?

In equation 4, γid\gamma_\text{id} and βid\beta_\text{id} are undefined. What are these set to in the experiments? Are they trainable parameters or constants?

In equation 5, yny_n is defined in the expression, but this quantity is not clarified with a name or purpose (i.e., it is not mentioned again in other equations). How is yny_n later used in the rest of the framework?

In equation 6, the meaning of the notation NlN_l and dlnd_{ln} are undefined and unclear. Note that while dnd_n is defined in equation 5, it is unclear what exactly it means to add another subscript for the expression dlnd_{ln}. It appears that the intention is to express how the earlier descriptions of Section 3.3 is meant to describe the behavior of one FactorNorm module, and that now in Section 3.4, this is dealing with "four FactorNorm modules", so ll appears to be indexing across these 4 modules. However, then why is the subscript nn (which presumably may be intended to be over the NN tokens) now averaged over NlN_l indices. What is NlN_l?

In equation 7, BB, zsess,iz_{\text{sess}, i}, and paip_{a_i} are undefined. While the similar zsessz_{sess} was previously defined, it is unclear what adding the extra subscript ii is meant to index, or why there is an apparent mean across BB values. Is this meant to denote batch size? Similarly, pkp_k was defined, but it becomes unclear what replacing that subscript with aia_i means? How is aia_i defined?

In equation 8 (and the immediately preceding text), the notation aa, ee, and η\eta are not defined. What are these meant to represent? Why is it given that e⊥p=0e^\perp p = 0?

In equation 9, the loss terms LidL_\text{id} and LsessL_\text{sess} (note: these are actually "\mathcal{L}" in the paper, but it seems that this review form that does not support the mathcal font) are not clearly defined, nor is it clear how everything ultimately feeds into these terms. Figure 1 seems to imply that LidL_\text{id} is related to "ArcFace" as an "Identity head", but the details are unclear, and LsessL_\text{sess} seems to be a cross-entropy loss, however, it is unclear how the "Residual pool x4" goes through a "Session head" and is ultimately used as part of a cross-entropy loss. The "Session head" does not appear to be mentioned or further clarified across the methodology section 3, nor is it clear how this "Residual pool x4" is formed.

Other relatively minor presentation issues (again, not an exhaustive list)

On lines 78-80 (in Section 2.1), the text "However, discriminative feature extraction does not by itself guarantee stability across sessions," is repeated twice in a row.

One lines 262 and 263, there are spaces missing after commas in "recognition,component" and "ACC,macro-F1".

Figures 1 and 2 both have small text that is difficult to read in a printed copy. Figure 2 has some very small text that is also light grey over the white background of the page (in the top-left section). This is completely unreadable on a printed page, and other small text in that Figure is just barely readable. A mitigating factor is that both Figures seem to be vector graphics, so viewing them on a computer screen and zooming in a lot will make them legible. However, another possibly related issue if that maybe the Figures (especially Figure 2) is not so efficiently encoded. When viewing this PDF on my (reasonably capable) laptop, the PDF rendering even becomes a bit sluggish and unresponsive when trying to zoom and pan around on this Figure. Overall, these issues make it quite inconvenient when trying to extract information from these Figures.

Further, Figure 2 is really multiple things thrown together. I recommend breaking it up into separate figures, each presented in larger format (to aid readability), and perhaps moving some into the appendix

On line 377 (end of page 7), there is a broken reference to "Appendix ??". It seems that this should point to Appendix B.2.

In line 632 (Appendix A, before equation 11), there is an inline equation "w=ap+e" that is in plain text mode, but should instead be put in math mode.

An additional caption below Table 7, states "Source: parameter counts from the existing implementation. This table measures model size, not runtime." Both of these statements are unnecessary and seem like the sort of gibberish that is common in AI-generated text.

Questions for authors

Please see the weaknesses box for specific issues about unclear notation and descriptions. Could you please clarify these issues?

Could you please clarify the overall set of trainable parameters for the FactorNorm modules?

It seems that the FactorNorm modules involves some hyperparameters, such as the rank value q mentioned in Equation 3. What was this and other hyperparamters set to for the experiments and how were these chosen?

Is the apparent advantage of SIFR statistically significant compared to the baselines?

Could you please check the reported results in light of the validity concerns that I raise above?

Reviewer_fBjtOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces a Session-Invariant Factorized Representation (SIFR) framework for cross-session EEG identity recognition. The backbone is a Transformer with a FactorNorm that consists of a shared affine path plus a gated residual in the last two blocks. The paper provides numerical results on three datasets (PhysioNet, MCD, and SEED-IV) and compares SIFR against 15 EEG decoding models.

Critical strengths

None

Critical weaknesses

The paper lacks important information:

  • The Transformer backbone (cited as Vaswani et al.) is not described; later, the authors refer to a CLS token and a final LayerNorm. This does not align with the original encoder–decoder Transformer, but rather with a Vision Transformer–style model. The position of the LayerNorm is important, since the main contribution is its replacement by the proposed FactorNorm.
  • The preprocessing needs more explanation, in particular “Feature prep.” and “Windowing”.
  • The MCD dataset is not cited.
  • For the experiments in Sec. 4, the data splits need more explanation (e.g., sample counts, training/test split).

Evaluation:

  • Test data leaks into early stopping (lines 293 ff. and 702 ff.).
  • SEED-IV: Sliding windows are extracted with a stride of 1 (line 702), so overlapping windows can end up in both the training and test data (lines 294 and 464).
  • PhysioNet: Trials are grouped into pseudo-sessions, which does not support the cross-session claim.
  • MCD: Memory and Reading share the same physical recording.
Questions for authors

Can the authors provide a detailed description of the Transformer backbone and the preprocessing? Can the authors comment on the evaluation of the model?

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

Related papers

Discussion (0)

Sign in to comment.