acceptodds
Under review as a conference paper at ICLR 2027

Representation, Normalization, and Sequential Calibration for Cross-Subject EEG-to-Image Retrieval

Abstract

EEG-to-image retrieval is commonly formulated as a cross-modal representation-learning problem, but cross-subject deployment introduces an additional train-test mismatch: subject-specific normalization statistics available during source training are unavailable for an unseen subject at cold start. We present the Representation-Normalization-Sequential Calibration (RNSC) framework, which treats shared representation learning, source-subject normalization, and unseen-subject calibration as coordinated stages of a single retrieval space. RNSC couples a shared, subject-agnostic visual reference with an EEG encoder that factorizes temporal structure before channel-semantic mixing and aggregates repeated trials after latent encoding. During source training, EEG latents are optimized in subject-standardized coordinates; at deployment, target moments are estimated sequentially from a finite source prior and past unlabeled target EEG under a strict predict-before-update protocol. On THINGS-EEG2, the shared representation achieves state-of-the-art intra-subject 200-way retrieval, with 93.67% Top-1 and 99.36% Top-5, while the complete RNSC framework reaches 48.10% Top-1 and 76.34% Top-5 under inter-subject evaluation. Controlled ablations, latent-geometry diagnostics, and cold-start analyses disentangle the complementary roles of representation learning, training-time normalization, and sequential target-statistic estimation.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG-to-image retrieval, cross-subject EEG, multimodal representation learning, brain-computer interfaces

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_PRNLOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2(+1)
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper proposes RNSC for cross-subject EEG-to-image retrieval. The framework combines a multi-level visual reference (UMVA), an EEG encoder (STAR-EEG), training-time subject normalization (TC-SAW), and sequential test-time calibration (SPPC). The main idea is to learn EEG representations in subject-standardized coordinates and estimate the corresponding statistics for unseen subjects using source statistics and past unlabeled target EEG. On THINGS-EEG2, the authors report 93.67% intra-subject Top-1 and 48.10% inter-subject Top-1 under 200-way retrieval. The paper also provides component ablations, latent geometry analysis, and cold-start experiments.

While the unified coordinate-system view could be useful for this research area, and the paper reports very high intra- and inter-subject retrieval accuracy, several important concerns remain regarding evaluation validity, performance attribution, and scientific novelty. These issues substantially weaken the evidence supporting the main claims and the overall contribution of the paper.

Critical strengths
  1. The component ablations are reasonably well structured. Table 5 uses a 2×2 comparison of TC-SAW and SPPC under the same STAR-EEG representation. The authors also report paired results across multiple runs. This is useful for separating the effects of training-time normalization and deployment-time calibration, rather than only comparing the full system against previous methods. (Sec. 4.5, Table 5, Appendix Tables 13 and 15)

  2. The cold-start analysis provides some useful insights. Figure 3 and Appendix Table 21 show how the benefit of the source prior changes as target observations accumulate. The authors also examine different prior strengths and report that an overly strong prior can hurt performance. These analyses help explain when sequential calibration is useful and where its gains become limited. (Sec. 4.6, Fig. 3, Appendix Tables 21–22)

Critical weaknesses
  1. Checkpoint selection and evaluation validity. The supplementary implementation appears to select the best checkpoint using Top-1 accuracy on the final test set. If this is the code used for the reported experiments, it would raise a substantial concern about test-set independence, especially for the leave-one-subject-out evaluation. It also seems inconsistent with the model selection description in Sec. 4.1. If the reported results were obtained using this procedure, the performance estimates may be optimistic. The SPPC ablation could also be affected, since the selected checkpoints are evaluated under different calibration settings. This issue needs to be clarified before the main results can be considered reliable. (Sec. 4.1, Tables 1–2 and 5, Appendix Table 13, supplementary implementation)

  2. The reported performance gains are not clearly attributed to the proposed components. Table 3 shows that EEGProject with raw-mean aggregation already achieves 92.05% intra-subject Top-1. Under the same encoded-mean aggregation, STAR-EEG improves Top-1 from 92.81% to 93.67%, a gain of only 0.86 percentage points. Similarly, Table 5 shows that the complete TC-SAW+SPPC configuration improves inter-subject Top-1 from 46.00% to 48.10% compared with the configuration without TC-SAW and using History-only calibration. These improvements are positive, but they do not clearly explain the much larger performance gap against published inter-subject methods, such as SAMGA (Jiang et al., 2026) in Table 2. Since the previous methods follow their original protocols, I think stronger matched comparisons are needed to clarify where the reported improvements actually come from. (Sec. 4.3–4.5, Tables 1–3 and 5)

  3. The motivation and additional scientific insight are not sufficiently convincing. In the Introduction, the authors describe two main directions of prior work: representation learning and handling subject variation. However, it is not clearly explained why treating these two aspects separately is insufficient, or what specific problem requires them to be coordinated in a unified coordinate system. The distinction between the two directions is also somewhat simplified. For example, SAMGA involves both multi-level visual supervision and subject-aware visual routing, rather than fitting only into the second direction. More importantly, although the proposed combination is reasonable, Table 5 mainly supports the individual benefits of TC-SAW and SPPC, with limited evidence for a distinct coordination effect. I think the paper needs a clearer motivation and stronger scientific insight beyond combining existing techniques. (Sec. 1–3, Table 5)

  4. The cold-start evaluation protocol and robustness need further clarification. The main evaluation uses repeated-trial aggregation, while sequential calibration operates over 200 concept-level queries. The exact preprocessing information available from the unseen subject is not sufficiently clear, especially whether target-subject statistics are used before the first query. Also, the large early SPPC gains are reported under a fixed query order, while the overall gain over 200 queries is only about 1.2 percentage points. More evidence is needed to understand how robust these results are under different query orders and initialization conditions. (Sec. 4.1, Sec. 4.6, Fig. 3, Appendix Table 6 and Table 21, supplementary preprocessing)

Questions for authors
  1. Could the authors clarify how the final checkpoints were selected? The supplementary implementation appears to use test Top-1 accuracy for checkpoint selection. Were the main results obtained using this procedure? Could the authors provide results using an independent validation protocol without using the final test labels for model selection?

  2. What mainly explains the very high retrieval accuracy? Table 3 shows that STAR-EEG improves only 0.86 percentage points over EEGProject under the same encoded-mean aggregation. Could the authors better separate the contributions of the visual representation, EEG encoder, and training protocol? More closely matched baselines would help clarify the large differences against previous methods.

  3. Why is a unified coordinate system needed, and what new insight does it provide? The Introduction describes representation learning and handling subject variation as two main directions of prior work, but it is not clear how this motivates the need for a unified coordinate system. Could the authors further explain what specific limitations arise when these two aspects are handled separately, why the proposed coordination is necessary, and what new scientific insight it provides beyond combining existing techniques?

  4. Could the authors clarify the cold-start information budget and robustness? How many EEG repetitions are used per query, and are target-subject statistics available during preprocessing before the first query? Also, do the SPPC gains remain stable under different query orders and a fixed Source-only normalization baseline?

Reference

Jiang, L., She, Q., Xu, J., Xu, H., Wu, D., and Kuang, Z. (2026). Subject-Aware Multi-Granularity Alignment for Zero-Shot EEG-to-Image Retrieval. arXiv preprint arXiv:2604.17782.

9 comments

We appreciate the reviewer’s detailed scrutiny of checkpoint selection, component attribution, and the cold-start information budget. These are important questions for evaluating the validity and interpretation of our results. We address each point below through explicit protocol descriptions, empirical comparisons, and their interpretation.

Q1. Checkpoint selection and evaluation validity

1.Could the authors clarify how the final checkpoints were selected? The supplementary implementation appears to use test Top-1 accuracy for checkpoint selection. Were the main results obtained using this procedure? Could the authors provide results using an independent validation protocol without using the final test labels for model selection?

We agree that independent checkpoint selection is important for evaluation validity. The original checkpoints were selected using target-test Top-1; the 27-study protocol audit and common-epoch comparisons quantify and explain the reasons for and effects of checkpoint selection.

(1) Checkpoint practices in 27 related studies. Based on reported methods and publicly available implementations:

Publicly verifiable protocol evidenceNumberShare
Code paths involving test metrics in checkpoint selection or best-epoch tracking829.6%
Independent checkpoint selection in the main evaluation path622.2%
Insufficiently verifiable or inapplicable1348.1%

Examples: NICE selects by validation loss; SATTC by source-validation Top-5; public EEGiT cross-subject code reuses its test loader for validation; Shallow Alignment tracks test loss/Top-1; SAMGA reports its final retrieval metrics on the test set.

(2) Validation mismatch. THINGS-EEG2 has 16,540 training images (4 EEG repetitions each) versus 200 test images (80 repetitions each). Our encoded-mean aggregator encodes repetitions individually before averaging, and test concepts also differ. Source-only validation is independent but does not exactly match the final 80-repetition, 200-way test condition.

(3) Checkpoint audit and fixed-epoch evaluation. We audited 300 run–fold records. The original best.pt checkpoints were selected by held-out target-test Top-1. To quantify performance under a common training budget, we report epoch 50 for intra-subject retrieval and epoch 20 for inter-subject LOSO:

ExperimentOriginal best Top-1 (%)Common epochTop-1 / Top-5 at that epoch (%)
STAR-EEG, intra, encoded-mean93.675091.94 / 99.42
STAR-EEG, LOSO, TC-SAW on48.102044.21 / 75.93
STAR-EEG, LOSO, TC-SAW off46.902043.25 / 75.58

Calibration alternatives within each run/fold share one checkpoint. Epochs 50/20 were chosen after inspecting historical trajectories; these are descriptive common-epoch comparisons, not prospectively selected independent-validation results. Separate 60-epoch intra-subject runs are labeled accordingly.

In summary, the 27-study audit documents protocol heterogeneity; the 4-vs-80 repetition mismatch explains the gap between source-validation and test conditions; and the fixed-epoch comparison demonstrates the model's performance under a common training budget. We will state the checkpoint rule once in the evaluation protocol and label both sets of reported results accordingly.

Q2. Attribution of the high retrieval accuracy

2.What mainly explains the very high retrieval accuracy? Table 3 shows that STAR-EEG improves only 0.86 percentage points over EEGProject under the same encoded-mean aggregation. Could the authors better separate the contributions of the visual representation, EEG encoder, and training protocol? More closely matched baselines would help clarify the large differences against previous methods.

As summarized in the General Response, visual representation quality contributes materially to retrieval performance alongside calibration. The epoch-50 intra-subject visual comparisons (66Jr Q3(b)) yield 92.15% UMVA versus 89.93% parameter-matched independent projection and 86.02% single L28. For EEG encoders at epoch 50, STAR-EEG = 91.94% and EEGProject = 90.92%.

At epoch 50, encoded-mean yields 91.94% versus 90.93% for raw-mean. Table 5 separately reports 48.10% for full RNSC versus 46.00% for no TC-SAW + History-only, under the original test-selected checkpoint rule.

The separate 60-fold Alljoined LOSO comparison (mirK AQ(a)) yields SPPC 16.025% versus History-only 15.433% Top-1 (+0.592 pp) at the same checkpoints selected using target-test Top-1. This provides additional deployment-calibration evidence under the reported selection rule; PRNL Q1 describes the checkpoint-selection protocol. The source-side normalization strategy comparison in PRNL Q3 further examines the training component of the framework.

Q3. Why a unified coordinate system is necessary and what scientific insight it provides

3.Why is a unified coordinate system needed, and what new insight does it provide? The Introduction describes representation learning and handling subject variation as two main directions of prior work, but it is not clear how this motivates the need for a unified coordinate system. Could the authors further explain what specific limitations arise when these two aspects are handled separately, why the proposed coordination is necessary, and what new scientific insight it provides beyond combining existing techniques?

The need for a unified coordinate system arises because source representation learning and unseen-subject calibration act on the same EEG–visual retrieval geometry. TC-SAW standardizes trial-level EEG latents using source-subject statistics inside contrastive training. The shared encoder is thus optimized to preserve stimulus-discriminative information under reduced subject-specific statistical variation, rather than relying solely on subject-dependent latent moments.

At deployment, an unseen subject has no directly available matching latent statistics. SPPC combines the source prior with strictly past unlabeled target queries, estimating the subject's normalization coordinates under predict-before-update, without updating the encoder or accessing future queries.

Crucially, SPPC does not invert TC-SAW or restore raw features; it maps target latents into coordinates compatible with those used to train the retrieval model. Deployment-only calibration cannot change an encoder learned without training-time standardization, whereas training-time standardization alone cannot supply an unseen subject's latent statistics. This is the concrete limitation addressed by coordinating the two stages.

This emphasis on sequentially available statistics also complements the batch-oriented structural expert/PoE discussed in SATTC’s own Limitations (Huang and Zhu, 2026): RNSC makes each prediction before incorporating the current query into the calibration state. Consistent with their complementary roles, Table 5 shows 46.90% → 48.10% Top-1 when adding TC-SAW under the same SPPC setting, and 46.89% → 48.10% when replacing History-only with SPPC under TC-SAW training.

As a separate-dataset deployment check, Alljoined-1.6M LOSO yields +0.592 pp SPPC-over-History-only across 60 folds at shared, test-selected best checkpoints (mirK AQ(a)). This provides additional evidence of deployment calibration on a second dataset; the checkpoint-selection protocol is addressed in PRNL Q1.

Alternative training-time normalization strategies. To test whether the underlying principle extends beyond one source-statistics estimator, we compared None, Source-subject (the paper's TC-SAW implementation), and Global (one pooled set of statistics from all nine source subjects in each batch) during training. At evaluation, all three retain the same online diagonal SAW/SPPC (κ=8), encoded-mean queries, and cosine retrieval. With five seeds × ten LOSO target subjects, the common epoch-20 results are:

Training-time normalizationTop-1 (%)Top-5 (%)
None43.2575.58
Source-subject (TC-SAW)44.2175.93
Global (pooled source statistics)45.3976.47

Global improves Top-1 over None by +2.14 pp and over Source-subject by +1.18 pp, with subject-level paired 95% CIs of [+0.65, +3.63] pp and [+0.02, +2.34] pp, respectively. Together, these results show that the choice of source-side normalization coordinates matters, and that training-time latent standardization has substantial room for further development. In particular, the strong pooled-global result suggests that the RNSC perspective can accommodate a family of training-time estimators, motivating hierarchical and shrinkage-based alternatives alongside subject-conditioned statistics. The comparison uses shared subjects, seeds, and evaluation settings at epoch 20.

The paired seed-99 analysis (see ieJB Q2) further distinguishes raw scale changes from stimulus-relevant geometry: the same-stimulus versus different-stimulus cosine margin rises from 0.0111 to 0.4661 after online SAW/SPPC. This measures the combined deployment transformation, not TC-SAW alone. The scientific insight is that training-time latent normalization shapes what the encoder learns, while causal deployment calibration determines how an unseen subject enters that learned standardized retrieval space. We will articulate this specific training–deployment connection in the Introduction.

Q4. Cold-start information budget, preprocessing, query order, and Source-only baseline

4.Could the authors clarify the cold-start information budget and robustness? How many EEG repetitions are used per query, and are target-subject statistics available during preprocessing before the first query? Also, do the SPPC gains remain stable under different query orders and a fixed Source-only normalization baseline?

We distinguish three information sources: (i) the 80 EEG repetitions of the current image, encoded individually and averaged into a single query; (ii) upstream session-wise MVNN whitening from the UBP preprocessing pipeline used in our experiments, whose whitening matrix is fitted on each participant's training EEG partition only (sigma_tot = sigma_part[1]) and then applied to that participant's training and test EEG; and (iii) SPPC's sequential, unlabeled target-test-query latent history.

Importantly, the UBP whitening matrix is not fitted using target test EEG covariance or test labels. For the held-out participant, the upstream whitening can use that participant's training-split EEG as preprocessing data, but the learned encoder is trained only on source subjects. This preprocessing does not contradict our methodological contribution: TC-SAW operates on learned EEG latents during source training, whereas SPPC calibrates latent moments using only past target-test queries.

Before the first retrieval query, SPPC initializes from the source prior and has no target-test-query latent history; each query is retrieved before its latent representation updates the online state. Separately, upstream input whitening uses the held-out participant’s training EEG partition as described above.

Our earlier five random orderings × five training seeds × ten held-out subjects provide a matched-checkpoint analysis of causal calibration:

Query regimeHistory-only Top-1SPPC Top-1Difference
Original order, all 20046.8948.10+1.21 pp
Random orders, first 815.8023.80+8.00 pp
Random orders, first 1629.9031.35+1.45 pp

Early-history composition and order. We additionally compare three configurations of the first 32 target queries, reporting the Top-1 gain of SPPC over History-only at different query horizons:

First-32-query configurationFirst 8 gainFirst 32 gainAll 200 gain
Original set, original order+18.00 pp+5.81 pp+1.21 pp
Same set, shuffled internally+15.50 pp+5.31 pp+1.13 pp
Random set+13.00 pp+1.56 pp+0.30 pp

Retaining the same first-32 set while shuffling its internal order largely preserves the gain (+5.81 → +5.31 pp over the first 32; +1.21 → +1.13 pp over all 200). Changing the early set itself has a larger effect (+1.56 pp over the first 32; +0.30 pp over all 200). These results suggest that which observations enter the early calibration history matters, beyond history length alone. The two analyses probe different order/composition perturbations and their prefix gains should not be pooled.

A fixed Source-only normalization reference obtains 20.42% Top-1 over 200 queries, regardless of order. For the earlier five-permutation analysis, the first-eight-query paired, subject-level 95% CI is [+3.33, +12.67] pp. Overall, these comparisons identify causal cold-start calibration as SPPC's main strength and show that its effectiveness depends on the information contained in the early target history, not only the number of accumulated queries.

A fixed source prior can be especially useful before reliable target statistics emerge, while the optimal balance may evolve as evidence accumulates. These observations also motivate two future directions: reliability-aware online moment fusion, which adjusts source-versus-target weighting according to the uncertainty and stability of sequential target estimates; and adaptive switching, which progressively shifts from source-prior-dominated to target-history-dominated calibration when the available evidence becomes sufficiently reliable. Both would retain the predict-before-update information constraint; they are prospective extensions rather than mechanisms evaluated in this work.

Reviewer_66JrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes RNSC, a framework for cross-subject EEG-to-image retrieval that integrates representation learning, subject-wise normalization, and sequential calibration. Specifically, STAR-EEG learns EEG representations through shared temporal projection and post-encoding repetition aggregation, UMVA combines multi-level InternViT features, TC-SAW performs subject-wise standardization during training, and SPPC uses source-domain priors and historical target observations for online calibration. Experiments on THINGS-EEG2 demonstrate strong intra-subject and inter-subject retrieval performance.

Critical strengths
  1. The paper addresses the important challenge of cross-subject distribution shifts in EEG-based visual decoding. The strict predict-before-update protocol provides a meaningful setting for evaluating online calibration.
  2. The proposed system achieves competitive performance, reaching 93.67% and 48.10% Top-1 accuracy in intra-subject and inter-subject retrieval, respectively.
Critical weaknesses
  1. Limited methodological novelty. The proposed components largely build upon established techniques. STAR-EEG employs temporal projection and feature-space aggregation, UMVA relies on multi-level visual feature fusion, TC-SAW uses feature standardization, and SPPC adopts existing statistical moment-merging operations. Although their integration is reasonable, the paper does not sufficiently demonstrate a distinctive methodological contribution beyond combining these techniques.
  2. Insufficient attribution of performance improvements. Table 5 shows that the baseline STAR-EEG system with History-only calibration already achieves 46.00% Top-1 accuracy, while the complete RNSC framework reaches 48.10%. Thus, the normalization and calibration components contribute only 2.10 percentage points, despite the much larger improvement over external baselines reported in Table 2. This raises concerns that the overall performance advantage mainly comes from stronger representations rather than the proposed calibration framework. Moreover, comparisons with previously reported results under different experimental settings make it difficult to isolate these contributions.
  3. Inadequate comparisons for sequential calibration. SPPC is primarily compared against History-only calibration, which cannot reliably estimate target variance during the earliest queries. Consequently, the substantial cold-start improvements may partly reflect the advantage of source initialization rather than the proposed moment-merging strategy. Comparisons with stronger online normalization approaches, such as source-initialized EMA or other source-target statistical fusion methods, are needed to establish the independent value of SPPC.
  4. Insufficient validation of UMVA. The feature-masking and Shapley analyses in Table 4 demonstrate that the trained model utilizes multiple visual layers, but do not establish that the proposed fusion strategy outperforms simpler or parameter-matched alternatives. The contribution of UMVA remains difficult to distinguish from the benefits of using a stronger pretrained visual backbone.
Questions for authors
  1. What are the fundamental methodological differences between TC-SAW/SPPC and existing normalization or test-time adaptation approaches?
  2. Given the relatively small improvement from 46.00% to 48.10% in Table 5, how do the authors justify attributing the overall performance gains to the proposed normalization and calibration framework?
  3. Can SPPC outperform stronger causal calibration baselines, and can UMVA demonstrate advantages over appropriately controlled visual representation alternatives?
5 comments

We thank the reviewer for engaging with the prequential evaluation setting and for challenging us to clarify both the methodological novelty and the sources of performance improvement. Your questions prompted a clearer account of STAR-EEG’s design and additional evidence on visual fusion and causal calibration.

Q1. Methodological novelty

1.What are the fundamental methodological differences between TC-SAW/SPPC and existing normalization or test-time adaptation approaches?

STAR-EEG is motivated by two EEG representation challenges: stimulus-related temporal patterns may be reusable across electrodes despite channel-specific responses, and repeated recordings contain trial-specific variability. Its channel-shared temporal basis separates temporal feature extraction from channel-semantic mixing, while post-encoding repetition aggregation combines independently encoded trials in latent space rather than averaging sensor signals before nonlinear encoding.

These structural inductive biases enable STAR-EEG to capture reusable temporal patterns, preserve channel-specific semantics, and aggregate repeated observations in latent space. The Alljoined intra-subject comparison (mirK AQ(a)) further supports the aggregation choice: at epoch 20, encoded-mean improves over raw-mean by +4.56 pp Top-1 / +5.83 pp Top-5 across four seeds × 20 subjects, providing direct evidence for the aggregation pathway. Separately, its 60-fold LOSO results provide deployment-side context: SPPC improves over History-only by +0.592 pp Top-1 at shared best checkpoints (see mirK AQ(a)); this additional result supports the deployment calibration motivation.

Please see the General Response (Core contribution and novelty). The central claim is the training–deployment coordination: TC-SAW optimizes source-side representations under subject-normalized coordinates, while SPPC estimates compatible statistics from a source prior and past unlabeled target queries, with the encoder fixed and the current query processed before the online update. The distinctive contribution is the joint design of training-time representation coordinates and sequential deployment-time calibration. The normalization-strategy comparison (see PRNL Q3) demonstrates that this view naturally extends to different source-side normalization rules, opening a broader design space for training-time standardization.

How does RNSC differ from existing normalization/TTA methods? Source-only normalization does not provide a continuously updated estimator for unseen subjects, while deployment-only normalization/TTA does not train the encoder under the corresponding subject-standardized contrastive geometry. RNSC jointly designs representation learning and causal deployment calibration within a compatible retrieval space.

MethodTraining stageDeployment mechanism
Source-only normalizationSource-statistics standardizationNo updates from target-query history
Deployment-only normalization/TTAAdjusts representations after encoder trainingDoes not shape source-side contrastive learning
SATTC: SAW+CSLSSAW calibration and retrievalOne-time unlabeled calibration followed by frozen inference
SATTC: Structural expert + PoESimilarity-structure refinementMore naturally suited to batch refinement
RNSC: TC-SAW + SPPCTrial-level latent standardization during source-side contrastive trainingFrozen encoder; source prior + past unlabeled queries; predict-before-update

TC-SAW shapes the standardized representation geometry during training, while SPPC causally estimates compatible target coordinates from source priors and past unlabeled queries. Together, they connect training and deployment without updating model weights or accessing future target queries.

Q2. Attribution of performance improvements

2.Given the relatively small improvement from 46.00% to 48.10% in Table 5, how do the authors justify attributing the overall performance gains to the proposed normalization and calibration framework?

Table 5 reports 46.00% → 48.10% Top-1 (+2.10 pp) when moving from no TC-SAW + History-only to full TC-SAW + SPPC at the reported best checkpoints. This quantifies the combined training–deployment contribution. The matched-checkpoint causal calibration comparison (see ieJB Q1) separately identifies the deployment estimator’s contribution.

For encoded-mean intra-subject retrieval at epoch 50, EEGProject = 90.92% and STAR-EEG = 91.94%. On Alljoined (mirK AQ(a)), holding the STAR-EEG architecture and reporting at the same epoch 20, encoded-mean exceeds raw-mean (28.94% vs 24.38% Top-1; four seeds × 20 subjects). This supports latent-space repetition aggregation as a design choice. The visual-fusion comparisons appear in 66Jr Q3(b) (below); checkpoint-selection details are discussed in PRNL Q1.

Q3(a). Stronger causal calibration baselines

3.a Can SPPC outperform stronger causal calibration baselines?

Direct comparison with stronger causal calibration baselines. Under identical checkpoints, EEG query embeddings, candidate galleries, and sequential information access, we evaluated the following deployment estimators across five seeds × ten held-out subjects:

Causal calibration methodTop-1 (%)
Source-only (fixed source statistics)20.42
History-only46.89
Source-initialized EMA (α = 0.10)43.67
Fixed source–target interpolation (source weight = 0.50)35.77
SPPC (κ = 8)48.10

Under the evaluated configurations and shared checkpoints, SPPC achieves the highest Top-1 among these causal calibration baselines, improving over History-only by +1.21 pp. A target-test-swept interpolation weight can approach SPPC in full-sequence accuracy; since that weight is selected using target-test outcomes, we report it as sensitivity analysis (see ieJB Q1).

Cold start. Across five randomized query orders, SPPC improves Top-1 over History-only by +8.00 pp for the first eight queries. The magnitude of its full 200-query gain varies with query order (see PRNL Q4), highlighting the importance of early-history composition for online calibration.

Q3(b). Controlled alternatives for UMVA

3.b Can UMVA demonstrate advantages over appropriately controlled visual representation alternatives?

We compared four visual targets for 10 intra-subject participants × three seeds at the same epoch 50:

Visual representationTop-1 (%)
Single InternViT L2886.02
Five-layer equal-weight fusion87.08
Five-layer independent projection89.93
Original UMVA92.15

The independent-projection alternative matches the original linear projector's parameter count; UMVA exceeds it by 2.22 pp. We additionally conducted a single-seed comparison of a SAMGA-inspired visual-prior component: UMVA achieved 92.20% Top-1, while the SAMGA-style fixed visual prior achieved 83.25% Top-1.

References

  1. Jiang, L., et al. (2026). Subject-Aware Multi-Granularity Alignment for Zero-Shot EEG-to-Image Retrieval (SAMGA). arXiv:2604.17782. https://arxiv.org/abs/2604.17782

I thank the authors for their detailed rebuttal, which has successfully resolved some of concerns (Q3 and Weakness 4). Nevertheless, the fundamental issues regarding the actual inter-subject contribution remain unresolved. Under a strict common training budget, the real improvement is marginal (~1 point), and the headline 48.10% figure relies on test-set checkpoint selection, compromising its validity under strict LOSO. Moreover, the proposed TC-SAW mechanism is outperformed by simpler pooled global standardization in the authors' own ablations, demonstrating that the advertised coordination does not drive the results.

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

This paper addresses subject-dependent distribution shifts between training and deployment in EEG-to-image retrieval. It presents a unified framework that coordinates shared representation learning, source-subject normalization, and unseen-target-subject calibration. Specifically, UMVA constructs a shared visual reference, while STAR-EEG uses a shared temporal basis and aggregates repeated trials during EEG latent encoding. TC-SAW standardizes EEG latents for each source subject, and SPPC combines a source-statistics prior with previously observed target EEG latents. Extensive experiments and ablations on THINGS-EEG2 demonstrate improvements in both intra-subject and inter-subject retrieval.

Critical strengths
  • The motivation is clear and well justified. EEG signals exhibit inter-subject variability, making distribution shift an important challenge for cross-subject EEG-to-image retrieval. Addressing this issue is relevant to deployment on unseen users.
  • TC-SAW and SPPC form a coherent framework for training and deployment. TC-SAW learns EEG representations in subject-standardized coordinates, while SPPC addresses the lack of reliable target-subject statistics at cold start by combining a finite source prior with accumulated target observations.
  • The experimental evaluation is extensive, covering intra-subject and inter-subject retrieval, component ablations, and multiple subjects and independent runs. The cold-start analysis is particularly informative, showing that SPPC provides notable benefits when target-subject history is limited.
Critical weaknesses
  • The connection between the observed distribution shifts and the proposed normalization could be established more clearly. Although Table 20 provides quantitative latent-space diagnostics, the reported metrics do not uniformly improve: the absolute mean-vector gap decreases, whereas the standardized mean-shift RMS increases. Figure 2 also makes it difficult to assess how the held-out subject differs from the source subjects and which differences are corrected by TC-SAW and SPPC. Further analysis that distinguishes changes in feature scale from improvements in representation alignment would strengthen the motivation and interpretation.
  • Subject-adaptive whitening has already been used in this task in SATTC. The paper should more clearly distinguish the contribution of RNSC from existing normalization and test-time calibration approaches. In particular, the motivation for the specific SPPC design, and its advantages over simpler source–target statistics interpolation require further justification.
  • Table 5 provides a limited view of ablations compared with Table 14 in the appendix. Including the transductive results in the main text would help readers assess the remaining gap under different subject information-providing settings. The transductive setting functions as an information-rich reference upper bound for the comparison of SPPC.
  • Table 2 does not report per-subject results for SATTC. The authors should clarify why these results are unavailable. Since both methods use subject-adaptive normalization, a clearer comparison would help isolate the contribution of RNSC.
Questions for authors
  • What are the principal methodological differences between RNSC and prior subject-adaptive normalization methods, particularly SATTC? Can the benefits of SPPC be demonstrated against simpler source-initialized or interpolated-statistics baselines?
  • Can the authors clarify the interpretation of Figure 2 and the mixed trends in Table 20? In particular, how much of the apparent alignment improvement reflects feature-scale contraction, and how much reflects better alignment of stimulus-relevant representations across subjects in train and test?
  • Could the main text include the transductive reference results and explain the missing per-subject SATTC results in Table 2? By the above clarifications, based on the rational of the motivation, the completeness in experiments design, I’m willing to raise my score.
4 comments

We appreciate the reviewer’s careful assessment of our training–deployment formulation, particularly the requests to distinguish RNSC from SATTC and to clarify the geometry diagnostics and reference baselines. These questions have helped us explain the motivation, information constraints, and comparisons more precisely.

Q1. Differences from SATTC and stronger statistical baselines

1.What are the principal methodological differences between [...] benefits of SPPC be demonstrated against simpler source-initialized or interpolated-statistics baselines?

As explained in the General Response (Core contribution and novelty), RNSC differs from SATTC through the joint optimization of source-side standardized representations and causal, query-by-query target calibration, which keeps the learned retrieval geometry compatible across training and deployment. SATTC explicitly distinguishes these deployment modes in its Limitations (Huang and Zhu, CVPR 2026): its SAW+CSLS branch allows one-time calibration followed by frozen inference, whereas “the structural expert and PoE are most naturally applied in a batch refinement stage” when multiple unlabeled queries are available.

This identifies a complementary question motivating SPPC: how can an unseen subject be calibrated from the first retrieval onward when future test queries are not yet available? Rather than refining a precomputed multi-query similarity matrix, SPPC estimates latent normalization statistics from a source prior and only previously observed, unlabeled target queries; each current query is retrieved before updating the statistics. TC-SAW further brings latent standardization into source-side contrastive training, making this sequential deployment calibration compatible with the learned retrieval coordinates.

Using identical checkpoints, candidates, query embeddings, and causal information access over five seeds × ten held-out subjects:

CalibrationTop-1 (%)
Source-only20.42
History-only46.89
Source-initialized EMA (α = 0.10)43.67
Fixed interpolation (source weight = 0.50)35.77
SPPC (κ = 8)48.10

In this shared-checkpoint comparison, SPPC improves over History-only by +1.21 pp Top-1. An interpolation weight selected by scanning target-test outcomes approaches SPPC on full-sequence accuracy; we report this separately as a sensitivity analysis.

Q2. Latent geometry, feature scale, and stimulus-relevant alignment

2.Can the authors clarify [...]how much reflects better alignment of stimulus-relevant representations across subjects in train and test?

Table 20's absolute source–target mean gap decreases 4.479 to 2.742, while standardized mean-shift RMS increases 0.459 to 0.590. In a paired ten-subject, seed-99 analysis, source variance contracts 0.378 to 0.082 while the absolute gap decreases 4.50 to 2.65: raw distance reduction alone is not evidence of scale-invariant alignment.

The stimulus-conditioned online SAW/SPPC diagnostics add a task-relevant view:

MetricBeforeAfter
Effective rank, TC-SAW enabled15.1421.43
Same–different stimulus cosine margin, TC-SAW enabled0.01110.4661
Alignment-retrieval margin correlation, TC-SAW enabled0.2790.596
Same–different stimulus cosine margin, TC-SAW disabled0.35040.4462

The online transformation improves stimulus-conditioned discrimination in both training-time SAW groups. The complementary training-time normalization comparison (see PRNL Q3) examines None, per-subject, and pooled-global normalization, revealing further opportunities to design source-side statistical coordinates. We will clarify in Figure 2/Table 20 that absolute-distance changes partly reflect feature scale, while stimulus-conditioned measures capture retrieval-relevant geometry.

Q3. Transductive reference and SATTC per-subject reporting

3.Could the main text include the transductive reference results and explain the missing per-subject SATTC results in Table 2?

The revised Table 5 includes a transductive upper-information reference (49.42 ± 0.46% Top-1, 78.33 ± 0.47% Top-5) alongside SPPC (48.10 ± 0.67%, 76.34 ± 0.20%) and History-only (46.89 ± 0.79%, 75.68 ± 0.34%). Unlike SPPC, the transductive setting accesses all unlabeled target queries, explaining the 1.32 pp Top-1 gap.

SATTC provides aggregate results (14.8% Top-1 / 38.4% Top-5) but no per-subject scores in its paper/supplement; the revised Table 2 caption explains the dashes.

References

  1. Huang, Q., & Zhu, W. (2026). SATTC: Structure-Aware Label-Free Test-Time Calibration for Cross-Subject EEG-to-Image Retrieval. CVPR 2026, 16887–16896. https://openaccess.thecvf.com/content/CVPR2026/html/Huang_SATTC_Structure-Aware_Label-Free_Test-Time_Calibration_for_Cross-Subject_EEG-to-Image_Retrieval_CVPR_2026_paper.html

Thanks for your detailed responses to my questions. Some concerns, including the comparison with SATTC, the motivation for the specific SPPC design, and other clarifications on experiments, have been addressed. Regarding Q2, can you briefly explain the meaning of the metric in the table in plain language?

Regarding Q2, can you briefly explain the meaning of the metric in the table in plain language?

Thank you for your positive feedback and follow-up question. The three metrics can be interpreted as follows:

  • Effective rank: Measures how many feature dimensions are meaningfully utilized. A higher value indicates a less concentrated representation.
  • Same–different stimulus cosine margin: Measures how much more similar EEG representations are for the same visual stimulus than for different stimuli. A larger margin indicates better stimulus discrimination.
  • Alignment–retrieval margin correlation: Measures how strongly stimulus-level alignment is associated with the separation between correct and incorrect image candidates during retrieval.

After online SAW/SPPC calibration, effective rank increases from 15.14 to 21.43, the stimulus cosine margin from 0.0111 to 0.4661, and the alignment–retrieval correlation from 0.279 to 0.596.

Additionally, under the same online calibration procedure, the post-calibration stimulus cosine margin is 0.4661 with training-time TC-SAW versus 0.4462 without TC-SAW. This modest difference provides additional evidence of stimulus-relevant representation changes associated with training-time normalization.

We sincerely appreciate your constructive feedback, which has helped us clarify the roles of our components and improve the rigor and clarity of our experimental analysis.

Dear Reviewer ieJB,

Thank you again for your constructive feedback and follow-up question. Before the discussion period closes, we wanted to briefly share that we have completed our multi-seed, strictly leak-free validation experiments (3 seeds ×\times 10 subjects on THINGS-EEG2, and 3 seeds ×\times 20 subjects on Alljoined1.6M), summarized in General Response #2:

  • Leak-Free Causal vs. Transductive Benchmark (Q1 & Q3): Under strictly source-only, concept-disjoint checkpoint selection (never accessing target test labels), SPPC achieves 44.68% ±\pm 0.98% causal Top-1—recovering 95.5% of the offline Transductive reference upper bound (46.80% ±\pm 0.89%) while delivering +13.75 pp cold-start gain over History-only (27.50% vs. 13.75%) across the first 8 queries.
  • Task-Relevant Geometry & Training Standardization (Q2): Alongside our plain-language explanation of the stimulus-conditioned cosine margin (0.0111 →\to 0.4661) and effective rank (15.14 →\to 21.43), we evaluated the bias–variance trade-off across training-time standardization estimators (Global vs. subject-wise TC-SAW) in the same normalized retrieval space.

We hope our metric explanations and these completed multi-seed experiments address all of your remaining concerns, and we would be grateful to know if any further clarification would be helpful.

Reviewer_mirKOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper addresses a very hard sub-problem in EEG-image retrieval algorithms of unseen subject calibration. Many EEG-image models are heavily dependent on source-subject normalization of EEG-image models which has been critical to build subject-specific EEG-image models. This paper presents Representation–Normalization–Sequential Calibration (RNSC) framework, which treats shared representation learning, source-subject normalization, and unseen-subject calibration as coordinated stages of a single retrieval space and achieves very high accuracy in comparison to several SoTA EEF-image retrieval algorithms.

Critical strengths
  1. Hard problem to solve relevant for the ICLR community
  2. Neat set of the algorithmic and ML contributions
  3. Thorough evaluation with better results than several SoTA algorithms
  4. Well written paper
Critical weaknesses
  1. THINGS-EEG2 is one of the best datasets for studying EEG-to-image retrieval, but it has important limitations for cross-subject adaptation. It has only 10 subjects across a broad array of scenarios. The key distinction is that THINGS-EEG2 is excellent for measuring visual representation decoding, while generalization to unseen individuals requires a more careful experimental design. The original THINGS-EEG dataset has 50 subjects but has its own limitations which were better addressed for this study but 10 subjects is too small for cross-subject generalization especially for the claim of unseen subjects.

  2. There is limited discussion of EEG foundation models. Some missing refs in this space include:

W.-B. Jiang, L.-M. Zhao, and B.-L. Lu. Large brain model for learning generic representations with tremendous eeg data in bci. arXiv preprint arXiv:2405.18765, 2024.

D. Kostas, S. Aroca-Ouellette, and F. Rudzicz. Bendr: Using transformers and a contrastive self-supervised learning task to learn from massive amounts of eeg data. Frontiers in Human Neuroscience, 15:653659, 2021.

J. Wang, S. Zhao, Z. Luo, Y. Zhou, H. Jiang, S. Li, T. Li, and G. Pan. Cbramod: A criss-cross brain foundation model for eeg decoding. arXiv preprint arXiv:2412.07236, 2024.

Questions for authors

The biggest question I have is the difference between the size of the subject sample size and the generalist claims made in the paper.

Specifically, how reliable are the accuracy numbers and how statistically significant are these numbers over a small sample size of patients. Are there bigger datasets for this task?

Can the authors answer the two weakness queries:

  1. THINGS-EEG2 is one of the best datasets for studying EEG-to-image retrieval, but it has important limitations for cross-subject adaptation. It has only 10 subjects across a broad array of scenarios. The key distinction is that THINGS-EEG2 is excellent for measuring visual representation decoding, while generalization to unseen individuals requires a more careful experimental design. The original THINGS-EEG dataset has 50 subjects but has its own limitations which were better addressed for this study but 10 subjects is too small for cross-subject generalization especially for the claim of unseen subjects.

  2. There is limited discussion of EEG foundation models. Some missing refs in this space include:

W.-B. Jiang, L.-M. Zhao, and B.-L. Lu. Large brain model for learning generic representations with tremendous eeg data in bci. arXiv preprint arXiv:2405.18765, 2024.

D. Kostas, S. Aroca-Ouellette, and F. Rudzicz. Bendr: Using transformers and a contrastive self-supervised learning task to learn from massive amounts of eeg data. Frontiers in Human Neuroscience, 15:653659, 2021.

J. Wang, S. Zhao, Z. Luo, Y. Zhou, H. Jiang, S. Li, T. Li, and G. Pan. Cbramod: A criss-cross brain foundation model for eeg decoding. arXiv preprint arXiv:2412.07236, 2024.

Additional Questions: a. Why not try other datasets: THINGS-EEG1 (50 subjects) and Alljoined-1.6M (20 subjects) and Alljoined1 (8 subjects)? Can you train on one dataset and test on another? b. Can you use the EEG-ImageNet dataset for a relevant experiment in your pipeline?

At a surface level, the paper results look very impressive. I would be more supportive of this paper if the authors can answer these queries convincingly.

4 comments

Q1–Q2. Subject sample size and statistical reliability

1.the difference between the size of the subject sample size and the generalist claims made in the paper.

2.how reliable are the accuracy numbers and how statistically significant are these numbers over a small sample size of patients. Are there bigger datasets for this task?

Five runs across ten held-out subjects yield a +1.21 pp SPPC gain over History-only in the original query order (9/10 subjects improve). Treating subjects, rather than run-folds, as independent units gives a 95% CI [+0.41, +2.01] pp; leave-one-subject-out sensitivity gives +1.00 to +1.41 pp. The main finding is summarized in the General Response.

Our additional 20-subject, three-seed Alljoined-1.6M LOSO evaluation (60 held-out folds) obtains 16.03%/36.96% Top-1/Top-5 using best checkpoints, or 13.87%/37.08% at epoch 10. At those same best checkpoints, SPPC obtains 16.025% Top-1 versus 15.433% for causal History-only (Pure), an average +0.592 pp. This is complementary within-dataset unseen-subject evidence. Checkpoint-selection details are discussed in PRNL Q1.

Q3. EEG foundation models

3.limited discussion of EEG foundation models.

We added BENDR, LaBraM, and CBraMod to Related Work. Their EEG pretraining strategies are complementary to the training–deployment normalization and sequential calibration studied here.

AQ(a)–AQ(b). Larger datasets, cross-dataset transfer, and EEG-ImageNet

THINGS-EEG2 is one of the best datasets for [...] the claim of unseen subjects.

a. Why not try other datasets: THINGS-EEG1 (50 subjects) and Alljoined-1.6M (20 subjects) and Alljoined1 (8 subjects)? Can you train on one dataset and test on another?

b. Can you use the EEG-ImageNet dataset for a relevant experiment in your pipeline?

The new Alljoined-1.6M 20-subject LOSO results are compared with published 200-way retrieval references; settings are not strictly matched (see the General Response).

Intra-subject reference (200-way retrieval):

MethodEvaluated subjectsTop-1 (%)Top-5 (%)Selection/evaluation note
ATMS + ProCA2015.5038.60100-way
ATMS + ProCA2012.4031.60150-way
ATMS + ProCA2010.5026.30200-way
RNSC, raw-mean, epoch 202024.3849.544 seeds; common-epoch measurement
RNSC, encoded-mean, epoch 202028.9455.364 seeds; common-epoch measurement

Within the RNSC Alljoined intra-subject runs (four seeds × 20 subjects, shared STAR-EEG architecture, epoch 20), encoded-mean exceeds raw-mean by +4.56 pp Top-1 and +5.83 pp Top-5. This supports the design choice to aggregate repeated EEG observations after latent encoding, rather than averaging raw signals before nonlinear encoding.And the RNSC encoded-mean best-checkpoint result is 32.02%/57.08%.

Cross-subject reference (200-way retrieval, 20 held-out subjects):

Method or configurationTop-1 (%)Top-5 (%)Information/selection note
SAMGA original encoder (reported by SCORE)6.0820.55Without target-domain coordinate recovery
SAMGA + subject-adaptive whitening (SCORE comparison)8.9327.54Whitening-only reference
SCORE, recovery-aware source training only9.0826.78No deployment coordinate recovery
SCORE, deployment coordinate recovery only10.6730.51Without recovery-aware source training
Complete SCORE12.0132.16Published label-free coordinate-recovery framework; 3 seeds
RNSC, diagonal SAW + cosine, epoch 2012.5034.763 seeds; common-epoch measurement
RNSC, diagonal SAW + cosine, epoch 1013.8737.083 seeds; additional epoch-based comparison

Within-method LOSO diagnostics (diagonal online SAW + SPPC κ=8 + cosine; 3 seeds × 20 held-out subjects = 60 folds):

Evaluation window (60-fold mean)Top-1 (%)Top-5 (%)
Best checkpoint16.02536.958
Epoch 1013.86737.083
Epoch 2012.50034.758

At the same best checkpoint (mean selected epoch: 10.00):

Calibration settingTop-1 (%)
History-only (Pure)15.433
Transductive15.858
SPPC16.025
SPPC − History-only+0.592 pp

The +0.592 pp is a within-checkpoint causal SPPC-versus-History-only difference averaged across 60 folds. History-only (Pure) and noncausal Transductive reuse the same SPPC best checkpoint; the transductive reference uses all unlabeled target queries and is not a permissible prequential deployment result.

The comparison shows that RNSC is competitive with published Alljoined retrieval results, including SCORE. The methods differ in visual/EEG pipelines, checkpoint selection, and target-information access. SCORE reports label-free target-side coordinate recovery, while RNSC uses a causal predict-before-update history; the exact temporal information budget in SCORE remains to be verified from its inference implementation.

A separate single-seed RNSC diagonal-plus-low-rank configuration with causal CSLS reaches 16.93% Top-1 / 42.60% Top-5 at epoch 10. Because it changes both the normalization geometry and retrieval rule, we list it separately from the three-seed cosine results.

Looking ahead, we plan to investigate THINGS-EEG2–Alljoined A→B cross-dataset transfer and examine EEG-ImageNet protocol adaptation, including channel mapping, preprocessing, and candidate-set alignment. These future evaluations will help characterize generalization beyond within-dataset LOSO.

References

  1. Gifford, A. T., et al. (2022). A large and rich EEG dataset for modeling human visual object recognition. NeuroImage, 264, 119754. https://doi.org/10.1016/j.neuroimage.2022.119754
  2. Xu, J., et al. (2025). Alljoined-1.6M: A Million-Trial EEG-Image Dataset for Evaluating Affordable Brain-Computer Interfaces. arXiv:2508.18571. https://arxiv.org/abs/2508.18571
  3. Kostas, D., Aroca-Ouellette, S., & Rudzicz, F. (2021). BENDR: Using Transformers and a Contrastive Self-Supervised Learning Task to Learn From Massive Amounts of EEG Data. Frontiers in Human Neuroscience, 15, 653659. https://doi.org/10.3389/fnhum.2021.653659
  4. Jiang, W.-B., Zhao, L.-M., & Lu, B.-L. (2024). Large Brain Model for Learning Generic Representations with Tremendous EEG Data in BCI (LaBraM). arXiv:2405.18765. https://arxiv.org/abs/2405.18765
  5. Wang, J., et al. (2024). CBraMod: A Criss-Cross Brain Foundation Model for EEG Decoding. arXiv:2412.07236. https://arxiv.org/abs/2412.07236
  6. Zhang, K., et al. (2025). CognitionCapturer: Decoding Visual Stimuli from Human EEG Signal with Multimodal Information. AAAI, 39(13), 14486–14493. https://doi.org/10.1609/aaai.v39i13.33587
  7. Zhou, K., et al. (2026). ProCA: Progressive Contrastive Alignment for Robust EEG Visual Decoding. arXiv:2609.05094. https://arxiv.org/abs/2609.05094
  8. Cui, Z., et al. (2026). SCORE: Subject Coordinate Recovery for Label-Free Cross-Subject EEG-to-Image Retrieval. arXiv:2608.19134. https://arxiv.org/abs/2608.19134
  9. Jiang, L., et al. (2026). Subject-Aware Multi-Granularity Alignment for Zero-Shot EEG-to-Image Retrieval (SAMGA). arXiv:2604.17782. https://arxiv.org/abs/2604.17782

Dear Reviewer mirK,

Thank you again for your encouraging review and constructive questions regarding cohort scale, statistical reliability, and EEG foundation models. Before the discussion period closes, we wanted to briefly share our completed multi-seed, strictly leak-free evaluation (summarized in General Response #2), which directly addresses your concerns and updates the Alljoined-1.6M checkpoints reported in our earlier reply:

  • Leak-Free Larger-Cohort Evaluation on Alljoined-1.6M (Q1–Q2 & AQ(a)): Moving beyond the best-checkpoint Alljoined numbers (16.025%) in our earlier comment, we completed a strictly leak-free evaluation at a single common epoch (epoch 10) across all 3 seeds ×\times 20 held-out subjects (60 LOSO folds) without using target test performance for selection. SPPC achieves 13.87% ±\pm 0.56% full-sequence Top-1 (competitive with published SCORE at 12.01%) while boosting early cold-start accuracy over the 0.50% chance baseline by +53.9% relative (4.17% vs. 2.71% over the first 8 queries).
  • Multi-Seed Leak-Free Stability on THINGS-EEG2 (Q1–Q2): Under strictly source-only, concept-disjoint validation across 3 seeds ×\times 10 held-out subjects, SPPC achieves 44.68% ±\pm 0.98% causal Top-1 and +13.75 pp cold-start gain over History-only (27.50% vs. 13.75%) across the first 8 queries.
  • Scoped Generalist Claims & Literature/Cross-Dataset Outlook (W1–W2 & AQ(a)–(b)): In light of these 10-subject and 20-subject multi-seed findings, we will tighten our framing in the revised manuscript to focus specifically on causal cold-start calibration rather than broad generalist claims over long sequences, alongside our added BENDR, LaBraM, and CBraMod discussion and planned THINGS-EEG2$\to$Alljoined / EEG-ImageNet cross-dataset extensions.

We truly appreciate your support and welcome any final thoughts you may have.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response

Dear PCs, SACs, ACs, and Reviewers,

We sincerely thank all reviewers for their careful reading and constructive feedback. We appreciate their recognition of the importance and difficulty of retrieving visual information from previously unseen EEG subjects, as well as the value of a predict-before-update protocol for latent-space calibration. Their comments have helped us sharpen the central scientific question and its empirical evaluation.

For the reviewers' convenience, we have highlighted the revisions in blue in the revised manuscript.

Core contribution and novelty. RNSC addresses a training–deployment mismatch that is central to cross-subject EEG-to-image retrieval: source-subject representations are learned under particular statistical coordinates, while a new subject's compatible coordinates must be estimated from limited, sequentially available observations. STAR-EEG provides the EEG representation foundation through channel-shared temporal factorization and post-encoding repetition aggregation.

Our framework links three complementary mechanisms: UMVA constructs an EEG-aligned visual target by integrating multi-level InternViT representations; TC-SAW incorporates subject-conditioned normalization into source-side contrastive representation learning; and SPPC combines a source statistical prior with previously observed, unlabeled target queries to estimate deployment-time normalization.

Under predict-before-update, the current query is retrieved before its EEG representation enters the online statistics, with the encoder parameters fixed. The distinctive methodological contribution is the joint design of source-trained representation coordinates and causally estimated target coordinates for retrieval.

Additional evidence and clarifications. To address the reviewers' questions, we add (i) visual-fusion controls, including 92.15% versus 89.93% Top-1 for UMVA and a parameter-matched alternative in the epoch-50 intra-subject comparison; (ii) matched-checkpoint causal calibration baselines (48.10% SPPC versus 46.89% History-only Top-1) and randomized-query cold-start analysis; (iii) stimulus-conditioned geometry diagnostics and subject-level reliability estimates (SPPC gains in 9/10 subjects, with a subject-level 95% CI of [+0.41,+2.01] pp under the original order); and (iv) 20-subject Alljoined-1.6M intra-/inter-subject evaluations with published-method context.

We also present a 27-study checkpoint-protocol audit, common-epoch comparisons and the 4-versus-80 repetition consideration; clarify preprocessing and target-information access; and update the manuscript's EEG foundation-model discussion, transductive reference, and SATTC reporting. Detailed settings, limitations, and evidence for each point are given in the reviewer-specific responses below. The UBP input-level MVNN is fitted using each participant's training EEG partition only; SPPC separately enforces predict-before-update on the target-test latent-query stream. This information distinction is detailed in PRNL Q4.

The new None / Source-subject / Global comparison (see PRNL Q3) further highlights the potential of training-time latent standardization as an extensible design choice for cross-subject retrieval.

Once again, thank you for the time and thoughtful consideration you have devoted to this paper. Your feedback has helped us clarify our contributions, present the supporting evidence more transparently, and identify productive directions for further research.

General Response #2: Summary of Completed Multi-Seed Leak-Free Evaluation & Final Manuscript Revisions

Dear PCS, SACs, ACs, and Reviewers,

We sincerely thank all four reviewers (mirk, ieJB, 66Jr, and PRNL) for their constructive engagement throughout the discussion period. In particular, we appreciate Reviewer PRNL for raising the rating after reviewing our leak-free validation protocol, and Reviewers ieJB, 66Jr, and mirk for their insightful feedback on evaluation rigor, visual/encoder controls, and dataset scale.

To provide a unified summary for the Area Chair and all reviewers at the close of the discussion period, we consolidate below the completed multi-seed, strictly leak-free evaluations and the corresponding revisions we will incorporate into the camera-ready/revised manuscript (detailed per-seed tables and prefix curves are posted in our follow-up comments to Reviewers PRNL and 66Jr):

1. Completed Multi-Seed Leak-Free Evaluation

Following reviewer feedback, we replaced test-set checkpoint selection with a concept-disjoint, source-only pseudo-target validation protocol on THINGS-EEG2 (3 seeds ×\times 10 held-out subjects) and a single common-epoch protocol (epoch 10) on Alljoined (3 seeds ×\times 20 held-out subjects), never accessing target test labels for model selection:

Dataset (Leak-Free Protocol)Pure Top-1SPPC Top-1TransductiveFirst-8 Gain (SPPC −- Pure)First-16 Gain (SPPC −- Pure)
THINGS-EEG2 (3 seeds ×\times 10 subj.)44.70% ±\pm 1.18%44.68% ±\pm 0.98%46.80% ±\pm 0.89%+13.75 pp (27.50% vs. 13.75%)+6.25 pp (32.71% vs. 26.46%)
Alljoined (3 seeds ×\times 20 subj.)13.91% ±\pm 0.54%13.87% ±\pm 0.56%14.57% ±\pm 0.50%+1.46 pp / +53.9% rel. (4.17% vs. 2.71%)+1.35 pp / +17.4% rel. (9.17% vs. 7.81%)
  • Theoretical consistency across query horizons: Across both benchmarks, SPPC delivers substantial early-query improvements during cold start—doubling First-8 accuracy on THINGS-EEG2 (+13.75 pp) and improving First-8 accuracy by +53.9% relative on Alljoined—before smoothly decaying its prior weight κ/(κ+n)\kappa / (\kappa + n) to reach asymptotic parity with unbiased Pure estimation over the full 200-query stream.

2. Summary of Key Revisions in the Manuscript

  1. Leak-Free Primary Evaluations: We will replace test-selected checkpoints across our main cross-subject tables with the multi-seed leak-free validation results above (44.68% ±\pm 0.98% on THINGS-EEG2; 13.87% ±\pm 0.56% on Alljoined), and clearly separate leak-free intra-subject evaluations from the originally reported checkpoints.
  2. Scoped Contribution (Causal Cold-Start Calibration): We will refine the abstract, introduction, and conclusions to position SPPC specifically as a causal cold-start calibration mechanism under strict predict-before-update constraints, rather than claiming large gains over long query sequences.
  3. Training-Time Standardization Positioning: We will present the unified standardized coordinate system as our core design principle and directly report the Global vs. subject-wise TC-SAW comparison (45.39% vs. 44.21% at epoch 20) as a bias–variance trade-off without claiming subject-wise normalization is universally superior.
  4. Controlled Baselines & Literature Context: We have incorporated the parameter-matched visual fusion controls (UMVA 92.15% vs. 89.93%), causal calibration baselines (Source-only, EMA, fixed interpolation, and transductive upper bounds), plain-language geometry diagnostics, and EEG foundation model discussions (BENDR, LaBraM, CBraMod).

We thank all reviewers and the Area Chair once again for their time and rigorous guidance, which have substantially strengthened the scientific validity and clarity of this paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.