acceptodds
Under review as a conference paper at ICLR 2027

UGPA-AAD: Uncertainty-aware Gaussian Graph and Prototype Adaptation for Cross-Subject Auditory Attention Decoding

Abstract

Auditory Attention Decoding (AAD) plays a vital role in enabling intelligent auditory brain-computer interfaces by decoding attention-related neural responses from EEG signals. Currently, cross-subject AAD has gained increasing attention, aiming to learn subject-invariant attention-related representations that generalize to target subjects. Despite recent advancements, cross-subject AAD still faces two core challenges under subject-specific variations and signal disturbances: unreliable spatial dependency modeling and cross-subject semantic inconsistency. To address these challenges, we propose an Uncertainty-aware Gaussian Graph and Prototype Adaptation framework (UGPA-AAD), which improves representation reliability and cross-subject semantic consistency through uncertainty-aware structural learning and prototype-guided semantic alignment. Specifically, we first introduce a Multi-Scale Feature Encoder (MSFE) to capture informative neural representations by integrating complementary representations across multiple temporal receptive fields. Then, to mitigate unreliable channel interactions under subject variability and signal disturbances, we design an Uncertainty-Aware Gaussian Graph Encoder (UGGE) that leverages channel-wise uncertainty to modulate spatial dependencies and construct robust structural representations. Finally, Prototype-Guided Cross-Subject Learning (PCSL) is designed to alleviate class-wise semantic inconsistency across subjects by constructing global class prototypes as semantic anchors to align representations across subjects toward consistent class structures, thereby enhancing semantic discriminability. Experimental results on four benchmarks demonstrate that our method consistently outperforms existing approaches in both audio-only and audio-visual scenarios. Our code will be released upon acceptance.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Electroencephalography (EEG), Cross-subject, Auditory Attention Decoding (AAD), Graph Convolutional Network (GCN), Uncertainty

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_hCShOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes UGPA-AAD for spatial auditory attention decoding across different subjects, to discriminate left versus right attention purely from EEG signals. The framework consists of three core modules: MSFE, UGGE, and PCSL. The authors adopt leave-one-subject-out evaluation on DTU, MM-AAD, AVED and AV-GC datasets, using 1-second and 2-second time windows. Seven baseline methods are reimplemented for comparison.

Critical strengths
  1. Cross-subject AAD is practically valuable.
  2. Instead of directly copying numbers from existing papers, the authors re-run all baseline methods under a unified preprocessing and data partition pipeline.
  3. Weighting graph edges by uncertainty is an intuitive and reasonable design.
Critical weaknesses
  1. From what I can see, all seven baselines are trained only on source subjects. The paper also does not clarify whether these adaptation windows are exactly the same samples used for testing. The ablation results in the manuscript itself show how critical this issue is. When L_{tpcl} is removed (Table 7b): DTU: 56.2 / 56.9, lower than S2Former at 57.7 / 59.7. AV-GC: 57.8 at 1 s, nearly on par with PCRNet (57.9); 57.6 at 2 s, roughly +1.0 above the best baseline MHANet (56.6). So most performance advantage reported over baselines likely comes from access to target-subject data, rather than the UGGE module or prototype design.
  2. Accuracy values stay close to random guess level, and statistical significance tests are missing.I ask the authors to supply: paired subject-wise significance tests (e.g. Wilcoxon test) against the strongest baseline, per-subject accuracy records, and performance variance across different random seeds.
  3. For AV-GC, gaze direction is a well-known confounding factor. The paper notes that its strongest learned graph connections are fronto-central channels (Fig.7). The largest performance gains appear in audiovisual settings: AV-GC (+2.7 / +3.8), AVED-AV (+3.3 / +3.1). Taken together, these observations suggest the improvement may come from eye movement signals rather than auditory attention itself.
  4. Hyperparameter selection and checkpoint picking appear to rely on test-set performance. Please clarify this.
  5. Several inconsistencies exist in method description. For examples, "SA" is ablated in Table3 but never defined; Equation (14) computes cross entropy over f_m, while Section3.1 states the classifier operates on z. Please clarify which feature feeds the classifier, and how the classification head on z is optimized; The paper says \sigma is refined into \hat{\sigma}, but Equation (8) only updates \mu.
Questions for authors
  1. Are the unlabeled target windows used for adaptation exactly the same samples used for testing? What are the performance of all methods without target data access?
  2. Can you share per-subject accuracy, paired significance test against the best baseline, and performance variance across random seeds?
  3. What are AV-GC results split by condition, especially gaze-incongruent trials? Was any ocular artifact removal applied? Why were 6 out of 16 participants excluded?
  4. Is there empirical evidence showing that \sigma indeed captures subject-wise or channel-wise signal unreliability?
4 comments

The authors have not responded to my comments. Therefore, I keep my original rating.

AuthorsOpenReview ↗

Response - Part 1

We thank the reviewer for the careful assessment and constructive requests. We agree that data access, subject-level reliability, potential ocular confounds, and implementation consistency must be made explicit. We address each concern below and will incorporate the corresponding clarifications and verified analyses into the revision.

1. Data Access, Adaptation Fairness, and the Role of TPCL (W1, Q1)

We respectfully clarify that UGPA-AAD and all seven baselines draw their data from the same pool of K non-test subjects in each outer leave-one-subject-out (LOSO) fold, while the held-out test subject is reserved exclusively for final evaluation. As defined in Sec. 3.1, each fold contains K+1 subjects, comprising K−1 source subjects, one adaptation subject, and one held-out test subject (K−1:1:1). The adaptation subject is randomly selected from the K non-test subjects, with 60% of its trials used for unlabeled TPCL adaptation and the remaining 40% for labeled validation. The seven original baselines are trained using labeled data from K non-test subjects, whereas UGPA-AAD uses K−1 labeled source subjects and the designated adaptation subject. All methods follow the same LOSO evaluation protocol across all datasets. We have clarified these important details and the ambiguous use of "target subject" in the revised manuscript, with changes highlighted in blue.

To assess adaptation effectiveness under matched data access, we further compare representative baselines augmented with DANN, Deep CORAL, pseudo-label self-training (PL), and prototype alignment (PA), using identical source, adaptation, and validation partitions. The results are presented in Table R1. UGPA-AAD consistently outperforms all adaptation-enhanced baselines, indicating that its gains are not solely attributable to access to unlabeled adaptation data. Furthermore, the w/o TPCL results and Table 3 ablations support the complementary contributions of TPCL, UGGE, and PCSL.

Table R1. Adaptation comparison under matched data access (Acc.% ± SD).

MethodDTU 1sDTU 2sAV-GC 1sAV-GC 2s
S²M-Former57.7 ± 4.2159.7 ± 5.2554.0 ± 3.7853.2 ± 3.89
S²M-Former+DANN52.8 ± 8.4354.6 ± 7.9251.7 ± 6.8150.9 ± 8.17
S²M-Former+CORAL54.3 ± 9.2655.1 ± 8.7450.8 ± 7.6352.4 ± 7.19
S²M-Former+PL53.6 ± 7.5855.9 ± 6.9354.6 ± 5.8254.1 ± 6.47
S²M-Former+PA55.2 ± 8.1753.8 ± 9.1251.4 ± 8.3652.3 ± 7.84
ListenNet56.8 ± 7.3257.2 ± 5.8353.2 ± 4.9055.0 ± 7.06
ListenNet+DANN51.2 ± 9.0651.9 ± 8.9750.4 ± 0.1250.0 ± 8.09
Listennet+CORAL53.1 ± 10.1753.2 ± 9.7949.9 ± 8.2250.1 ± 7.81
ListenNet+PL50.5 ± 9.6051.1 ± 9.4552.4 ± 5.4753.0 ± 5.06
ListenNet+PA52.1 ± 9.1653.2 ± 9.8149.8 ± 8.9350.2 ± 8.77
UGPA-AAD w/o TPCL56.2 ±4.1556.9 ±3.8157.8 ±8.1657.6 ±9.18
UGPA-AAD58.8 ± 3.7359.2 ± 3.4460.6 ± 7.5860.4 ± 8.68

2. Eye-Gaze Confounds and AV-GC Participant Selection (W3, Q3)

We thank the reviewer for raising the important concern regarding potential gaze-related confounds in AV-GC. Our study focuses on cross-subject auditory attention decoding under a unified leave-one-subject-out (LOSO) protocol. Specifically, all four experimental conditions of AV-GC are included in the evaluation of each held-out test subject, and the reported accuracy reflects the overall decoding performance across these conditions. No condition-specific models are trained, and gaze-direction labels are not used as inputs or supervision during training or adaptation.

The consistent improvements on auditory-only datasets (DTU and MM-AAD) demonstrate the effectiveness of our method beyond audiovisual settings. Moreover, Fig. 7 reveals prominent connections across fronto-central, temporal, and parietal regions, suggesting that the learned representations are not restricted to regions potentially associated with ocular activity.

For AV-GC, we followed the official preprocessing pipeline and S²Former's experimental settings without additional ocular-artifact removal. As explained in Appendix A.1 (Dataset Details), we retained 10 participants with complete recordings across all four conditions. Specifically, participants 2, 5, and 6 were excluded due to consent restrictions, while participants 1, 3, and 14 had incomplete recordings, consistent with the official dataset documentation.

AuthorsOpenReview ↗

Response - Part2

3. Subject-Wise Significance and Random-Seed Robustness (W2, Q2)

We conducted a subject-wise comparison between UGPA-AAD and the strongest baseline, ListenNet, on AVED-AV using a two-sided exact Wilcoxon signed-rank test. UGPA-AAD achieves mean improvements of 3.22% and 3.18% for 1 s and 2 s windows, respectively, with statistically significant differences (p = 0.005859 and p = 0.013672). Furthermore, UGPA-AAD outperforms ListenNet on 9 of 10 subjects in both settings, with significance maintained after Holm correction (adjusted p = 0.011719 and 0.013672).

Table R2. Subject-wise classification accuracy and paired tests on AVED-AV (%).

SubjectUGPA-AAD 1 sListenNet 1 sUGPA-AAD 2 sListenNet 2 s
S0155.0553.6155.4253.29
S0255.0153.7255.2253.70
S0359.7356.5360.3153.83
S0458.0352.7459.3553.98
S0552.4053.4052.6152.24
S0660.6852.3864.2052.65
S0759.5252.3859.6454.97
S0857.1654.5655.7155.32
S0958.8255.5357.0056.47
S1053.2352.5952.9454.14
Mean56.9653.7457.2454.06
Wilcoxon W2.04.0
Two-sided p0.0058590.013672

To assess robustness to random initialization, we conducted experiments with three random seeds (42, 1024, and 3407) across all datasets and both window lengths. As shown in Tables R3 and R4, the standard deviation of mean accuracy across seeds ranges from 0.12% to 0.46%, demonstrating consistent performance across random initializations.

Table R3. Performance across random seeds on DTU and MM-AAD (Acc.% ± SD).

SeedDTU 1 sDTU 2 sMM-AAD AO 1 sMM-AAD AO 2 sMM-AAD AV 1 sMM-AAD AV 2 s
4258.8 ± 3.7359.2 ± 3.4458.1 ± 5.4258.5 ± 5.1857.0 ± 4.3557.2 ± 4.26
102458.4 ± 4.3659.0 ± 3.8558.2 ± 5.1358.7 ± 5.2957.4 ± 4.9757.5 ± 4.33
340759.1 ± 3.8259.6 ± 3.1757.9 ± 4.9158.5 ± 5.3657.6 ± 4.2357.6 ± 3.90
Mean58.7759.2758.0758.5757.3357.43

Table R4. Performance across random seeds on AVED and AV-GC (Acc.% ± SD).

SeedAVED AO 1 sAVED AO 2 sAVED AV 1 sAVED AV 2 sAV-GC 1 sAV-GC 2 s
4254.1 ± 1.2554.0 ± 1.4457.0 ± 2.7457.2 ± 3.4360.6 ± 7.5860.4 ± 8.68
102454.4 ± 2.3154.2 ± 1.3656.8 ± 2.9657.3 ± 3.3360.0 ± 7.1960.0 ± 8.31
340753.7 ± 1.3853.7 ± 1.2257.0 ± 2.7056.9 ± 3.1159.7 ± 6.1459.9 ± 7.57
Mean54.0753.9756.9357.1360.1060.10

4. Hyperparameters and Checkpoint Selection (W4)

No test-subject data or test accuracy was used for hyperparameter tuning or checkpoint selection. Hyperparameters were determined using the training set, with key settings detailed in Sec. 4.1. In each LOSO fold, the optimal checkpoint was selected based on the lowest supervised loss on the 40% labeled validation trials of the non-test adaptation subject. The remaining 60% unlabeled trials were used for TPCL adaptation, and the held-out test subject was evaluated only after checkpoint selection.

5. Empirical Support for σ as a Reliability Estimate (Q4)

Within our Gaussian representation framework, σ models feature-level uncertainty that may reflect inter-subject variability. As illustrated in Fig. 1(b), σ exhibits more pronounced variations across subjects than μ, highlighting its sensitivity to subject-dependent feature differences. The w/o RGL ablation in Appendix B.1 (Table 5) demonstrates the contribution of reliable graph learning. To further isolate the role of σ, we introduce a μ-only graph variant under the same architecture and training settings. As shown in Table R5, UGGE consistently outperforms this variant across all four settings, with accuracy gains of 0.90% to 1.7%.

Table R5. Graph construction comparison (Acc.% ± SD).

GraphDTU 1sDTU 2sAV-GC 1sAV-GC 2s
Random Graph58.0 ± 3.8658.4 ± 3.5959.5 ± 7.7859.6 ± 8.86
μ\mu-only57.9 ± 3.9958.1 ± 4.1258.9 ± 7.3458.8 ± 9.25
UGGE58.8 ± 3.7359.2 ± 3.4460.6 ± 7.5860.4 ± 8.68
AuthorsOpenReview ↗

Response - Part3

6. Notation, Classification Features, and Uncertainty Updates (W5)

We thank the reviewer for identifying these presentation inconsistencies. We clarify the following points and have revised the corresponding descriptions for consistency:

  • SA in Table 3: SA denotes the subject-invariance learning module, which employs an adversarial discriminator to suppress subject-specific information and encourage subject-invariant representations. Its objective is defined in Eq. (12), with details provided in L270–284 and the module architecture illustrated in the lower-right part of Fig. 2.

  • Classification features (Eq. 14): The classification loss in Eq. (14) is applied to zz, rather than fmf_m. The final linear classifier is trained jointly with the feature extractor through Lcls\mathcal{L}_{\mathrm{cls}}. We have corrected Eq. (14) and the corresponding formulations to ensure consistency throughout the manuscript. The labeled adaptation trials are used only for checkpoint selection, not classifier training (see our response to W1 & Q1 for details).

  • Uncertainty updates (Eq. 8): We apologize for the ambiguity in our description. As shown in Eq. (8), only μ\mu is directly refined by the GCN, while σ\sigma characterizes uncertainty and determines node reliability for graph construction, thereby influencing the adjacency matrix and subsequent updates of μ\mu. Both Gaussian parameters remain trainable through end-to-end optimization with KL regularization. We have revised the description of σ^\hat{\sigma} to distinguish parameter learning from graph refinement.

Concluding Remarks

We sincerely thank the reviewer again for the encouraging assessment and constructive feedback. We hope that the additional graph construction analysis, clarified experimental protocol, and subject-wise statistical comparisons address your main concerns and provide stronger evidence for the effectiveness and reliability of UGPA-AAD.

Reviewer_X7LoOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes UGPA-AAD for cross-subject auditory attention decoding from EEG under leave-one-subject-out evaluation. A multi-scale dilated convolution encoder feeds an uncertainty-aware graph encoder. Each channel feature is mapped to a Gaussian with a mean and a log-variance. Channel affinities come from a Wasserstein-style distance between these Gaussians. They are scaled by a reliability term r=\exp(-\overline{\sigma^2}) and sparsified with top-k. A GCN refines the means. A prototype module keeps two global class prototypes and applies a contrastive loss on source embeddings. It adds a gradient-reversal subject discriminator on the source subjects. It also applies a pseudo-label prototype loss on the unlabeled target subject. Experiments cover DTU, AV-GC, MM-AAD and AVED, with 1 and 2 second windows. The method gains 1 to 4 points over seven reproduced baselines. Ablations, t-SNE plots, topographic maps and connectivity plots follow.

Critical strengths

Four datasets and both audio-only and audio-visual conditions. The set of settings is wide. The ablations cover each module and each loss, and show the choices made.

Critical weaknesses

The comparison is not fair. The method trains on the target subject’s unlabeled windows through \mathcal{L}{\mathrm{tpcl}}. The evaluation text says all methods share preprocessing and splits. It does not say that baselines see target data. They look like plain source-only models. The target adaptation loss is the largest single loss effect in Table 7. Without the whole prototype module, DTU accuracy is 57.1 and 58.0. S2Former reports 57.7 and 59.7. So the spatial design alone does not beat a baseline on DTU. Whatever margin exists seems to come from access to target data. No adaptation baseline is run. DANN, CORAL, pseudo-labelling and prototype alignment on top of the baselines are all missing. It is also unclear if the unlabeled adaptation windows are the same windows that get scored. If they are, this is transductive on the test set. The gains are small and no statistics are given. Absolute accuracy is 54 to 61 percent, close to chance. Gains are about 1 point on DTU, with subject SD of 3.7 to 5.4. On AV-GC there are 10 subjects and SD of 7.6 to 8.7. No paired test over subjects, no seeds, no confidence intervals. Many ablation steps are below 1 point. Hyperparameters were chosen by LOSO test accuracy. Fig. 5 picks the layer count and Table 8 picks the kernel set. The threshold, loss weights and stopping epoch are not described. No validation split is described. Baselines are reproduced by the authors, and some score at or below chance. DBPNet gets 48.8, DARNet 49.1 and DHGCN 50.0. Poor tuning is likely. The claims about uncertainty and prototypes are not tested. \sigma is predicted from the same features and regularised toward a unit Gaussian. Nothing shows it tracks channel noise or subject variation. No plot of \sigma or of r is given. Fig. 7 shows edge weights from 0.010 to 0.016 on DTU and 0.0152 to 0.0158 on AVED. The strongest edges are nearly equal. The graph may be close to flat, which would explain the small gain from the graph module. The claim that \mu is subject-invariant is not measured. The subject discriminator is trained on source subjects only, so it does not align the target. Removing only \mathcal{L}{\mathrm{tpcl}} gives 56.2 and 56.9 on DTU. Removing the whole prototype module gives 57.1 and 58.0. Removing more gives a better score. The target loss may hurt without its partners, which suggests it depends on pseudo-label quality. No pseudo-label accuracy is reported. With two classes, two prototypes act as a cosine classifier. The compactness story rests on t-SNE and topographic maps. Fig. 3 gives no scale and no quantitative consistency measure. The learned maps are an average over feature dimensions with an unstated procedure. The text links long-range edges to temporal and frontoparietal sources. Scalp EEG cannot support this reading, and edges depend on noise. The method description does not match itself. The implementation section lists a target entropy loss with weight 0.01. Eq. 15 has no such term. \mathcal{L}{\mathrm{cls}} in Eq. 14 acts on f_m. The final prediction comes from z through a linear classifier. How this classifier is trained is not stated. The confidence p(c|z^t) used for pseudo labels is not defined. Eq. 9 samples noise. It is not stated whether inference is stochastic. Eq. 1 sums over K minus 1 source subjects, while the text says K are used. Table 3 and Table 7 differ on AV-GC 2 second without \mathcal{L}{\mathrm{dom}}, 59.6 against 59.5.

Questions for authors

Do baselines see the target subject’s unlabeled data? If not, please run baselines with DANN, CORAL and pseudo-label self-training under the same access. Are the adaptation windows the same as the scored windows? Please report a version that adapts on one half of the target data and tests on a held-out half. How were thresholds, loss weights, layer count, kernel set and stopping epoch chosen? On which split? What is the pseudo-label accuracy over training? Why does removing \mathcal{L}_{\mathrm{tpcl}} alone give lower accuracy than removing the whole prototype module? What do per-subject paired tests, and several seeds, show for the main table and each ablation? What is the distribution of \sigma and of the edge weights? Does \sigma correlate with channel noise level? What happens with a random graph, a fixed uniform graph, and a graph from \mu only? Where does the target entropy loss with weight 0.01 enter the objective? How is the final classifier trained? Is inference stochastic? How were baselines tuned? Why do several score below chance? Please compare with published cross-subject numbers on the same datasets. How are the maps in Fig. 3 produced from D-dimensional features? Please give a quantitative measure of cross-subject consistency.

4 comments

The author has provided no further response, I am maintaining my score.

AuthorsOpenReview ↗

Response - Part 1

We appreciate your time and thank you for the careful assessment of our experimental protocol and method details. Below we address the weaknesses and questions with additional analysis and clarifications that we will incorporate into the revised version.

1. Data Access and Experimental Fairness

We respectfully clarify that UGPA-AAD and the baselines use the same training-data pool consisting of K non-test subjects in each outer fold, and the held-out test subject is reserved exclusively for final evaluation. As defined in Sec. 3.1, each fold contains K+1 subjects: K−1 source subjects, one adaptation subject, and one test subject (K−1:1:1). The adaptation subject is randomly selected from the K non-test subjects, with 60% of its trials used for unlabeled adaptation and 40% for labeled validation. The held-out test subject is never used for training, adaptation, validation, or model selection. We have clarified these important details and the ambiguous use of "target subject" in the revised manuscript, with changes highlighted in blue.

Importantly, the baselines use labeled data from all K non-test subjects, whereas UGPA-AAD uses K−1 labeled source subjects and one subject for adaptation and validation. TPCL accesses only the unlabeled adaptation trials from this non-test subject. Therefore, our gains cannot simply be attributed to additional data access or greater training-label supervision. To further assess fairness, we additionally compare DANN, Deep CORAL, pseudo-label self-training (PL), and prototype alignment (PA) under matched source, adaptation, and validation partitions (Table R1a).

Table R1a. Adaptation comparison under matched data access (Acc.% ± SD).

MethodDTU1sDTU2sAV-GC1sAV-GC2s
S²M-Former57.7 ± 4.2159.7 ± 5.2554.0 ± 3.7853.2 ± 3.89
S²M-Former+DANN52.8 ± 8.4354.6 ± 7.9251.7 ± 6.8150.9 ± 8.17
S²M-Former+CORAL54.3 ± 9.2655.1 ± 8.7450.8 ± 7.6352.4 ± 7.19
S²M-Former+PL53.6 ± 7.5855.9 ± 6.9354.6 ± 5.8254.1 ± 6.47
S²M-Former+PA55.2 ± 8.1753.8 ± 9.1251.4 ± 8.3652.3 ± 7.84
ListenNet56.8 ± 7.3257.2 ± 5.8353.2 ± 4.9055.0 ± 7.06
ListenNet+DANN51.2 ± 9.0651.9 ± 8.9750.4 ± 7.1250.0 ± 8.09
ListenNet+CORAL53.1 ± 10.1753.2 ± 9.7949.9 ± 8.2250.1 ± 7.81
ListenNet+PL50.5 ± 9.6051.1 ± 9.4552.4 ± 5.4753.0 ± 5.06
ListenNet+PA52.1 ± 9.1653.2 ± 9.8149.8 ± 8.9350.2 ± 8.77
UGPA-AAD w/o TPCL56.2 ±4.1556.9 ±3.8157.8 ±8.1657.6 ±9.18
UGPA-AAD58.8 ± 3.7359.2 ± 3.4460.6 ± 7.5860.4 ± 8.68

Our original evaluation does not involve transductive adaptation on test data. We additionally introduce a separate calibration experiment using half of the test subject's trials for unlabeled adaptation and the remaining half for evaluation. All methods share identical calibration and evaluation partitions (Table R1b).

Table R1b. Comparison under held-out trial calibration (Acc.% ± SD).

BackboneCalibrationDTU1sDTU2sAV-GC1sAV-GC2s
S²M-FormerNone58.4 ± 5.7760.0 ± 5.9856.2± 4.5357.1 ± 4.99
S²M-FormerDANN60.8 ± 6.7462.3 ± 6.1858.6 ± 5.9259.2 ± 6.47
S²M-FormerDeep CORAL62.1 ± 7.1663.5 ± 6.8259.4 ± 6.3560.7 ± 5.83
S²M-FormerPseudo-labeling63.4 ± 5.8764.2 ± 6.3161.8 ± 5.4662.3 ± 6.12
S²M-FormerPrototype alignment64.7 ± 5.3965.1 ± 5.7463.2 ± 5.8162.8 ± 5.26
ListenNetNone60.2 ± 6.8161.1 ± 6.2758.7 ± 5.9459.3 ± 6.52
ListenNetDANN58.6 ± 8.1359.4 ± 7.8956.8 ± 7.2457.5 ± 7.61
ListenNetDeep CORAL59.7 ± 8.6260.3 ± 8.1457.4 ± 7.8358.6 ± 7.42
ListenNetPseudo-labeling61.5 ± 7.3862.7 ± 7.1660.4 ± 6.6961.2 ± 6.84
ListenNetPrototype alignment62.3 ± 6.9263.1 ± 7.3461.6 ± 6.4860.8 ± 7.12
UGPA-AADTPCL off63.8 ± 4.8664.5 ± 5.1265.2 ± 6.7364.7 ± 7.16
UGPA-AADTPCL on67.2 ± 4.3568.1 ± 4.6768.4 ± 6.2169.3 ± 6.58

The improvements are not solely driven by target adaptation. Table 3 shows that removing UGGE results in accuracy decreases of 2.4% to 2.9% across the four settings, compared with 1.2% to 2.1% when removing PCSL. Table 6 further demonstrates UGGE's contribution under controlled spatial modeling comparisons.

AuthorsOpenReview ↗

Response - Part 2

2. Training Strategy

Several questioned hyperparameters are already specified in Sec. 4.1, including the confidence threshold (0.95 to 0.90), prototype temperature (0.1), temporal kernels ({1,3,5,7}), GCN depth (3), and maximum training epochs (100). Loss coefficients are also reported, subject to the terminology correction below.

No test-subject data or test accuracy was used for hyperparameter tuning or checkpoint selection. Hyperparameters were determined using training data. In each LOSO fold, the optimal checkpoint was selected based on the lowest supervised validation loss on the 40% labeled trials of the non-test adaptation subject, while the remaining 60% unlabeled trials were used for TPCL adaptation. The held-out test subject was evaluated only after checkpoint selection. Appendix B.2 (Table 8 and Fig. 5) further examines architectural sensitivity. We have clarified the model-selection procedure in the revised manuscript.

All baselines follow their original architectures and recommended hyperparameter configurations under identical preprocessing and evaluation protocols. Performance near chance can occur under challenging short-window cross-subject EEG decoding and does not necessarily indicate improper implementation. We have clarified the reproduction settings and provide comparisons with published results under directly comparable protocols.

3. Prototype Learning

PCSL requires coordinated source semantic structuring and target adaptation, rather than simply imposing additional prototype constraints. SPAL establishes shared source class structures, while TPCL aligns confident target embeddings. Removing TPCL retains source prototype constraints without corresponding target alignment, whereas removing PCSL eliminates these constraints altogether. This provides a plausible explanation for the observed performance difference. Importantly, Tables 3 and 7 show that the complete framework consistently outperforms both variants.

Unlike a simple cosine classifier, the proposed framework integrates subject-balanced prototypes, source-subject adversarial learning, and confidence-aware target adaptation. The discriminator suppresses source-subject information, while TPCL explicitly aligns target semantics. Fig. 1(b) further provides qualitative evidence of more consistent cross-subject patterns in μ\mu.

4. Performance Reliability

Cross-subject AAD remains challenging due to substantial inter-subject EEG variability. Nevertheless, UGPA-AAD outperforms the strongest baselines by 2.9/2.6 percentage points on MM-AAD (audio-only) and 2.7/3.8 percentage points on AV-GC. We further conduct experiments using three random seeds, demonstrating consistent performance across datasets.

Table R3. Performance across random seeds on DTU and MM-AAD (Acc.% ± SD).

SeedDTU 1 sDTU 2 sMM-AAD AO 1 sMM-AAD AO 2 sMM-AAD AV 1 sMM-AAD AV 2 s
4258.8 ± 3.7359.2 ± 3.4458.1 ± 5.4258.5 ± 5.1857.0 ± 4.3557.2 ± 4.26
102458.4 ± 4.3659.0 ± 3.8558.2 ± 5.1358.7 ± 5.2957.4 ± 4.9757.5 ± 4.33
340759.1 ± 3.8259.6 ± 3.1757.9 ± 4.9158.5 ± 5.3657.6 ± 4.2357.6 ± 3.90
Mean58.7759.2758.0758.5757.3357.43

Table R4. Performance across random seeds on AVED and AV-GC (Acc.% ± SD).

SeedAVED AO 1 sAVED AO 2 sAVED AV 1 sAVED AV 2 sAV-GC 1 sAV-GC 2 s
4254.1 ± 1.2554.0 ± 1.4457.0 ± 2.7457.2 ± 3.4360.6 ± 7.5860.4 ± 8.68
102454.4 ± 2.3154.2 ± 1.3656.8 ± 2.9657.3 ± 3.3360.0 ± 7.1960.0 ± 8.31
340753.7 ± 1.3853.7 ± 1.2257.0 ± 2.7056.9 ± 3.1159.7 ± 6.1459.9 ± 7.57
Mean54.0753.9756.9357.1360.1060.10

The small variations across random seeds have no substantial impact on the overall performance, indicating the stability of UGPA-AAD.

AuthorsOpenReview ↗

Response - Part 3

5. Uncertainty-Aware Graph Learning

Uncertainty directly guides spatial dependency learning, as formulated in Eqs. (4–8). Fig. 1(b) qualitatively illustrates more consistent cross-subject patterns in μ and stronger subject-dependent variations in σ, consistent with our Gaussian modeling design.

The narrow edge-weight ranges in Fig. 7 reflect softmax normalization and do not necessarily imply uniform connectivity. Existing ablations in Tables 5–6 support UGGE's effectiveness. To further isolate the contribution of uncertainty modeling, we additionally evaluate graph construction using μ\mu alone. As shown in Table R5, UGGE consistently outperforms both random graph construction and the μ\mu-only variant, demonstrating the effectiveness of our uncertainty-aware graph design.

Table R5. Graph construction comparison (Acc.% ± SD).

GraphDTU 1sDTU 2sAV-GC 1sAV-GC 2s
Random Graph58.0±3.8658.4±3.5959.5±7.7859.6±8.86
μ-only57.9±3.9958.1±4.1258.9±7.3458.8±9.25
UGGE58.8±3.7359.2±3.4460.6±7.5860.4±8.68

Fig. 3 is generated by averaging absolute feature magnitudes across feature and temporal dimensions for each EEG channel, followed by MNE Topomap visualization. These maps provide qualitative sensor-space observations rather than quantitative evidence of subject invariance or cortical source connectivity. We have revised the corresponding interpretation to avoid overstatement.

6. Implementation Clarifications

We thank the reviewer for identifying these inconsistencies. We have carefully reviewed the corresponding formulations and implementation details and made the following clarifications and corrections, with all revisions highlighted in blue:

  • Target entropy loss: The term "target entropy loss" is an incorrect reference to the existing Ltpcl\mathcal{L}_{tpcl}, rather than an additional loss. Its weight is 0.01, compared with 0.1 for other auxiliary losses and 1.0 for classification.

  • Classifier and pseudo-label probabilities: The classification loss in Eq. (14) is applied to zz, rather than fmf_m. The final linear classifier is trained jointly with the feature extractor through Lcls\mathcal{L}_{\mathrm{cls}}. The pseudo-label probabilities p(c∣zit)p(c|z_i^t) are computed via softmax over the similarities between zitz_i^t and class prototypes pc\mathbf{p}_c:

p(c∣zit)=π(c∣zit)=Softmax⁡c(sim⁡(zit,pc)τp).p(c|z_i^t)=\pi(c|z_i^t)=\operatorname{Softmax}_c\left(\frac{\operatorname{sim}(z_i^t,\mathbf{p}_c)}{\tau_p}\right).

We have corrected Eq. (14) and the corresponding formulations to ensure consistency throughout the manuscript.

  • Inference procedure: Eq. (9) employs the reparameterization trick during training to incorporate uncertainty into spatial representation learning, enhancing robustness. To ensure stable and reliable feature representations, we use only the refined mean μ^\hat{\mu} as the reliable feature representation during inference. Thus, inference is deterministic and does not involve random sampling or repeated prediction averaging. We have clarified this distinction in the revised manuscript.

  • Source-subject notation: We have corrected the source-subject summation and corresponding description to consistently reflect the K−1:1:1K-1:1:1 protocol, comprising K−1K-1 source subjects, one adaptation subject, and one held-out test subject. Thus, KK denotes the total number of non-test subjects, rather than the number of source subjects alone.

  • AV-GC 2 s discrepancy: We apologize for the inconsistency between Tables 3 and 7. After double-checking the original experimental logs, we confirm that the correct accuracy is 59.6%. The 59.5% in the supplementary material is a transcription error and has been corrected.

We apologize for these presentation inconsistencies and appreciate the reviewer's careful examination.

Concluding Remarks

We sincerely thank the reviewer for the constructive feedback. We hope these corrections, clarifications, and supplementary analyses adequately address your core concerns and provide stronger support for the effectiveness of UGGE and PCSL.

Reviewer_znCWOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The main goal of this article is to explore how to make the auditory attention decoding model trained on one group of subjects better adapt to new subjects. The authors combines multi-scale feature extraction, Gaussian graph learning, and prototype alignment to adjust the connections between EEG channels on the one hand, and to utilize the unlabeled data of new subjects on the other hand, making it easier for representations of the same class across subjects to align. The experimental results on four datasets are generally positive.

Critical strengths

I think the strengths of this work are:

  1. The problem and motivation are clearly stated. The EEG of different individuals varies, which indeed affects model transfer. The article addresses this by focusing on two aspects: spatial relationships and category representations. The approach is relatively straightforward in relation to the problem it aims to solve.

  2. The components fit together well. From multi-scale feature extraction to graph modeling between channels, and then to cross-subject prototype alignment, there are relatively natural connections between each part. Among them, the idea of adjusting channel connections using variance is also quite interesting.

  3. The experiments are quite comprehensive. The authors tested four datasets and also considered different scenarios and time windows. Under most settings, performance improvements could be observed, which strengthened confidence in the effectiveness of the method.

  4. The ablation study was conducted in detail. The main modules, training losses, and some parameters have corresponding analyses, which can help readers understand which designs contribute to the results and provide support for the overall scheme.

Critical weaknesses
  1. The explanation of the roles of mean and variance could be supported with more direct evidence.

Modeling channel features with Gaussian representations and adjusting connection strengths via variance is a reasonable approach. However, interpreting the mean as subject-invariant features and the variance as subject-related uncertainty seems to still remain an assumption at present. While existing ablation studies support the module's effectiveness, further demonstration of how variance and connection weights change with channel perturbations may make this interpretation more compelling.

  1. Some of the methods and experimental details can be elaborated further.

The paper has explicitly adopted the LOSO evaluation protocol. It is recommended to supplement the selection method for the final model within each fold, including whether to use a fixed number of training epochs or to select the model based on the validation set; if using the validation set, please also explain its construction method. Additionally, it is hoped to clarify whether the unlabeled target is adapted to use the entire evaluation samples of the left-out subjects, or a separately divided subset. In the method description, the handling method of variance in the graph propagation process, as well as the corresponding relationship between "target entropy loss" and Equation (15) in the implementation details, are also recommended to be further explained. These are mainly issues related to description completeness, and their supplementation will make it more convenient for readers to understand and reproduce.

  1. The stability of performance improvement could be further elaborated.

The method achieved higher average accuracy across most settings, but the improvement in some cases was modest. In addition to the current subject-averaged mean and standard deviation, providing subject-wise paired performance differences or corresponding confidence intervals would make it easier to assess the stability of the gains and further support the conclusion of overall performance advantage.

  1. It is suggested that a further review of the spelling, layout, cross-references and consistency of symbol naming throughout the text be conducted. At the same time, the annotations for the illustrations should be improved and the readability of the legends enhanced to make the overall presentation clearer.
Questions for authors

Basically, the primary question has been stated in the previous weakness section, which is the evidence for the different role of mean and variance, this part confuses me most. Sometimes we believe the statistical metrics have different meanings, but I think we still need evidence, or proof to support this. Other questions have also been posted in the weakness section.

2 comments
AuthorsOpenReview ↗

Response - Part 1

We sincerely thank the reviewer for the positive assessment of our motivation, methodological design, and experimental evaluation. We particularly appreciate the constructive suggestions regarding the interpretation of Gaussian representations and the statistical reliability of performance improvements. We address these concerns below.

1. Evidence for the Roles of Mean and Variance (Main question & W1)

We appreciate the reviewer's insightful question. Our design aims to learn relatively subject-invariant representations through μ\mu, while using σ\sigma to characterize feature uncertainty and guide reliable spatial dependency modeling. The LclsL_{cls} and LdomL_{dom} encourage discriminative and subject-invariant representations, while KL regularization constrains the Gaussian distribution and σ\sigma guides uncertainty-aware graph construction.

First, Fig. 1(b) provides qualitative evidence of their distinct characteristics. We randomly selected EEG samples from four subjects in the DTU dataset and visualized the learned distributions of μ\mu and σ\sigma. As illustrated, μ\mu exhibits relatively consistent feature patterns across subjects, whereas σ\sigma shows more pronounced inter-subject variations. These observations are consistent with our intended Gaussian representation design, although the visualization alone does not constitute formal proof of subject invariance.

Second, uncertainty directly contributes to graph construction, as formulated in Eqs. (4–8). The estimated uncertainty σ\sigma guides channel reliability and influences spatial connection weights, while the mean representations μ\mu are progressively refined through graph learning. To further verify this contribution, we additionally compare UGGE with random graph construction and a μ\mu-only variant that excludes uncertainty-guided graph construction.

Table R1. Graph construction comparison (Acc.% ± SD).

GraphDTU 1sDTU 2sAV-GC 1sAV-GC 2s
Random Graph58.0 ± 3.8658.4 ± 3.5959.5 ± 7.7859.6 ± 8.86
μ\mu-only57.9 ± 3.9958.1 ± 4.1258.9 ± 7.3458.8 ± 9.25
UGGE58.8 ± 3.7359.2 ± 3.4460.6 ± 7.5860.4 ± 8.68

As shown in Table R1, UGGE consistently outperforms both random graph construction and the μ\mu-only variant across all four settings, further supporting the effectiveness of uncertainty-guided spatial dependency modeling.

2. Experimental Protocol and Implementation Clarifications (W2)

We appreciate the reviewer for highlighting these implementation details and clarify them as follows.

  • Data access and LOSO protocol: As defined in Sec. 3.1 (L141), each fold contains K+1K+1 subjects: K−1K-1 source subjects, one adaptation subject, and one held-out test subject (K−1:1:1K-1:1:1). The adaptation subject is selected from the KK non-test subjects, with 60% of its trials used for unlabeled TPCL adaptation and 40% for labeled validation. The held-out test subject is never accessed during training, adaptation, validation, or model selection. Thus, our original evaluation does not involve transductive adaptation on test samples.

  • Hyperparameter and checkpoint selection: No test-subject data or test accuracy was used for hyperparameter tuning or checkpoint selection. Hyperparameters were determined using training data, with key settings specified in Sec. 4.1. In each LOSO fold, the optimal checkpoint was selected based on the lowest supervised validation loss on the 40% labeled trials of the non-test adaptation subject. The held-out test subject is evaluated only after checkpoint selection.

  • Variance refinement (Eq. 8): We apologize for the ambiguity in our description. As shown in Eq. (8), only μ\mu is directly refined by the GCN, while σ\sigma characterizes uncertainty and determines node reliability for graph construction, thereby influencing the adjacency matrix and subsequent updates of μ\mu. Both Gaussian parameters remain trainable through end-to-end optimization with KL regularization. We have revised the manuscript to clarify that σ\sigma participates in uncertainty-guided graph construction rather than direct GCN refinement.

  • Target entropy loss: The term "target entropy loss" in the implementation details incorrectly refers to the existing Ltpcl\mathcal{L}_{\mathrm{tpcl}}, rather than an additional loss. Its weight is 0.01, compared with 0.1 for other auxiliary losses and 1.0 for classification. We have corrected the terminology to ensure consistency with Eq. (15).

We have clarified these implementation details in the revised manuscript, with the corresponding changes highlighted in blue.

AuthorsOpenReview ↗

Response - Part 2

3. Stability of Performance Improvements (W3)

We appreciate this valuable suggestion. To further evaluate the consistency of performance improvements, we conduct subject-wise paired comparisons on three representative datasets: DTU, AV-GC, and AVED-AV. We report the mean paired accuracy differences (Δ\DeltaAcc), 95% confidence intervals, and two-sided Wilcoxon signed-rank test results.

Table R2. Subject-wise paired performance comparison between UGPA-AAD and the strongest baselines. (Δ\DeltaAcc in percentage points; 95% CI of paired differences; unadjusted two-sided Wilcoxon p-values; improved subjects as n/N.)

DatasetWindowBaselineΔAcc (pp)95% CI (pp)Wilcoxon pImproved Subjects
DTU1sS²M-Former+1.11[0.08, 2.14]0.035913/18
DTU2sS²M-Former-0.50[-1.98, 0.97]0.79278/18
AV-GC1sPCRNet+2.74[-2.22, 7.71]0.32237/10
AV-GC2sMHANet+3.82[-1.02, 8.66]0.16027/10
AVED AV1sListenNet+3.22[1.12, 5.31]0.00599/10
AVED AV2sListenNet+3.18[0.42, 5.94]0.01379/10

As shown in Table R2, UGPA-AAD achieves positive mean paired improvements in five of the six settings. In particular, DTU 1s and both AVED AV settings exhibit positive 95% confidence intervals excluding zero, with unadjusted two-sided Wilcoxon p-values below 0.05. Furthermore, performance improvements are observed for 13/18 DTU subjects at 1s, 7/10 AV-GC subjects at both window lengths, and 9/10 AVED-AV subjects at both window lengths.

Although the gains vary across datasets and are not statistically significant in every setting, these paired analyses provide more direct evidence of the performance advantages of UGPA-AAD while revealing the variability of improvements across subjects and datasets.

4. Presentation and Notation Consistency (W4)

We thank the reviewer for these helpful suggestions. We have carefully reviewed the manuscript and corrected typographical errors, inconsistent mathematical notation, and cross-references. We have also revised the figure annotations and legends to improve readability and clarify the corresponding methodological interpretations. All revisions are highlighted in blue.

Concluding Remarks

We sincerely thank the reviewer again for the encouraging assessment and constructive feedback. We hope that the additional graph construction analysis, clarified experimental protocol, and subject-wise statistical comparisons address your main concerns and provide stronger evidence for the effectiveness and reliability of UGPA-AAD.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.