acceptodds
Withdrawn from ICLR 2027

DGSIN: Discrepancy-Gated Source–Instance Normalization for Cross-Domain Time-Series Semantic Segmentation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Cross-domain time-series semantic segmentation aims to label temporal segments in unseen domains. Representative approaches use feature alignment or statistical normalization to reduce feature discrepancies across domains. However, these discrepancies can reflect both domain factors and semantic variation, suppressing them may also weaken the features that help distinguish semantic states. Normalization methods in visual recognition learn a trade-off between batch and instance normalization, but their input-independent mixing coefficients do not explicitly respond to changes between temporal segments.We propose discrepancy-gated source–instance normalization (DGSIN), which learns segment-dependent trade-offs through the segmentation objective. DGSIN uses source–instance statistical discrepancies to generate layer–channel coefficients that control the correction from source-normalized to instance-normalized representations. At inference, a shared network averages prediction probabilities under different frozen source statistics. Across five leave-one-dataset-out sleep-staging tasks, DGSIN achieves 72.43% mean macro-F1, 74.30% balanced accuracy, and 0.688 Cohen’s κ. Its mean macro-F1 exceeds DSON’s 72.14% and PSDNorm’s 71.47%, with the highest macro-F1 on four targets. DGSIN provides a concrete approach to learning segment-dependent source–instance normalization trade-offs for cross-domain temporal segmentation.Source code is available at https://anonymous.4open.science/r/Dgsin-E9DB.

Primary area:
learning on time series and dynamical systems
Keywords:
Domain Generalization, Time-Series Semantic Segmentation, Adaptive Normalization, Source–Instance Normalization, Sleep Staging

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Fpo6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DGSIN, an adaptive normalization mechanism that uses the discrepancy between the current segment statistics and source-domain statistics to generate source-, layer-, and channel-specific normalization coefficients. The resulting representation dynamically interpolates between source-normalized and instance-normalized features, while inference averages predictions from multiple source-conditioned branches.

Critical strengths
  • The motivation is reasonably strong. The diagnostic experiments suggest that statistical discrepancies may encode both domain-related and task-relevant semantic information, so aggressively minimizing domain discrepancy is not necessarily beneficial for target-domain prediction.
  • The matched comparisons among coarse-static, fine-static, and discrepancy-conditioned variants are useful for separating the effects of parameter granularity from those of input-dependent conditioning.
  • The evaluation covers five leave-one-dataset-out sleep-staging settings and includes analyses of normalization behavior, perturbation robustness, and source-conditioned prediction fusion.
  • The proposed module is conceptually simple and can be integrated into the underlying segmentation architecture without introducing an additional adaptation stage at test time.
Critical weaknesses
  • The methodological contribution is relatively incremental. DSON (ECCV 2020) already studies domain-specific trade-offs between batch and instance statistics, while PSDNorm (ICLR 2026) considers normalization specifically for cross-domain sleep staging. Domain-generalizable sleep staging has also been extensively studied, for example by SIDA (TIST 2024). DGSIN mainly extends this line by making the source–instance trade-off fine-grained and input-dependent through a discrepancy-conditioned gate, which is useful but represents a relatively modest conceptual step.
  • Compared with the matched fine-static control, the mean macro-F1 improvement is only around 0.6 percentage points, making the practical benefit of dynamic discrepancy conditioning relatively limited.
  • A substantial portion of the final performance gain also comes from multi-source probability fusion rather than the dynamic normalization mechanism itself.
  • The four-branch source-conditioned inference incurs roughly four times the inference latency, which makes the cost–accuracy trade-off less compelling.
  • The source train/validation split does not enforce subject- or recording-level disjointness, and each target configuration is trained only once. Given the small absolute performance gains, stronger control of training variance would be important.
Questions for authors

Please see Weaknesses

Reviewer_h7WtOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes dynamically adjusting the balance between source and instance normalization based on segment-level feature-statistics discrepancies, demonstrating modest average improvements over static gating and existing normalization methods across five cross-dataset sleep-staging tasks.

Critical strengths
  1. A clear and lightweight formulation of source-conditioned dynamic normalization. The method learns the adjustment strength through the supervised task objective and explicitly connects source-reference statistics with instance statistics. This provides a concrete, testable mechanism for learning normalization trade-offs without relying on discrepancy reduction as a surrogate objective.

  2. A meaningful matched comparison between dynamic and static gating, with transparent reporting of limitations. Table 4 controls the backbone, task objective, and main inference configuration through independently retrained variants. The appendix also uses paired cluster resampling and reports negative results on HMC and under strong white noise. These experiments provide substantive empirical support while acknowledging the method’s operating limits.

Critical weaknesses
  1. The separation between method development and unseen-domain evaluation is insufficiently established. Section 5 acknowledges that benchmark results informed architecture development, but does not specify whether this included target-domain performance or which decisions were affected. If target results influenced architecture or hyperparameter choices, source-only checkpoint selection cannot establish an independent final evaluation. Moreover, the main experiments split source data by 20-epoch blocks, allowing the same recording or subject to appear in both training and validation. This weakens validation as a basis for selecting models that generalize beyond observed recordings. Such overlap does not itself constitute target-domain leakage, but its impact requires evaluation.

  2. The modest gains have not been validated across repeated training runs. The average improvements are approximately 0.46 percentage points over DSON and 0.64 points over the matched fine-static control, while Appendix D states that each configuration is trained only once per target. The confidence intervals in Appendix E quantify evaluation-sample uncertainty for fixed checkpoints; they do not establish stability across initialization and optimization randomness. Corresponding uncertainty estimates for the comparison with DSON are also missing.

  3. The comparison with the strongest baseline mixes normalization and aggregation effects. Appendix H states that DSON averages logits, whereas DGSIN averages softmax probabilities. The DSON implementation follows its original inference procedure, so this is not evidence of an incorrect reproduction. Nevertheless, given the small performance gap, exchanging the aggregation rules is necessary to determine how much of the ranking is attributable to the normalization design. This concern does not directly invalidate the matched dynamic-versus-static comparison in Table 4.

  4. The discrepancy descriptor and monotonic response assumption lack targeted validation. Table 4 shows that the proposed dynamic formulation outperforms a static counterpart, but does not establish why mean and scale discrepancies should be compressed into a single scalar, or why larger discrepancies should monotonically increase reliance on instance normalization. The monotonic trend in Figure 3(a) follows from the parameterization and provides no independent validation of this assumption. The paper discusses related dynamic normalization methods, including DTIN, which also addresses the loss of discriminative information caused by instance normalization, but does not experimentally compare against a dynamic alternative.

  5. Several consequential data-processing details remain unspecified. Appendix B defines epoch-wise z-scoring but does not describe the recording-level preprocessing operations that require full-record statistics. Table 9 does not explain how the 150 SHHS1 participants and 150 P2018 recordings were selected. Removing invalid epochs can also create temporal gaps, yet neither the prevalence nor the impact of sequences spanning these gaps is reported. These are primarily reporting and validation gaps; the available evidence does not establish that they constitute methodological errors.

Questions for authors

The following questions concern the validity of the current conclusions rather than general future-work suggestions.

  1. Under recording- or subject-disjoint source splits, what are the paired multi-seed results for DGSIN, the fine-static control, and DSON? Do the two reported average advantages remain stable?

  2. When both logit averaging and probability averaging are applied to the same DSON and DGSIN checkpoints, do the target-level results and overall ranking change?

  3. How does performance change when the gating slope is unconstrained, or when mean and scale discrepancies are provided separately? What evidence supports a monotonic function of source–instance discrepancy over a more general dynamic gate?

Reviewer_65bqOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses domain shift in cross-domain time-series semantic segmentation and proposes DGSIN. The method uses the discrepancy between source statistics and current instance statistics to dynamically generate segment-dependent, source–layer–channel normalization mixing coefficients, and fuses multiple source-conditioned branches during inference.

Critical strengths
  1. The paper focuses on the fact that statistical discrepancies may contain both domain-related and semantic factors. The research motivation is relatively clear.

  2. The method uses source distribution information during training but does not require source labels at inference time, making it suitable for applications with unknown source-domain identities.

Critical weaknesses
  1. The experimental scope is limited. The paper is mainly evaluated on one type of time-series task, sleep staging, and five datasets, which is insufficient to support the claim that the method can generally address cross-domain time-series segmentation.

  2. Compared with the strongest baseline, DGSIN provides only a small performance improvement with limited evidence of stability. The paper does not provide sufficient repeated-training results, making it difficult to determine whether the improvement exceeds the variation caused by random initialization and the training process.

Questions for authors
  1. DGSIN does not outperform all normalization methods on target datasets such as HMC. Can the authors analyze which types of domain shift or data characteristics cause DGSIN to fail, and explain what its dynamic gate learns in these cases?

  2. How much would performance decrease if only one source-conditioned branch were used during inference, or if a shared branch with low-cost fusion were adopted? Is multi-branch inference necessary for DGSIN to obtain its improvement?

  3. Discrepancy gating essentially generates adaptive coefficients from the source–instance statistical discrepancy, and its difference from existing adaptive gating mechanisms seems to lie mainly in the choice of the gating input. Can the authors further clarify the substantial novelty of this design compared with existing methods?

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

Related papers

Discussion (0)

Sign in to comment.