acceptodds
Withdrawn from ICLR 2027

SAIL: Stable Anchoring with Individual-Conditioned Learning for Cross-Subject Time Series Classification

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Cross-subject classification of biomedical time series is challenging because the same task semantics can be expressed differently across individuals, causing models trained on seen subjects to generalize poorly to unseen subjects. Existing methods typically suppress subject-specific information to learn invariant representations, which may compromise task-relevant semantics. In this paper, we rethink the relationship between task semantics and individual variation, viewing semantic representations as shared task semantics shaped by individual-specific variation. We further find that cross-subject variation can be decomposed into reusable shift modes, which can be recombined to adapt semantic representations to unseen individuals. Based on these observations, we propose SAIL, which extracts a physiology-grounded semantic anchor and employs a mixture-of-experts (MoE) architecture to model reusable variation patterns. Individual context conditionally selects and combines these patterns, while feature-wise gating controls their contribution to the final representation. Across five benchmark medical datasets, SAIL improves the average macro-F1 over the lightweight TeCh baseline by 5.12% , while achieving approximately 9× faster inference.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
cross subject, time series classfication, medical time series

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 4 reviews, above 27% of the others.
Reviewer_GZZZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses cross-subject generalization in biomedical time-series classification by modeling individual variation as reusable conditional shifts. The authors introduce a subject-conditioned refinement framework that constructs a semantic anchor with individual-conditioned MoE adaptation. Experiments on seven datasets against 15 baselines demonstrate strong classification performance, supported by ablation and representation analyses.

Critical strengths

Interesting problem and clear presentation. The paper starts from a meaningful problem and tells a clear story. I particularly like how the authors use visualizations to explain their motivation and ideas. Overall, the paper is well written and easy to follow.

Strong experimental results. The authors have clearly put substantial effort into the experiments, covering seven datasets, 15 baselines, and several ablation studies. The results are convincing in showing that the full pipeline works well and is computationally efficient.

Critical weaknesses

The experiments do not fully support the original motivation. The authors provide extensive results showing that the proposed architecture works well, but it is unclear to me why the semantic anchor and individual context actually represent the claimed information, or how the learned residual captures reusable individual variation. The experiments mainly validate the architectural design, the overall pipeline, rather than establishing a clear connection to the original research question.

Unclear novelty and incomplete related work. I like the two motivating observations: task semantics and individual variation can overlap, and individual variation may contain reusable structure. However, these ideas are not entirely new. The related work mainly discusses approaches that suppress individual variation but does not sufficiently cover existing methods. Although the proposed architecture is well motivated, it is still unclear what the main scientific contribution is compared with prior individual-aware approaches.

Questions for authors

Representation decomposition. Why can the proposed decomposition be interpreted as separating task-related semantics from individual information? I do not see clear evidence that these representations actually correspond to the claimed physiological meanings. Can the authors justify this interpretation beyond showing better classification results of the overall pipeline?

Scientific contribution. What is the main scientific contribution of this work relative to prior individual-aware and subject-adaptive learning? I would appreciate a clearer discussion of related approaches that already model or leverage individual variation, and an explanation of what this work adds beyond them. This would help me better assess the novelty.

Reviewer_W1dkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper explores an alternative perspective by modeling time-series representations as shared task semantics modulated by individual-specific variations. Its motivation highlights that individual-specific information should not be treated merely as removable noise, and that individual information from others can be reused for unseen test users. Building on these insights, the paper proposes SAIL, a framework in which TSE extracts meaningful latent features, ACP decomposes them into a semantic anchor and complementary context, SCA aggregates subject-level information to facilitate its reuse, and CEM integrates these components into the final representation. Experiments across multiple datasets, comparisons with diverse baselines, and extensive ablation studies demonstrate the effectiveness of SAIL and support its potential for improving generalization in time-series tasks.

Critical strengths

The paper provides extensive experimental evidence, including analyses that try to empirically align with the motivation, comparisons with diverse baselines, ablation studies, and hyperparameter sensitivity analyses. The exhaustive results provided in the appendix further strengthen the empirical support for the proposed method and increase my confidence in the validity of the paper's findings.

Critical weaknesses

The motivation, method design, and experiments are not aligned.

Motivation 1 argues that task semantics and individual-specific variations share spectral support, suggesting that individual information cannot be simply treated as a removable nuisance. However, the proposed method separates these components using a frequency cutoff, and Figure 5 explicitly labels the corresponding regions as “Task-oriented” and “Subject / nuisance,” which appears inconsistent with the original motivation.

Furthermore, the ablation results show that the anchor-only variant (V7) outperforms most baselines on APAVA, TDBRAIN, and PTB. This raises questions about the necessity of modeling individual-specific variations and whether the proposed method's central design choices are sufficiently justified by the empirical evidence.

Questions for authors
  1. Could the authors provide more details about Figure 3, including the encoders used to obtain the F1 scores, the train/test split protocol, and the number of experimental runs?

  2. What are the key takeaways from the top two plots in Figure 1(b), given that both the subject and channel differ?

  3. The residual-only ablation (V4) achieves results comparable to those of the baselines. Could the authors clarify what information is retained in the residual representation and explain why it alone is sufficient to achieve comparable performance?

  4. Figure 9 shows that K=2 achieves the best results. Could the authors clarify whether this finding supports the necessity of the proposed mixture-of-experts structure, or whether a simpler architecture could achieve similar performance?

Reviewer_75jnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

SAIL classifies EEG and ECG signals from people not seen during training. It combines a low-frequency signal representation with a small correction based on subject information and predicted class. It leads on 6 of 7 datasets and runs about 9 times faster than prior work TeCh.

Critical strengths

The evaluation is broad: 15 baselines, subject-separated splits, overall solid. Performance seems good. SAIL scores 98.35 on TDBRAIN versus TeCh’s 92.38, and 90.65 on PTB versus BioFormer’s 84.71. It is also much faster: 6.4 ms versus 57.3 ms per batch (Table 16).

Critical weaknesses

My main issue is that most of the gain seems to come from the base representation, not the proposed correction. The base alone beats all baselines on APAVA, TDBRAIN, and PTB. The correction adds only 1.1–1.8 points there and 0.35 on PTB-XL. On PTB, it separates classes much better than subjects within each class (0.81 versus 0.07). Removing the predicted class hurts APAVA by 5.5 points, while subject prototypes add only 0.5. Table 2 also leaves out UCI-HAR, where three variants beat SAIL (Table 8).

Also, it will be nice to see how settings were chosen, especially with only four APAVA test subjects.

Questions for authors
  1. How were the Table 4 settings chosen, and on which split? Please add validation macro-F1 to Tables 9 and 10 and state the baselines' tuning and augmentation.
  2. We might need a new experiment. With trained weights fixed, zero the correction at inference and report macro-F1 on all seven datasets; Figure 12c suggests it lowers true-class probability on APAVA. Also recompute the routing affinities in Figures 6 and 10 within class.
Reviewer_SfKQOpenReview ↗
Rating
2(−1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work presents SAIL for cross‑subject biomedical time‑series classification. Different from prior methods that suppress subject‑specific signals, it observes individual variations contain reusable structured modes. SAIL builds a physiology‑inspired stable semantic anchor from spectral statistics, then uses a mixture‑of‑experts module conditioned on subject context to generate gated residual corrections for adapting representations to unseen persons. The authors conduct comprehensive ablations to verify each component’s function and analyze learned expert routing behaviour. Evaluated across seven EEG/ECG/inertial datasets under disjoint test‑subject settings, SAIL outperforms many strong baselines including TeCh and BioFormer, meanwhile delivering faster inference speed. Additional geometry analysis, hyperparameter sensitivity and efficiency profiling further back the design choices.

Critical strengths

Rather than following the common practice of suppressing subject‑specific signals via adversarial or disentanglement techniques, this paper puts forward a fresh viewpoint, i.e., subject‑level variation should be treated as reusable conditional modifiers for shared task semantics, not pure noise to be eliminated. The authors back this idea with concrete empirical probe studies, demonstrating that task‑relevant information and individual subject signatures co‑occur across frequency bands and that subject residuals possess reusable low‑rank structure.

SAIL is evaluated on seven diverse datasets against a large suite of competitive baselines, achieving state‑of‑the‑art macro‑F1 performance. Multiple orthogonal lines of evidence consistently validate how each design element contributes to final performance. This is consequential because it raises confidence that the performance gains originate from the proposed subject‑conditioned refinement mechanism instead of random tuning luck, and it demonstrates practical usability for real‑world biomedical deployment.

Critical weaknesses

SAIL builds subject prototypes by aggregating multiple samples belonging to the same test‑set subject. Even though performance only drops slightly for query‑only (batch‑size‑1) inference, the method still implicitly assumes access to multiple samples from the unseen test subject during evaluation. For real‑world cold‑start clinical scenarios where only a single sample per new subject is available, the model cannot reliably construct good subject‑context prototypes.

Important knobs including spectral cutoff ρ, expert count K need dataset‑wise tuning. Sensitivity experiments are only reported for one dataset (APAVA). The paper does not supply a heuristic or validation‑free recipe to choose these hyperparameters for a brand‑new biomedical dataset.

Questions for authors

SAIL relies on subject‑level prototypes aggregated over multiple samples from a test subject. Even though the B=1 (query‑only) setting yields small performance drops on reported datasets. Can you report results for a strict single‑sample cold‑start setting where absolutely no other samples from the test subject are available to build subject prototypes?

Hyper‑parameters spectral cutoff ρ and expert count K are tuned per‑dataset, and sensitivity analysis is only provided for the APAVA dataset. Can you provide sensitivity results for ρ and K on at‑least one additional held‑out dataset (e.g. PTB)?

Your router takes the anchor‑predicted task probability pbasep_{base} as one input for expert selection. What happens if the semantic anchor zbasez_{base} makes a wrong high‑confidence prediction on a test sample? Does this incorrect task hypothesis misdirect expert routing and further worsen the final prediction? Please provide quantitative metrics measuring this failure mode.

The paper demonstrates that experts learn distinct reusable correction directions, but does not quantify how much performance gain specifically comes from the MoE bank versus simpler fixed transformation banks. Can you compare SAIL against a baseline that replaces the learned expert bank with fixed, non‑learned correction directions while keeping the full routing‑gating‑residual pipeline unchanged?

Comments on the paper

1 comment

Thank you for the detailed response and the additional experiments. The clarifications regarding the Figure 3 protocol, the reframing of the anchor–context split as a functional division rather than strict disentanglement, and the additional controls on correction pathways and expert routing have addressed several of my concerns. I have increased my Soundness score accordingly. However, I remain unconvinced that the revisions sufficiently address my main concerns, particularly regarding the necessity of individual-conditioned refinement and the extent to which the motivation, proposed method, and supporting evidence are aligned. Therefore, I do not believe a change in my overall rating is warranted at this stage.

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

Related papers

Discussion (0)

Sign in to comment.