acceptodds
Under review as a conference paper at ICLR 2027

Hierarchical Topology-Aware Sparse Attention for Medical Time-Series Classification

Abstract

Medical time series (MedTS) exhibit complementary temporal and inter-channel dependencies that are structured and non-uniform. Existing Transformer-based methods increasingly introduce structured representations and interactions, but generally do not adapt their interaction patterns to the distinct characteristics of temporal and inter-channel dependencies. We propose , a sparse attention framework that combines branch-specific interaction structures with hierarchical local-to-global communication. In the channel branch, direct interactions are determined by groups derived from training-set Fast Fourier Transform (FFT)-based spectral similarities, whereas the temporal branch restricts interactions to local temporal neighborhoods; learnable hubs coordinate information across groups. Mixture-of-Experts (MoE)-based temporal embedding and subject-guided enhancement further enrich the complementary branch representations. Under subject-independent five-fold evaluation on five EEG and ECG datasets, TopoAttn achieves the best result in 27 of 30 dataset-metric comparisons among the evaluated methods. Ablation and sensitivity analyses support the benefit of structured sparse interaction and branch-specific grouping beyond sparsity alone. TopoAttn also achieves the best result across all metrics on two human activity recognition datasets, supporting transfer beyond MedTS. The code will be released upon acceptance.

Primary area:
learning on time series and dynamical systems
Keywords:
medical time-series, time-series classification, topology-aware attention, sparse attention, representation learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_nfwvOpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TopoAttn addresses medical time-series classification using separate channel and temporal representations, fixed training-derived spectral channel groups, local temporal neighborhoods, and hub-mediated communication. Temporal MoE embeddings and subject-context pooling complete the system. Five medical and two HAR datasets, component/topology comparisons, subject-level results and inference-access sensitivity provide evidence for the integrated configuration.

Critical strengths

S1. A coherent design incorporating distinct channel and temporal priors. TopoAttn combines training-derived spectral channel groups with local temporal neighborhoods in separate branches, using a shared hierarchical communication framework. The architecture provides an explicit, technically coherent way to encode these different interaction structures for medical time-series classification.

S2. Good reported gains over a close predecessor. Table 2 reports the highest mean performance in 27 of 30 medical dataset–metric comparisons, including comparisons with TeCh. These results support the empirical value of the integrated architecture.

Critical weaknesses

W1. The central TopoAttn contribution is insufficiently differentiated from close prior work. The first contribution emphasizes topology-aware local interactions and group-mediated global coordination (p.2, L69–87). RECTOR [1] already combines within-region local channel interactions, bidirectional channel–region exchange and global region attention for neural signals, but is not discussed. This overlap narrows the novelty of the general communication principle underlying TopoAttn, which is presented as the paper’s central methodological contribution.

W2. The topology comparisons do not fully isolate the benefits of TopoAttn’s specific design choices. The paper includes topology comparisons and group-count sensitivity analyses, but the FFT/prior comparisons change channel-group and therefore hub counts on PTB (8/3) and PTB-XL (4/2) (Table 6; A.5). The APAVA group-count sensitivity analysis (Table 11) does not resolve this confound. The random-control descriptions do not explicitly establish whether group sizes and permitted attention-connection counts are preserved (A.9.1–A.9.2). Moreover, the reported ablations do not isolate the additional benefit of hub–hub communication while retaining the global aggregation already provided by the sample core (§3.4; A.4.2, Eq.5). These comparisons support useful configurations more clearly than they establish the benefits attributable to particular grouping and communication choices.

References

[1] Jinhan Liu et al. RECTOR: Masked Region-Channel-Temporal Modeling for Affective and Cognitive Representation Learning. Proceedings of the 43rd International Conference on Machine Learning (ICML), PMLR 306:78389–78415, 2026.

Questions for authors

Q1. Given that TopoAttn is presented as the central methodological contribution, what advance does it provide relative to RECTOR? Please clarify how TopoAttn’s specific mechanisms differ from the shared principle of local interactions and group-mediated global communication, and identify what is new or informative in its design or empirical findings.

Q2. Does random channel grouping preserve the FFT partition’s group sizes and hub count, and does random temporal topology preserve the relevant connection counts or degrees? What evidence separates group membership from changes in the interaction budget in the FFT/prior comparisons? Please also clarify which hub operations remain in the dense variants and whether any comparison isolates hub–hub communication while retaining the sample core and other components. If existing controls do not isolate these effects, a targeted comparison would help.

Q3. Table 13 specifies two-sided t-tests at (p<0.05p<0.05). What observations enter each test, and are the tests paired or unpaired? How is dependence from repeated evaluation of the same subjects across seeds and overlapping training sets across folds addressed? Please clarify which sources of uncertainty these tests quantify.

Q4. Given the same-subject context and the independent-inference analysis in Table 14, which comparative conclusions extend to independent-sample inference? Please confirm that the retrained w/o Sub variant removes cross-recording enhancement during both training and inference while retaining within-sample cores. Please specify the input context available to TeCh and Medformer and their per-dataset numerical configurations and hyperparameter-selection procedures.

Q5. A.1 describes channel shuffling during training. How are these permutations handled in the channel-topology masks and positional encodings? When constructing and applying cached subject cores, how is correspondence between channel tokens across samples maintained?

6 comments
AuthorsOpenReview ↗

Response to Reviewer nfwv (1/2)

We thank the reviewer for the constructive comments. Related points are grouped and answered below. Tables A-C are provided in the General Response to Reviewers.

1. Difference from RECTOR [W1, Q1]

We now cite RECTOR in Related Work (Section 2). Both models share the general structure of local interaction with group communication, but differ in two essential aspects:

  • Channel structure. RECTOR groups channels into anatomy-initialized regions, which requires a known electrode layout. TopoAttn derives a spectral topology from training subjects only, without any layout, and outperforms the anatomical prior in all 42 dataset-metric comparisons (Table 10).
  • Temporal structure. RECTOR applies dense attention along time, whose cost grows quadratically with sequence length. TopoAttn uses a locality topology with linear cost (Appendix A.6), which outperforms dense attention on all five MedTS datasets (Table 4).

RECTOR has no public implementation, so an empirical comparison is not feasible in our rebuttal.

2. Isolating the benefits of Topology [W2, Q2]

  • Random and domain-prior controls. The random topologies differ from TopoAttn only in group membership, with identical group number, group sizes, hub count, and permitted connections, and thus isolate the effect of the topology from the interaction budget. The domain-prior topology follows fixed anatomical groupings, so its group count can differ from that of TopoAttn. TopoAttn outperforms both controls in all dataset-metric comparisons (Tables 10 and 11).

  • Hub operations. In w/o Topology ablation study, we directly replace the topology-aware sparse attention with dense attention in both channel and temporal branches, so no hub structure remains. All other components, including subject-guided aggregation module (subject and sample core) are kept unchanged. TopoAttn outperforms this w/o Topology variant on all MedTS datasets (Table 4). We also add an experiment that equips TeCh with our topology-aware sparse attention structure. The results also show consistent improvements on all MedTS datasets (See the Table below).

Table. F1 of TeCh equipped with the channel and temporal topologies of TopoAttn (TeCh+Topology) under subject-enhanced inference. Best in bold.

ModelADFTDAPAVATDBrainPTBPTB-XL
TeCh46.98±0.6274.81±1.1085.73±2.0183.47±1.4962.87±0.09
TeCh+Topology48.61±0.4777.64±0.6187.19±1.8784.43±1.1263.28±0.08
TopoAttn50.42±1.0380.13±0.2784.04±1.0586.45±1.0463.61±0.11

3. Significance tests [Q3]

[Corrected on 18 Nov 2026; see our follow-up response.] We have updated the description of the significance tests under two inference protocols (Tables 14 and 15). Each test compares the five seed-level averages (each over the five folds) of TopoAttn and TeCh for one dataset and metric using an unpaired two-sided Welch's t-test, and therefore reflects variability across training seeds. Variability across held-out subjects is analyzed with a paired fold-level test in our follow-up response and in Appendix A.11.

AuthorsOpenReview ↗

Response to Reviewer nfwv (2/2)

4. Per-sample inference and baseline settings [Q4]

  • We have updated per-sample inference results for TopoAttn (Table A) and completed the comparison of TopoAttn under two inference protocols for all MedTS datasets (Table C). The results support that the main conclusion extends to per-sample inference: TopoAttn achieves the best result in 22/30 evaluation metrics, and its performance changes only marginally from subject-enhanced inference on datasets with sufficient samples per subject. We confirm that w/o Sub ablation only removes the cross-recording enhancement in both training and inference.

  • We now ensure that all methods receive the same input under each protocol. Under per-sample inference, every method receives only the single input sample (Table A); under subject-enhanced inference, all methods receive the same subject core as TopoAttn in both training and inference (Table B). All baselines use the optimal hyperparameters of their official implementations, and the checkpoint with the best validation F1 is used for testing. The per-dataset configurations of TeCh and TopoAttn are listed below.

Table. Training configurations and model sizes of TeCh and TopoAttn. B: batch size; D: feature dimension; Nt/Nc: numbers of temporal/channel layers (0: branch disabled); Emax: maximum epochs; PT: early-stopping patience.

DatasetModelBDNtNcEmaxPT
ADFTDTeCh1281286666
TopoAttn1281286666
APAVATeCh128256664020
TopoAttn128256664020
TDBrainTeCh128128606060
TopoAttn128128316060
PTBTeCh128256038040
TopoAttn128256338040
PTB-XLTeCh128128502020
TopoAttn256128512020

5. Channel shuffling and token correspondence [Q5]

  • Topology masks and positional encodings. Channel shuffling is a training-only augmentation that permutes entire channels before tokenization, following Medformer and TeCh (Appendix A.1); positional encodings are added afterwards. Since channels are grouped by the FFT profile of the entire channel signal (Section 3.3), which shuffling leaves unchanged, each channel keeps its group after permutation, and the topology masks are unaffected.

  • Correspondence in subject cores. The subject core is built right after token embedding. For each anchor sample, we aggregate all samples with the same subject ID only once using Eq. 5 (Section 3.4). The aggregation is performed at each token position, so each core token corresponds to the same channel in all samples. The subject core is then concatenated to the anchor sample to form the enhanced sample. After that, each encoder layer performs only sample-level aggregation within the enhanced sample.

I thank the authors for the detailed response and revision. The clarifications concerning inference protocols (Q4), together with the additional controls and portability experiments (Q2), address several of my earlier concerns.

My principal concern remains the distinct contribution, alongside several questions about interpreting the results.

1. Novelty and scope (W1/Q1). The response acknowledges that TopoAttn shares the general principle of local interaction with group-mediated global communication. Related principles appear in local-plus-global token attention (Longformer, BigBird), core aggregation–redistribution (SOFTS, TeCh), and region-level channel interaction with global region attention (RECTOR). The dual tokenization and core module also build on TeCh and SOFTS. The differences from RECTOR (data-driven rather than anatomical groups, and local rather than dense temporal attention) clarify the design, but do not by themselves establish the significance of the methodological advance.

The distinctive contribution therefore appears to lie primarily in the specific topology construction: FFT-similarity channel clustering and sliding-window masks with segment hubs. Contribution (1), however, also emphasizes the hierarchical communication mechanism, whose incremental benefit remains insufficiently isolated:

  • As the response confirms, the w/o Topology ablation removes the masks and hubs together while retaining the sample core, which already performs global aggregation. It therefore does not separately establish the benefit of the hubs or hub–hub communication. A focused control removing the hubs, or only the hub–hub attention, would clarify this.
  • The interaction counts in A.6 support theoretical savings, but their practical benefit remains unclear without a matched runtime and memory comparison against w/o Topology.

Please distinguish the proposed topology construction from the communication principles inherited from prior work, and align contribution (1) with the evidence. I acknowledge the empirical improvements, but remain uncertain whether they establish a sufficiently distinct methodological advance.

2. Variability and statistical tests (Q3). Please clarify what the reported standard deviations measure (e.g., APAVA accuracy ±0.24, Table 3), and provide:

  • the 25 individual run scores;
  • the fold/seed averaging order;
  • the observations entering each Welch test.

In particular, please distinguish variability across seeds from variability across held-out subjects or folds. Since all seeds reuse the same fold assignment, please account for pairing and dependence among runs, for example with a paired fold-level or corrected resampled t-test.

3. Channel shuffling (Q5). Permuting channel order preserves each channel's FFT profile, but changes its positional index. Please explain how correspondence among the shuffled inputs, topology masks, and subject-core tokens is maintained, including:

  • whether the masks are permuted or recomputed;
  • whether cached cores use augmented or unaugmented samples;
  • how each shuffled input is aligned with its core.

The description in A.1 does not yet make this clear. Any misalignment could affect the interpretation of both the topology and the subject-enhancement results.

4. TeCh+Topology configuration. The supplied TeCh configurations disable the channel branch on TDBrain and PTB-XL and the temporal branch on PTB. Please report the branch and layer settings of TeCh+Topology, state whether they match TeCh's, and include the experiment and its configuration in the manuscript. This is needed to attribute the reported gains to the topology addition.

The revision strengthens the experimental evidence, and I credit the matched protocols, random controls, TeCh+Topology results and PTB-XL per-sample evidence. My main reservation remains the distinct contribution, with the questions above limiting the interpretation of some results. I therefore maintain my current score of 2. Focused clarification and evidence addressing these points would strengthen the case and prompt me to reassess the rating.

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

TopoAttn is a Transformer for EEG and ECG classification with two branches: one token per channel, and cross-channel temporal patches embedded by a mixture-of-experts layer. Attention is restricted in both branches. Channels are clustered by FFT band-energy similarity on the training subjects, temporal patches attend within a local window and fixed contiguous groups, and learnable hub tokens connect the groups. A subject-level core token, built from all samples of the same subject, is added to each sample, and a gate fuses the two branches. Under subject-independent 5-fold evaluation on five datasets against 11 baselines, it is reported best on 27 of 30 metrics, with a 4.5% relative F1 gain over TeCh. Two HAR datasets are used to claim transfer beyond medical data.

Critical strengths
  1. Subject-disjoint folds, five seeds and a shared baseline framework, with both sample- and subject-level results.
  2. The appendix reports the same-subject aggregation used at inference and a per-sample check (App. A.12), which is what exposes weakness 2.
Critical weaknesses
  1. Little is new. Every component is borrowed: dual channel/temporal tokens from TeCh, the core token from SOFTS, channel tokens from iTransformer, patches from Medformer, and a standard noisy-gated MoE. Local windows plus global hub tokens is the local-plus-global sparse attention of Longformer and BigBird, which the paper neither cites nor compares with. The only remaining idea is to cluster channels by FFT profile and cut the time axis into fixed segments. That is a design choice, not a contribution. Sparsity is also pointless here: the channel branch has 12 to 33 tokens (Table 5), and TeCh is lighter anyway (App. A.10). I learned nothing new.

  2. Table 2 is not a fair comparison. By default, TopoAttn builds its subject core from the subject ID and the other unlabeled test recordings of the same subject (Sec. 3.4, 3.6, App. A.12). No baseline gets this. The main text mentions it in one sentence (Sec. 3.6), and the pointer to the per-sample result is a broken reference ("App. ??", line 471). Per-sample results are given for only two of five datasets. On PTB-XL, F1 falls from 63.61 to 61.85, below TeCh (62.85), Medformer (62.69) and PatchTST (62.38). The claimed gain over TeCh there (0.76 F1) is smaller than the same-subject effect (1.76). On APAVA, per-sample inference changes nothing (80.13 vs 80.11), yet "w/o Sub" costs 3 F1 points: the module adds training capacity, not subject context.

  3. The gains are within noise. "27 of 30" counts six correlated metrics, and Table 13 shows TopoAttn significantly worse than TeCh on TDBrain recall and F1. The FFT channel topology is within about one standard deviation of random or domain-prior grouping on ADFTD (50.42±1.03 vs 49.48±1.36 and 49.91±0.79) and TDBrain (84.04±1.05 vs 82.69±1.26 and 83.40±0.80). The standard deviations and t-tests pool 25 runs over the same subjects, so they measure seed noise, not subject variation; APAVA has only 23 subjects. On APAVA, the accuracy std of TopoAttn (±0.24) is 2 to 10 times smaller than that of every baseline and every other setting in Table 11, which needs explanation. Gch and Gte are tuned per dataset on validation data (App. A.9.3); the paper does not say whether the baselines were tuned. Changing Gch, Gte or the window radius r moves F1 on APAVA by up to 3.5 points, more than the topology gains being claimed.

  4. The transfer claim is unsupported. It rests on a 0.72% average gain over TeCh on two HAR datasets (8 and 30 subjects), with no significance test.

Questions for authors

See weaknesses.

4 comments
AuthorsOpenReview ↗

Response to Reviewer kRdW (1/2)

We thank the reviewer for the constructive comments. Related points are grouped and answered below. Tables A-C are provided in the General Response to Reviewers.

1. Novelty [W1]

  • Components. We agree that the building blocks and the local-plus-global pattern are not entirely new, and we do not claim them as our contribution. Our contribution is a topology-driven interaction structure for modeling MedTS dependencies. This structure distinguishes TopoAttn from generic sparse attention and dual-tokenization methods.

  • Comparison with Longformer and BigBird. We now cite both models and add them as baselines. TopoAttn outperforms them on 26/30 and 28/30 dataset-metric pairs, respectively (Table A). This shows that topology-driven sparsity is more effective than generic sparsity for MedTS modeling.

  • Topology as a contribution. To verify that the topologies are a contribution rather than a design choice, we add an experiment that equips TeCh with our topologies. They improve TeCh on all five datasets (See the Table below). This shows that our topologies capture transferable MedTS structure. In addition, both our channel and temporal topologies outperform the alternative topologies in all 42 dataset-metric comparisons (Tables 10 and 11). These results show that our topologies capture useful and transferable MedTS structure.

  • Sparsity and efficiency. Sparsity serves different purposes in the two branches. In the temporal branch, tokens are numerous, and the local topology keeps the attention cost linear in sequence length (Appendix A.6). In the channel branch, tokens are few (12-33), and sparsity acts as a structural prior: it groups channels with similar spectral profiles, which improves dependency modeling (Table 10). Although TeCh is lighter, TopoAttn remains within the efficiency range of common Transformer baselines (comparable to MTST and faster than Medformer and Reformer) (Appendix A.10). Combined with its consistent accuracy gains, TopoAttn achieves the best accuracy-efficiency trade-off.

Table. F1 of TeCh equipped with the channel and temporal topologies of TopoAttn (TeCh+Topology) under subject-enhanced inference. Best in bold.

ModelADFTDAPAVATDBrainPTBPTB-XL
TeCh46.98±0.6274.81±1.1085.73±2.0183.47±1.4962.87±0.09
TeCh+Topology48.61±0.4777.64±0.6187.19±1.8784.43±1.1263.28±0.08
TopoAttn50.42±1.0380.13±0.2784.04±1.0586.45±1.0463.61±0.11

2. Fairness of the comparison [W2]

  • Matched protocols. We now evaluate the methods under two matched protocols (Section 3.6). (1) Per-sample inference: We remove TopoAttn's subject-guided enhancement at inference to ensure a direct comparison with all 13 SOTAs, which operate without subject-level information (Table 2). (2) Subject-enhanced inference: We equip Medformer and TeCh, the best MedTS-specific SOTAs, with the same subject-guided enhancement module as TopoAttn during both training and inference, ensuring identical access to subject information. TopoAttn achieves the best result in 22/30 (Table A, full results in Table 2) and 25/30 (Table B) dataset-metric pairs under the two inference protocols, respectively. Updated significance tests are reported in Appendix A.11.

  • PTB-XL. We update the comparison results of TopoAttn under two inference protocols on all five MedTS datasets (Table C). The effect of inference-time subject context depends on the number of samples per subject. PTB-XL has only 11 samples per subject on average, so the subject core provides limited context during training and the model still relies on test-time aggregation: per-sample inference lowers F1 by 1.76%, whereas datasets with 259-793 samples per subject change by only 0.04% (Table C). Moreover, when TeCh, the best SOTA, uses the same subject context, TopoAttn still outperforms it on all six metrics on PTB-XL (Table B), so its gain is not explained by subject context alone.

  • APAVA: subject context vs. training capacity. The gain of subject-guided enhancement mainly comes from subject information learned during training, not from added capacity. APAVA has sufficient samples per subject, so the model learns subject-level patterns during training and infers them from a single sample. This explains why per-sample inference barely changes F1 (80.13 vs. 80.11). Three evidences show that this gain is not added capacity: (1) the subject core uses only a lightweight projection with dimension D/4D/4 (Eq. 5); (2) our added experiments show that the same module also improves Medformer and TeCh (Table A vs. Table B); and (3) the benefit of inference-time aggregation generally grows as samples per subject decrease (Table C), which added capacity would not explain.

AuthorsOpenReview ↗

Response to Reviewer kRdW (2/2)

3. Gains beyond noise [W3]

  • Significance. We update the significance tests under both inference protocols (two-sided Welch's t-test, p < 0.05; Appendix A.11). TopoAttn significantly outperforms TeCh on 20/30 and 24/30 dataset-metric pairs, respectively. This shows that its consistent gains are beyond noise.

  • Seed noise vs. subject variation. Our variance reflects subject variation, not only seed noise. Each fold tests a disjoint set of unseen subjects, so the 25 runs (5 folds × 5 seeds) cover between-subject variation. Our subject-level evaluation (Table 13), which makes one prediction per subject, also supports that the gains hold across unseen subjects.

  • Low variance on APAVA. The low variance is reproducible under both inference protocols (Tables A and B). The std varies across hyperparameter settings, and the selected setting yields the lowest std (Table 12). This shows that the variance depends on the topology configuration rather than on an evaluation artifact. We attribute the low variance to two factors: the selected topology matches the dependency structure of APAVA, and subject-guided training helps the model learn stable subject-level patterns.

  • Topology gains. The FFT-derived topology outperforms the domain-prior and random topologies in all 42 dataset-metric comparisons across all MedTS and HAR datasets (Table 10). This consistency indicates that the topology gains are reliable rather than within noise.

  • Tuning. We follow the evaluation practice of TeCh. TopoAttn uses the same batch size, feature dimension, maximum epochs, and early-stopping patience as TeCh (See the table below). Only the layer numbers and topology parameters are selected on the validation set. Varying GchG_{\mathrm{ch}}, GteG_{\mathrm{te}}, and rr only changes the granularity of the topology. TopoAttn still outperforms TeCh in 69/72 comparisons across all 12 non-default configurations, and even its lowest F1 on APAVA remains above that of TeCh (76.67 vs. 74.81) (Appendix A.9.3). This shows that its advantage does not rely on narrow tuning.

Table. Training configurations and model sizes of TeCh and TopoAttn. B: batch size; D: feature dimension; Nt/Nc: numbers of temporal/channel layers (0: branch disabled); Emax: maximum epochs; PT: early-stopping patience.

DatasetModelBDNtNcEmaxPT
ADFTDTeCh1281286666
TopoAttn1281286666
APAVATeCh128256664020
TopoAttn128256664020
TDBrainTeCh128128606060
TopoAttn128128316060
PTBTeCh128256038040
TopoAttn128256338040
PTB-XLTeCh128128502020
TopoAttn256128512020

4. Transfer beyond MedTS [W4]

We re-evaluate the HAR datasets under per-sample inference (Table 8). TopoAttn achieves the best result in 10/12 dataset-metric pairs and outperforms the best SOTA, TeCh, on all 12 pairs with smaller standard deviations. This shows that its gains on HAR are consistent. The absolute margins are small, especially on UCI-HAR, where most SOTAs are already close to saturation. We therefore present HAR as supplementary evidence of transfer beyond MedTS.

Dear authors,

Thank you for the rebuttal and the new experiments. They confirm my main concerns rather than resolve them, and thus I keep my score.

i. Under the fair protocol, the advantage is gone on 2/5 datasets. With per-sample inference, TopoAttn is below TeCh on PTB-XL (F1 61.85 vs 62.85) and TDBrain (83.58 vs 84.73), and on both it is also below Longformer, which has no topology at all. The original headline results relied on pooling the unlabeled test recordings of the same subject.

ii. The significance claims have been withdrawn. Your follow-up to Reviewer nfwv concedes that the ± values and Welch tests reflect only seed variability. With the paired fold-level test, TopoAttn is not significantly better than TeCh on any metric (p = 0.24 to 0.62); on APAVA, fold-level F1 ranges from 63.04 to 96.00, against a reported ±0.27.

iii. Little of the method remains. The building blocks and the local-plus-global pattern are not new; the topology is a mask on dense attention with no substantial speed gain, and removing the hubs changes APAVA accuracy by 0.02 and PTB-XL F1 by 0.09. The contribution reduces to FFT-based channel clustering and a sliding temporal window.

AuthorsOpenReview ↗

Follow-up Response to Reviewer kRdW

Dear Reviewer kRdW,

We appreciate your continued engagement. However, several conclusions in your latest comments do not reflect the full experimental evidence or the clarifications provided in our rebuttal. We would like to correct these points to ensure that our contributions and results are assessed accurately.

1. Interpretation of the experimental results

(1) Your assessment focuses on a small number of metrics where TopoAttn performs slightly below individual baselines, without accounting for its overall performance. Under the two evaluation protocols, TopoAttn achieves the best results on 22/30 and 25/30 dataset–metric pairs (Tables 2 and 3). The highlighted F1 differences relative to TeCh are only 1.00 on PTB-XL and 1.15 on TDBrain. Similarly, TopoAttn outperforms Longformer on 26/30 dataset–metric pairs overall, including 4/6 metrics on each of PTB-XL and TDBrain (Rebuttal Table A). Focusing exclusively on the remaining metrics does not provide a representative assessment of the results.

(2) In particular, the statement that our advantage "is gone" on TDBrain and PTB-XL is not supported by Tables 2 and 3. Under per-sample inference, TopoAttn achieves the best results on 3/6 metrics on TDBrain and ranks among the top two on 3/6 metrics on PTB-XL. Under subject-enhanced inference, it ranks among the top two on 4/6 metrics on TDBrain and achieves the best results on all six metrics on PTB-XL. We therefore maintain that the overall experimental evidence supports the effectiveness of TopoAttn, even though it does not outperform every baseline on every metric.

2. Architectural novelty and source of performance gains

(1) Our contribution is not the invention of each individual component, as we have already clarified. Rather, it is their integration into a hierarchical, topology-driven interaction structure that captures token dependencies in medical time series without relying on domain-prior layouts. This distinguishes TopoAttn from generic sparse attention mechanisms such as Longformer and BigBird, as well as dual-tokenization approaches such as TeCh. These distinctions are explained in Sections 3.3-3.4 and our rebuttal, while Tables 4, 5, 12, 16, and 17 provide experimental evidence for the architectural design.

(2) We also disagree with the attribution of our performance gains to pooling unlabeled test recordings from the same subject. TopoAttn already achieves the best results on 22/30 dataset-metric pairs under per-sample inference, where every test sample is predicted independently (Table 2). Furthermore, removing topology groups or hub structures decreases performance even when subject-level aggregation is retained (Table 4). These controlled comparisons directly demonstrate that only subject-level aggregation cannot account for the observed gains and that the proposed architectural components contribute to performance. Both points were explicitly addressed in our rebuttal (W1, W2).

3. Statistical interpretation

The comparison between the fold-level F1 range on APAVA (63.04-96.00) and the reported ±0.27 conflates two different sources of variation. As explained in Appendix A.11, ±0.27 represents the standard deviation across five seed-level averages, each computed over five folds. These averages are also used in our Welch's t-tests. We additionally reported paired fold-level tests, which show that the difference from TeCh is not statistically significant. We have not claimed otherwise. Nevertheless, TopoAttn exhibits lower fold-level variability and a higher worst-fold F1 than TeCh.

Thanks again,

The Authors

Reviewer_1BBrOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TopoAttn for medical time-series classification. Its two branches model channel dependencies and temporal dependencies separately. The channel branch clusters channels according to the similarity of their FFT frequency-band energies in the training set, while the temporal branch restricts patch interactions to local neighborhoods. Both branches use learnable hub tokens for within-group aggregation, communication between groups, and redistribution. The model also includes MoE patch embedding, subject-guided enhancement that aggregates samples from the same subject, and gated fusion.

Critical strengths
  1. The experimental protocol is fairly rigorous: subject-independent five-fold evaluation across five seeds, 11 baselines rerun under a unified framework, six reported metrics, and additional subject-level evaluation and significance tests.

  2. The topology sensitivity analysis is well targeted. FFT-based grouping is compared with anatomical-prior and random grouping, while local temporal topology is compared with random topology. These comparisons suggest that the gains do not come solely from sparsity. This paper also discusses the effect of aggregating samples from the same subject at inference time.

Critical weaknesses
  1. The methodological novelty is limited. ''Local attention plus global hub tokens'' has been a common sparse-attention pattern since Longformer [1] and BigBird [2], while most of the other components come from TeCh and SOFTS. The benefit of the topology itself is modest: FFT grouping outperforms anatomical-prior grouping by only about one F1 point on average, and TeCh achieves significantly better F1 on TDBrain.
  2. The main-table comparison is not entirely fair: at inference time, subject-guided enhancement aggregates other samples from the same test subject, whereas the baselines make per-sample predictions. Appendix A.12 shows that switching to strictly per-sample inference on PTB-XL lowers F1 from 63.61 to 61.85, below TeCh’s 62.85. However, this setting is reported for only two datasets and is absent from the main text.

[1]: Iz Beltagy, Matthew E. Peters, and Arman Cohan. Longformer: The Long-Document Transformer. arXiv preprint arXiv:2004.05150, 2020.

[2]: Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontañón, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, and Amr Ahmed. Big Bird: Transformers for Longer Sequences. In Advances in Neural Information Processing Systems (NeurIPS), 2020.

Questions for authors
  1. Could the authors report results for strictly per-sample inference on all five datasets in Table 2 of the main text? How is the subject core constructed at test time, and do the results depend on how many samples from the same subject are available at once?

  2. Removing subject-guided enhancement lowers APAVA F1 by about three points (80.13 → 77.16), but switching to per-sample inference barely changes it (80.13 → 80.11). Does this module’s gain come from subject information, or mainly from regularization during training?

  3. TopoAttn’s DD, number of layers, patch length, GchG_{\mathrm{ch}}, and GteG_{\mathrm{te}} are selected separately for each dataset using validation results. Were the baselines given the same tuning budget? What are the parameter counts of the respective models?

  4. Could the authors compare TopoAttn with general “local plus global token” sparse attention, such as Longformer [1] and BigBird [2], or add the proposed topology directly to TeCh? This would help separate the contributions of the hub mechanism and the branch-specific topologies.

  5. Are the channel groups produced by FFT clustering stable across folds? Do the EEG groups correspond to brain regions?

[1]: Iz Beltagy, Matthew E. Peters, and Arman Cohan. Longformer: The Long-Document Transformer. arXiv preprint arXiv:2004.05150, 2020.

[2]: Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontañón, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, and Amr Ahmed. Big Bird: Transformers for Longer Sequences. In Advances in Neural Information Processing Systems (NeurIPS), 2020.

4 comments
AuthorsOpenReview ↗

Response to Reviewer 1BBr (1/2)

We thank the reviewer for the constructive comments. Related points are grouped and answered below. Tables A-C are provided in the General Response to Reviewers.

1. Fair comparison and subject context [W2, Q1, Q2]

  • W2 (Fairness). We updated the experiments under two matched protocols (Section 3.6), and TopoAttn remains the best under both. (1) Per-sample inference: We remove TopoAttn's subject-guided enhancement at inference to ensure a direct comparison with all 13 SOTAs, which operate without subject-level information. TopoAttn is best in 22/30 dataset-metric pairs among all 14 methods (Table A, full results in Table 2). (2) Subject-enhanced inference: we equip Medformer and TeCh, the two best MedTS-specific SOTAs, in our comparison with the same subject-guided enhancement module in both training and inference, and TopoAttn is best in 25/30 pairs (Table B). Significance tests are reported in Appendix A.11.

  • Q1 (Subject core at test time). Under subject-enhanced inference, the subject core is aggregated from all unlabeled test samples of the same subject using the same subject-guided aggregation module as in training (Eq. 5). The effect of this inference-time subject context depends on the number of samples per subject (Table C): performance is nearly unchanged on datasets with sufficient samples per subject (ADFTD, APAVA, and PTB; 259-793), and the effect becomes larger as samples become fewer, i.e., on TDBrain (87) and especially PTB-XL (11). The reason is that sufficient samples per subject allow the model to learn subject-level patterns during training and infer them from a single sample, whereas with fewer samples, inference-time aggregation supplies more of the subject context.

  • Q2 (Subject information vs. regularization). The gain mainly comes from subject information learned during training. As already shown in our ablation study (w/o Sub, Table 4), subject-guided enhancement significantly improves APAVA F1 from 77.16% to 80.13%. Removing it only at inference keeps F1 almost unchanged (80.13% to 80.11%). Thus, the gain is obtained during training rather than at inference. This gain reflects subject information rather than generic regularization: (1) our added experiments show that the same subject context also improves Medformer and TeCh (Table A vs. Table B), and (2) the benefit of test-time aggregation generally grows as samples per subject decrease (Table C), which generic regularization would not explain.

2. Novelty and the contribution of the topology [W1, Q4]

  • W1 (Novelty). We agree that local-plus-global sparsity comes from Longformer and BigBird, and that dual tokenization and aggregation-redistribution follow TeCh and SOFTS. However, we do not claim these as our contribution. Our contribution is topology-driven modeling of MedTS dependencies. Longformer and BigBird use general position-based sparse patterns, while TopoAttn is topology-driven: its sparse patterns follow the channel and temporal dependencies of MedTS, so each token interacts only with topology-related tokens. We add Longformer and BigBird as baselines, and TopoAttn outperforms them on 26/30 and 28/30 pairs, respectively (Table A).

  • W1 (Topology gain). The topology gain is consistent. The FFT-derived channel topology outperforms both the domain-prior and random topologies in all 42 dataset-metric comparisons (Table 10). The temporal topology also outperforms a random one in all 42 comparisons (Table 11).

  • W1 (Comparison with TeCh). TopoAttn outperforms TeCh, the best SOTA, on 22/30 and 25/30 pairs under per-sample and subject-enhanced inference, respectively (Tables A and B).

  • Q4 (Hub vs. topology). We add two comparisons for explanation. (1) To test whether the gains come from the hub mechanism alone, we add Longformer and BigBird as baselines. They use similar hub-coordinated sparse structure but no topology. TopoAttn outperforms them, respectively (Table A). This shows that our topologies contribute beyond the hub mechanism. (2) To test whether our topologies work on other models, we add them to TeCh. They improve TeCh on all MedTS datasets (See the Table below). This also shows that our topology-aware sparse modeling is effective.

Table. F1 of TeCh equipped with the channel and temporal topologies of TopoAttn (TeCh+Topology) under subject-enhanced inference. Best in bold.

ModelADFTDAPAVATDBrainPTBPTB-XL
TeCh46.98±0.6274.81±1.1085.73±2.0183.47±1.4962.87±0.09
TeCh+Topology48.61±0.4777.64±0.6187.19±1.8784.43±1.1263.28±0.08
TopoAttn50.42±1.0380.13±0.2784.04±1.0586.45±1.0463.61±0.11
AuthorsOpenReview ↗

Response to Reviewer 1BBr (2/2)

3. Tuning budget and model size [Q3]

TopoAttn has a tuning budget and model size comparable to TeCh. We follow the evaluation practice of TeCh, the strongest SOTA, so we compare with their training configurations (See the Table below). TopoAttn uses the same batch size, feature dimension, maximum epochs, and early-stopping patience as TeCh. The only exception is a larger batch size on PTB-XL for training efficiency. Only the numbers of temporal and channel layers and the topology parameters are selected on the validation set. Both models share the same feature dimension and use similar numbers of layers, so their sizes are comparable. Moreover, TopoAttn outperforms TeCh in 69/72 comparisons across 12 non-default topology configurations (Appendix A.9.3). These results show that its gains rely neither on narrow tuning nor on extra parameters.

Table. Training configurations and model sizes of TeCh and TopoAttn. B: batch size; D: feature dimension; Nt/Nc: numbers of temporal/channel layers (0: branch disabled); Emax: maximum epochs; PT: early-stopping patience.

DatasetModelBDNtNcEmaxPT
ADFTDTeCh1281286666
TopoAttn1281286666
APAVATeCh128256664020
TopoAttn128256664020
TDBrainTeCh128128606060
TopoAttn128128316060
PTBTeCh128256038040
TopoAttn128256338040
PTB-XLTeCh128128502020
TopoAttn256128512020

4. Stability and interpretability of the channel groups [Q5]

  • Stability across folds. The channel groups are stable across folds. On ADFTD (EEG), all five folds yield the identical grouping {Fp1, F8, O1, T6, T3, T5, F3}, {Fz, C3, C4}, {F4, T4, P3, O2}, {Cz, Fp2, Pz}, {P4, F7}. On PTB-XL (ECG), the five folds yield only two groupings, which share the precordial groups {V2, V3} and {V4, V5} and differ only in how the limb leads, V1, and V6 are partitioned. Thus, although the exact grouping can be slightly affected by the training subjects, the topology is not subject-specific and reflects stable dependency structures in MedTS.

  • Correspondence to brain regions. The EEG groups do not strictly follow scalp regions. Instead, they cluster channels with similar spectral profiles, which can span different regions (e.g., {Fp1, F8, O1, T6, T3, T5, F3} covers frontal, temporal, and occipital electrodes), since distant electrodes can exhibit similar oscillatory activity. This data-driven grouping captures functional relationships that anatomical partitions miss: it outperforms the anatomical-prior topology in all 42 dataset-metric comparisons (Table 10), and it requires no electrode layout, making it directly applicable to ECG and HAR.

Thanks for the detailed response. The matched inference protocols and TeCh+Topology experiments substantially address my concerns about comparison fairness and the contribution of the topology design. While there are still some concerns remain: please report actual parameter counts and soften the claim that subject-guided enhancement benefits specifically from subject information, since the current ablations do not rule out regularization or architectural effects. Despite these remaining reservations, the additional experiments address my main concerns and strengthen the paper. I have therefore raised my score.

Additionally, I suggest discussing [1–3] in the related work on modeling temporal and cross-channel dependencies in biosignals. These studies would help clarify the relationship to existing graph-based and spatiotemporal modeling approaches.

[1] MBrain: A Multi-channel Self-Supervised Learning Framework for Brain Signals. KDD 2023.
[2] Modeling Multivariate Biosignals With Graph Neural Networks and Structured State Space Models. PMLR 2023.
[3] CBraMod: A Criss-Cross Brain Foundation Model for EEG Decoding. ICLR 2025.

AuthorsOpenReview ↗

Follow-up Response to Reviewer 1BBr

We sincerely thank the reviewer for the positive feedback and for raising the score. In the final version, we will: (1) report the actual parameter counts of TopoAttn and all baselines; (2) soften the claim on subject-guided enhancement as suggested by the reviewer; and (3) include MBrain, GraphS4mer, and CBraMod in the related work, as our examination of these studies confirmed their relevance.

Thank you again for the constructive comments, which have helped improve our paper.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Response to Reviewers (1/3)

We thank all reviewers for their constructive comments and their recognition of TopoAttn's coherent design, strong empirical performance, and comprehensive evaluation. Here, we summarize our main responses and new experiments. Key results are reported in Rebuttal Tables A-C below, and the Revised Manuscript file has been uploaded.

1. Fair comparison with SOTAs

We now evaluate the methods under two matched protocols (Section 3.6).

  • Per-sample inference. We remove TopoAttn's subject-guided enhancement at inference to ensure a direct comparison with all 13 SOTAs, which operate without subject-level information (Table 2).
  • Subject-enhanced inference. We equip Medformer and TeCh, the best MedTS-specific SOTAs, with the same subject-guided enhancement module as TopoAttn during both training and inference, ensuring identical access to subject information.

While SOTA performance is not a prerequisite for acceptance under ICLR's reviewing guidelines, TopoAttn nevertheless achieves the best results in 22/30 and 25/30 dataset-metric pairs under the respective protocols, providing further empirical support for the proposed method. Updated significance tests are reported in Appendix A.11.

2. Role of subject context

  • The gain of subject context comes mainly from training. Adding subject-guided enhancement in training clearly improves TopoAttn (w/o Sub ablation, Table 4), and the same subject context also improves Medformer and TeCh (Tables A and B). In contrast, removing it at inference leaves most metrics largely unchanged (Table C), indicating that TopoAttn learns subject-level patterns during training that can be inferred from a single sample.
  • Inference-time subject context matters most when samples per subject are few. Its effect is negligible on ADFTD, APAVA, and PTB, and largest on PTB-XL, which has the fewest (only 11) samples per subject. When samples per subject are limited, subject-guided enhancement at inference supplies additional subject context, so it brings a larger gain (Appendix A.12).

3. Novelty

  • Contribution. Our contribution is topology-driven modeling of MedTS dependencies: TopoAttn identifies the channel-wise and temporal-wise topologies inherent in MedTS and captures them through topology-constrained interaction and hub-mediated coordination, while remaining efficient (Appendices A.6 and A.10). It is neither a general-purpose sparse attention mechanism nor a modality-specific model.
  • The gains stem from MedTS topologies rather than sparsity alone. As suggested by reviewers, we compare TopoAttn with generic sparse attention (Longformer and BigBird) and clarify its difference from RECTOR. Longformer and BigBird share the local-plus-global sparse pattern but use no topologies; TopoAttn outperforms them in 26/30 and 28/30 pairs, respectively (Table A). TopoAttn's topologies also consistently outperform domain-prior and random ones (Table 5). RECTOR relies on anatomical partitions specific to EEG/sEEG, whereas our topologies are derived from signal characteristics, so TopoAttn generalizes across EEG, ECG, and HAR (Tables A, B, and 8).
AuthorsOpenReview ↗

General Response (2/3): Rebuttal Table A

Table A. Results under per-sample inference (excerpt of Table 2; full results in the revised manuscript). Best in bold, second best in italics.

DatasetModelAccuracyPrecisionRecallF1AUROCAUPRC
ADFTDLongformer52.53±0.5449.80±0.5949.63±0.2548.82±0.5267.44±0.6751.54±0.83
BigBird49.96±0.7247.61±0.8047.16±0.7746.50±0.8165.27±0.5548.91±0.71
Medformer48.84±0.7745.47±2.8243.40±0.9740.22±1.7264.09±1.0747.68±1.03
TeCh50.71±0.7448.35±0.6147.28±0.6846.58±0.9766.57±0.6350.73±0.58
TopoAttn (ours)53.96±0.9851.72±1.2250.94±1.0350.35±1.0269.30±0.8754.04±1.02
APAVALongformer75.33±0.9571.79±0.6275.45±0.9570.48±0.9982.08±0.5676.32±1.10
BigBird74.05±1.2270.28±0.9174.14±1.0168.72±1.2681.17±0.9274.39±0.85
Medformer76.08±1.0071.97±0.6576.40±0.6670.93±0.9382.81±0.5074.99±0.58
TeCh80.25±1.1878.11±0.7280.22±1.9674.66±1.6888.46±0.5580.84±0.60
TopoAttn (ours)84.06±0.2379.91±0.2385.19±0.4680.11±0.2691.30±0.2289.46±0.69
TDBrainLongformer85.23±0.4684.85±0.6183.92±0.5683.73±0.6093.31±0.4092.36±0.50
BigBird85.08±0.4884.56±0.4683.60±0.5483.48±0.4792.82±0.6491.78±0.80
Medformer83.48±0.4083.79±0.3181.83±0.5881.87±0.4992.04±0.4691.21±0.53
TeCh86.23±0.4885.69±0.5084.87±0.4584.73±0.4693.68±0.3092.75±0.30
TopoAttn (ours)85.87±0.4886.83±0.4383.26±1.4383.58±1.3593.79±0.2993.17±0.42
PTBLongformer90.15±0.3882.20±0.4781.98±1.5081.40±0.9692.51±1.0086.74±0.92
BigBird89.85±0.4781.69±1.0281.78±0.6381.14±0.7591.82±0.7485.92±0.56
Medformer89.30±0.2880.96±0.2881.43±1.2380.54±0.6691.38±0.7386.50±0.74
TeCh89.67±0.5881.14±0.9882.92±0.6981.24±0.7393.17±0.7287.16±1.09
TopoAttn (ours)92.63±0.5286.87±1.3487.31±0.5286.41±0.7496.02±0.5492.78±0.84
PTB-XLLongformer72.95±0.1765.12±0.3461.14±0.4762.51±0.2890.03±0.1067.28±0.23
BigBird72.65±0.1964.88±0.3460.42±0.1461.76±0.1389.94±0.0567.21±0.09
Medformer73.36±0.1064.87±0.1461.19±0.1562.69±0.1389.93±0.0767.29±0.14
TeCh73.53±0.1665.33±0.4161.49±0.3162.85±0.1790.16±0.0867.82±0.17
TopoAttn (ours)73.33±0.3465.21±0.5960.09±0.4861.85±0.4290.23±0.2367.44±0.54
AuthorsOpenReview ↗

General Response (3/3): Rebuttal Tables B and C

Table B. Results under subject-enhanced inference (Table 3 of the revised manuscript). Medformer and TeCh use the same subject core as TopoAttn in both training and inference. Best in bold, second best in italics.

DatasetModelAccuracyPrecisionRecallF1AUROCAUPRC
ADFTDMedformer49.02±0.4845.87±0.3444.18±0.4641.16±0.4264.92±0.3248.27±0.41
TeCh50.82±0.6248.69±0.4847.56±0.5846.98±0.6266.72±0.4650.83±0.36
TopoAttn (ours)54.09±0.9951.78±1.2551.01±1.0550.42±1.0369.35±0.8754.10±1.01
APAVAMedformer77.82±2.6873.78±1.2178.89±2.1172.60±3.0485.47±1.0777.28±1.43
TeCh80.33±0.8478.45±0.3980.88±1.2474.81±1.1088.62±1.1180.93±0.97
TopoAttn (ours)84.08±0.2479.92±0.2385.20±0.4780.13±0.2791.30±0.2389.46±0.70
TDBrainMedformer85.84±0.5886.03±0.5084.06±0.5984.31±0.5993.32±0.2092.38±0.29
TeCh87.13±1.6986.81±1.5885.91±2.0585.73±2.0194.16±1.0993.31±1.22
TopoAttn (ours)86.14±0.3686.99±0.3683.67±1.2284.04±1.0593.90±0.3293.27±0.43
PTBMedformer90.04±0.2082.11±0.3883.40±0.8282.23±0.4591.97±0.4786.79±0.81
TeCh90.97±0.7783.97±1.4084.77±1.6183.47±1.4995.22±0.7991.46±1.26
TopoAttn (ours)92.72±0.7186.89±1.4287.37±0.7586.45±1.0496.05±0.2892.82±1.06
PTB-XLMedformer73.41±0.0764.88±0.2361.27±0.0962.73±0.1189.98±0.0567.34±0.14
TeCh73.58±0.1065.05±0.2461.66±0.2162.87±0.0990.25±0.0767.90±0.16
TopoAttn (ours)74.93±0.0767.12±0.2362.05±0.0963.61±0.1191.01±0.0569.51±0.14

Table C. Effect of the inference protocol on TopoAttn (Table 16 of the revised manuscript). "Sub-enhanced" builds the subject core from all test samples of the same subject, while "Per-sample" uses only the sample itself. Δ denotes per-sample minus subject-enhanced performance; "Samples/Subject" is the average number of samples per subject.

DatasetSamples/SubjectInferenceAccuracyPrecisionRecallF1AUROCAUPRC
ADFTD793Sub-enhanced54.09±0.9951.78±1.2551.01±1.0550.42±1.0369.35±0.8754.10±1.01
Per-sample53.96±0.9851.72±1.2250.94±1.0350.35±1.0269.30±0.8754.04±1.02
Δ−0.13−0.06−0.07−0.07−0.05−0.06
APAVA259Sub-enhanced84.08±0.2479.92±0.2385.20±0.4780.13±0.2791.30±0.2389.46±0.70
Per-sample84.06±0.2379.91±0.2385.19±0.4680.11±0.2691.30±0.2289.46±0.69
Δ−0.02−0.01−0.01−0.020.000.00
TDBrain87Sub-enhanced86.14±0.3686.99±0.3683.67±1.2284.04±1.0593.90±0.3293.27±0.43
Per-sample85.87±0.4886.83±0.4383.26±1.4383.58±1.3593.79±0.2993.17±0.42
Δ−0.27−0.16−0.41−0.46−0.11−0.10
PTB325Sub-enhanced92.72±0.7186.89±1.4287.37±0.7586.45±1.0496.05±0.2892.82±1.06
Per-sample92.63±0.5286.87±1.3487.31±0.5286.41±0.7496.02±0.5492.78±0.84
Δ−0.09−0.02−0.06−0.04−0.03−0.04
PTB-XL11Sub-enhanced74.93±0.0767.12±0.2362.05±0.0963.61±0.1191.01±0.0569.51±0.14
Per-sample73.33±0.3465.21±0.5960.09±0.4861.85±0.4290.23±0.2367.44±0.54
Δ−1.60−1.91−1.96−1.76−0.78−2.07

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.