acceptodds
Under review as a conference paper at ICLR 2027

StreamDC: Condensing Time Series into a Shared Temporal Memory

Abstract

Time-series dataset condensation compresses a training set into a few synthetic sequences. Under a tight signal storage budget, these sequences must preserve diverse temporal patterns and their useful positional variation. We introduce StreamDC, which learns a class-segmented temporal memory and reuses its signal fragments through overlapping training windows. Linear windows slide across the memory with targets given by temporal class occupancy; class-local cyclic windows wrap from each segment's end to its beginning, adding within-class temporal offsets. All windows jointly optimize the same stored values through an existing teacher-guided condensation objective. After synthesis, denser decoding supplies additional teacher-labeled windows without increasing signal storage. This construction makes temporal reuse part of synthesis and applies across condensation backends. Across five time-series classification datasets, three backends, and two storage budgets, StreamDC achieves higher mean test accuracy than independent-sequence synthesis at equal signal storage in all 30 configurations, with an average gain of 19.5 percentage points. With ShapeCond, StreamDC also averages 12.7 percentage points higher test accuracy than TexTSC across ten dataset–budget settings.

Primary area:
learning on time series and dynamical systems
Keywords:
time series, dataset condensation, dataset distillation, data-efficient learning

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_JTN4OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces StreamDC, a shared-memory approach to teacher-guided time-series dataset condensation. Instead of storing independently synthesized sequences, the method organizes the same number of signal values into class-specific segments within a concatenated temporal memory.

During synthesis, overlapping linear windows and class-local cyclic windows are jointly optimized through existing condensation objectives. Windows crossing class boundaries receive occupancy-proportional soft targets, while cyclic windows provide additional temporal origins within individual class segments. After synthesis, the memory is decoded at a finer stride, and a student model is trained using teacher predictions on the resulting windows.

Experiments across five datasets, three condensation backends, and two storage budgets evaluate the method against independent synthesis and existing time-series condensation approaches. The results generally favor the complete StreamDC pipeline. Additional experiments examine window design, post-hoc decoding, sample replication, synthesis and decoding strides, and cross-architecture transfer.

Critical strengths

1. A simple and practical approach to reusing a limited synthetic-data budget.

The paper addresses a useful question in time-series dataset condensation: how to expose synthetic signals at different temporal positions without storing additional copies. By organizing the synthetic data into a shared memory and applying overlapping and cyclic readouts, StreamDC separates the amount of stored signal from the number of training windows generated from it. The class-local cyclic readout is particularly useful at very small storage budgets, where ordinary within-class sliding windows offer few distinct temporal origins. The ability to combine this design with several existing condensation objectives makes the approach useful beyond a single backend.

2. A broad empirical evaluation supported by informative controls.

The evaluation across five datasets, three backends, and two storage budgets provides encouraging evidence for the effectiveness of the overall approach. The post-hoc decoding, sample replication, window-family ablations, and stride analyses help identify which aspects of the pipeline contribute to its performance. Although these experiments do not completely separate the effects of synthesis and student training, they provide useful evidence about the importance of the readout design. I view the combination of a simple method, compatibility with multiple backends, and broad empirical improvements as the main reason to support acceptance.

Critical weaknesses

1. Insufficient positioning relative to shared-canvas dataset distillation.

The shared-memory formulation has close connections to existing work on shared-canvas dataset distillation. In particular, PoDD (Shul et al., 2025) already studies distilling data into a shared poster, extracting overlapping crops, and jointly updating the underlying stored values. It also considers class ordering and labeling strategies for the shared layout.

These connections do not invalidate the practical value of adapting the approach to time series. The class-local cyclic readout and the separation of synthesis and decoding strides are relevant design choices in the proposed setting. However, the paper could more clearly distinguish which components build on prior work and which provide additional methodological contributions. A more explicit discussion would help assess the precise extent of the novelty.

2. Incomplete separation of the sources of empirical improvement.

The main comparisons change both the synthesis procedure and the data used for student training. The independent-synthesis baselines in Table 1 also use teacher supervision, but StreamDC provides the student with a larger collection of temporally shifted and cross-class windows. Consequently, the resulting performance gap does not by itself establish how much improvement comes specifically from joint optimization of the shared memory.

The post-hoc decoding control in Table 3 is particularly informative. Applying the StreamDC readout to independently synthesized samples already recovers a substantial part of the improvement. Relative to the reported post-hoc decoding baseline, StreamDC achieves additional gains of 3.37 and 0.51 percentage points on HAR, and 9.56 and 1.78 points on Sleep, at SPC 1 and SPC 5 respectively. These results suggest benefits beyond post-hoc decoding, with their magnitude varying across datasets and storage budgets. More extensive comparisons with matched readout augmentation, teacher supervision, and student optimization budgets would strengthen the interpretation of these gains.

There is a separate difference in student supervision in Table 2: Random, K-center, and TexTSC use hard-label training, while StreamDC uses teacher distillation. These results are useful as comparisons of complete pipelines, but they do not isolate the contribution of the condensed representation itself.

Overall, I lean toward Weak Accept. These concerns mainly affect the attribution and positioning of the contribution. Despite these limitations, I find the overall approach practically useful and the reported improvements across multiple settings encouraging. The simplicity and flexibility of the complete framework provide sufficient value to support acceptance.

Questions for authors
  1. Relationship to existing dataset distillation methods. Could the authors clarify the relationship between StreamDC and PoDD, particularly regarding shared storage, overlapping readouts, class ordering, and labeling? Which components constitute additional contributions for time-series condensation? A more explicit distinction would help assess the methodological novelty.

  2. Contribution of joint synthesis. Could the authors extend the post-hoc decoding control to representative additional datasets or backends, while matching the readout augmentation, teacher supervision, and student optimization budget? This would help determine whether the additional gains from joint optimization remain consistent beyond the currently reported settings.

  3. Comparability of student-training protocols. For Table 2, how would Random and K-center perform when trained with the same teacher supervision and temporal augmentation as StreamDC? It would be helpful to understand whether the main conclusions remain unchanged under a more closely matched training protocol.

  4. Role of cyclic windows at different stages. The current cyclic-window ablation changes both synthesis and decoding. Could the authors separately examine the effect of cyclic windows during memory optimization and during student training? This would help clarify whether their benefit primarily comes from improving the learned memory, enriching the training views, or the interaction between the two stages.

2 comments
AuthorsOpenReview ↗

Response to Reviewer JTN4 (1/2)

Thank you for your thoughtful assessment and support for the practical value of the framework. Your suggestions helped us sharpen its positioning and strengthen the evidence separating synthesis and student training. We address each point below.

A.1. Relationship to PoDD and the methodological contribution.

We appreciate the suggestion to make the connection to PoDD more explicit. Building on the discussion in Section 2 and Appendix C.1, we have expanded the main-text explanation of the shared foundations and the temporal problem that motivates our design.

At SPC1, a class has only one length-L segment, so a contained linear readout has just one start. Increasing linear overlap cannot expose the same class evidence at additional pure-class origins. StreamDC resolves this by combining a global linear memory with cycles confined to individual class segments: global windows expose contiguous and cross-class contexts, while cyclic completion supplies the missing same-class origins. Both families jointly optimize the same stored signal; decoding then uses an independently chosen stride to reuse that signal more densely.

Class ordering follows the existing idea of arranging compatible neighbors, instantiated here using distances between sampled time-series class means. For labeling, PoDD uses a fixed spatial layout or learned label field, whereas StreamDC assigns synthesis targets by temporal occupancy and uses the backend's teacher predictions for student training. PoDD, Section 4

The additional methodological contribution is a memory learned jointly for class-local temporal reuse and global contextual composition, with independent control over synthesis and decoding density. This changes which temporal contexts can be represented at fixed storage and how their supervision shapes the shared signal. In particular, ordinary contained-window overlap cannot supply the additional pure-class origins at SPC1. The linear-only and stage-separated controls directly test this design and its role in the complete synthesis-and-decoding framework.

A.2. Contribution of joint synthesis under the same hybrid readout.

We have extended the post-hoc control to all five datasets, three backends, and both budgets. Holding the proposed hybrid decoder fixed lets us examine the additional effect of optimizing the shared memory for its readouts. Joint synthesis gives higher means in 16/30 configurations; on HAR at SPC1, the gains are +3.37/+2.47/+5.51 for ShapeCond/SRe²L-style/G-VBSM. Applying the hybrid decoder after independent synthesis also improves over the original independent baseline in all 30 configurations and supplies the larger share of the overall gain. The differences below are StreamDC minus post-hoc in percentage points.

BackendDatasetSPC1 differenceSPC5 difference
ShapeCondHAR+3.37+0.51
ShapeCondFacesUCR+0.05−0.13
ShapeCondElectric−1.83−1.68
ShapeCondSleep+9.56+1.78
ShapeCondInsect−1.61+1.01
SRe²L-styleHAR+2.47+1.75
SRe²L-styleFacesUCR−0.05+0.20
SRe²L-styleElectric−10.36−1.25
SRe²L-styleSleep+11.66+0.56
SRe²L-styleInsect−3.90+0.72
G-VBSMHAR+5.51+0.21
G-VBSMFacesUCR−2.58−0.97
G-VBSMElectric−0.29−0.77
G-VBSMSleep−6.12+5.16
G-VBSMInsect+5.65−6.61

We retain joint synthesis plus hybrid decoding as the main method: synthesis adapts the shared signal to the contexts used for supervision, and decoding exposes the learned memory more densely. The complete procedure improves over the corresponding independent baseline in all 30 main configurations. The new discussion develops the interpretation of these complementary stages already provided by the original Table 3 (Section 5.4). The post-hoc controls retain the same hybrid readout, backend supervision, and epoch budget; implementation details are in Appendix E.1 (Tables 12–13).

AuthorsOpenReview ↗

Response to Reviewer JTN4 (2/2)

A.3. Random/K-center with the same teacher supervision and readout.

Following your suggestion, we gave Random and K-center both the same KL teacher supervision and the same hybrid temporal readout as StreamDC. These directly address whether the advantage persists when selected real examples receive those components too. Accuracy (%) is mean ± standard deviation; StreamDC uses ShapeCond. The final column is its margin over the stronger coreset.

DatasetSPCRandom + Hybrid + KLK-center + Hybrid + KLStreamDCDifference
HAR171.95 ± 5.3178.31 ± 2.2986.35 ± 2.10+8.04
HAR592.76 ± 0.9191.63 ± 0.4893.92 ± 0.65+1.16
FacesUCR193.07 ± 1.7691.54 ± 1.0893.80 ± 0.54+0.73
FacesUCR595.19 ± 0.7195.19 ± 0.7895.40 ± 0.62+0.21
Electric159.41 ± 7.7054.36 ± 8.3064.29 ± 3.23+4.88
Electric575.11 ± 2.6772.58 ± 1.6174.63 ± 2.76−0.48
Sleep127.99 ± 15.9213.16 ± 9.1847.86 ± 13.14+19.87
Sleep565.04 ± 2.6658.80 ± 2.8268.07 ± 5.13+3.03
Insect126.17 ± 10.8510.94 ± 0.6429.96 ± 6.25+3.79
Insect556.02 ± 1.4149.60 ± 0.5456.94 ± 1.39+0.92

With supervision and readout aligned, StreamDC has higher means in 19/20 comparisons: 9/10 against Random and 10/10 against K-center. At SPC1, it exceeds both on every dataset, by 0.73–19.87 points over the stronger coreset. Thus, the advantage of the complete synthesis-and-decoding framework persists beyond access to teacher supervision and temporal augmentation, most clearly at the tightest budget. The coresets' improvement also shows that our readout is useful beyond synthesized inputs. We expanded the discussion of the submitted Table 2 accordingly (Section 5.3); KL-only results and training details are in Appendix E.7 (Tables 22–23).

A.4. Cyclic windows during synthesis and decoding.

To separate the two stages as you suggested, we independently include or remove cyclic windows during synthesis and decoding on all five datasets with ShapeCond. L denotes linear-only and H hybrid readouts including cyclic completion; the first letter is synthesis and the second student decoding. SPC1 accuracy (%) is mean ± standard deviation.

DatasetLLLHHLHH
HAR76.73 ± 4.1684.89 ± 2.6279.07 ± 4.0686.35 ± 2.10
FacesUCR90.52 ± 1.5693.70 ± 0.6290.27 ± 0.9893.80 ± 0.54
Electric55.37 ± 10.3562.17 ± 8.9656.58 ± 3.9464.29 ± 3.23
Sleep46.84 ± 13.3846.43 ± 14.5745.06 ± 12.0247.86 ± 13.14
Insect17.67 ± 7.4529.50 ± 6.6319.96 ± 7.6629.96 ± 6.25

At SPC1, complete HH has the highest reported mean on all five datasets. With hybrid decoding fixed, cyclic synthesis gives comparable or better accuracy (HH−LH: 0.10–2.12 points). With hybrid synthesis fixed, cyclic decoding gives 2.80–10.00 points (HH−HL). These comparisons support learning the memory with the contexts in which it will be used and retaining those contexts during student training. SPC5 mean accuracies (%) are:

DatasetLLLHHLHH
HAR93.7994.0793.9193.92
FacesUCR95.5395.1995.2695.40
Electric74.7175.5874.3574.63
Sleep67.1667.5567.8868.07
Insect54.7556.6654.9656.94

At SPC5, the variants achieve broadly similar accuracy, with dataset-dependent differences. This complements the SPC1 result, where pure temporal origins are scarcest and complete HH has the highest mean throughout. We included both budgets in Appendix E.8 (Table 24) and connected this stage ablation to the storage–placement motivation in the main text (Section 5.4).

We hope these controls resolve the attribution concerns and strengthen the evidence for the practical value you identified.

Reviewer_9QEAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces a series condensation method called StreamDC. They proposed a shared time-series memory where fragments are reusable. The memory has two type of readouts to construct the global window and class local cyclic ones for class-specific patters. They jointly optimize the windows through existing condensation backends. Using a dense decoding stage which traverses the representations through a finer stride length they create richer representations to guide distillation to external student models.

Critical strengths
  • Shared-memory representation for time series both novel and practical
  • I appreciate the cross-architecture experiments that shows the feasibility of the work
  • Strong theoretical analysis and compatibility with existing condensation methods
  • I also like the guidance to distillation part.
Critical weaknesses
  • I am curious as to understanding why StreamDC maintains strong downstream performance? The analysis in the paper would be strengthened if they could empirically show feature or spectral preservation, since that is the route a lot of other time series condensation papers seem to go with.

  • UniTSC, Shao et. al, ICML 2026 (https://proceedings.mlr.press/v306/shao26b.html) is another time-series condensation framework that specifically emphasizes preserving spectral information and multivariate dependencies. A comparison against UniTSC with similar cross-architecture evaluations would strengthen the paper even more. (The GitHub: https://github.com/ZJU-DAILY/UniTSC)

  • The paper would benefit from more analysis on joint synthesis vs post-hoc concatenation ablations since at SPC 5, the benefits are much lower on both datasets. Is this a monotonically decreasing relationship with respect to samples per class? As SPC increases, the accuracy increases, but the benefits of the joint synthesis seem to be less?

  • There is no experiment showing how the generated dataset perform.

Questions for authors
  1. Please explain why StreamDR maintain downstream performance. Is it because of feature/spectral preservation or more so classification-specific information?
  2. Is StreamDC’s performance characterized by maintaining task-specific information rather than feature preservation?
  3. How will StreamDC compare against latest papers line UniTSC.
  4. Why with SPC increase, the accuracy increases but the benefits of the joint synthesis decreases?
  5. In section 3.1, the authors claim that occupancy specifies their supervision, not an observed transition probability. They then define a hybrid readout which I understand is a union of the global and class-local windows. Then equation 7 says that the synthetic targets using the hybrid windows are part of the CE classification loss. What is the motivation of doing this? It seems like in equation 8 for distillation, they do use the teacher labels. What is the benefit of using the labels for distillation but not for synthesis?
4 comments
AuthorsOpenReview ↗

Response to Reviewer 9QEA (1/2)

Thank you for your thoughtful review and helpful suggestions for understanding the learned data. We have added independent feature/spectral diagnostics, a joint-versus-post-hoc budget sweep, and UniTSC comparisons using both its native backbone and three shared student architectures. We have also expanded the explanation of supervision during synthesis and student training.

A.1. What information supports downstream performance?

We directly examined whether StreamDC readouts carry class information recognizable by a model trained independently on the full real training set, beyond the synthesis teacher.

We trained and froze a separate convolutional neural network using the full real training set; it did not participate in synthesis. Features are taken after the final convolutional block and before the final linear classification head. We analyze pure-class linear/cyclic windows, excluding cross-class compositions. The table summarizes three synthesized memories with the same evaluator. Section 5.7 reports these results (Table 5), and Appendix E.4 details the evaluation protocol.

We compare the complete method with Indep. using synthetic-to-real class-centroid distance, normalized by mean distance between real class centroids (smaller is closer). Evaluator recognition measures accuracy on pure synthetic windows under their intended labels.

DatasetSPCIndep. centroid distanceStreamDC centroid distanceStreamDC evaluator recognition (%)
HAR10.674 ± 0.0080.461 ± 0.03778.82 ± 12.88
HAR50.422 ± 0.0220.335 ± 0.01587.43 ± 3.49
Sleep12.992 ± 0.0551.215 ± 0.09865.88 ± 4.40
Sleep51.551 ± 0.0050.663 ± 0.01284.77 ± 1.99

The results show that StreamDC readouts retain class information recognizable by a model trained independently on the full real training set. Their class centroids are also closer to the corresponding real class centroids than those of independent samples in all four settings. For example, the distance on Sleep falls from 2.992 to 1.215 at SPC1 and from 1.551 to 0.663 at SPC5. This provides a direct characterization of useful feature structure beyond reporting downstream accuracy alone.

A.2. Task-specific information and feature/spectral preservation.

These perspectives are complementary. Joint synthesis optimizes the shared memory to carry useful class evidence across its linear and cyclic contexts, and dense decoding exposes that learned evidence at additional temporal positions. The feature results above show recognizable class structure under an independently trained evaluator, supporting this task-oriented interpretation of the complete method.

For the spectral question, we compute power spectra per sample before class averaging. At SPC1, a pure cyclic view is a full-length circular shift of the stored sequence and therefore preserves its discrete Fourier transform (DFT) magnitude. Correspondingly, independent samples and post-hoc pure-class readouts have identical spectral distances: 0.2727 on HAR and 0.4487 on Sleep. Yet applying StreamDC's complete decoder improves downstream accuracy from 72.77 to 82.98 on HAR and from 17.41 to 38.30 on Sleep. This illustrates why spectral similarity alone does not capture the utility of temporal reuse. In the full method, synthesis learns the shared signal under multiple class contexts, and decoding makes that learned signal available at more offsets; downstream accuracy evaluates the usefulness of this combined process.

The generated datasets are evaluated directly in Tables 1–4 of the submission by training fresh students and testing them on real held-out data. The new analysis complements this by examining synthetic signals through an independently learned real-data representation. We added the feature diagnostic to Section 5.7 (Table 5) and the spectral diagnostic to Appendix E.4 (Table 18), with their interpretation in the main text.

AuthorsOpenReview ↗

Response to Reviewer 9QEA (2/2)

A.3. Comparison with UniTSC and cross-architecture transfer.

Thank you for recommending UniTSC; it is a valuable addition to the comparison. We evaluated its official implementation on the same prepared splits and SPC budgets and added it to related work (Section 2). Below are accuracy (%) point estimates for its native PatchTST evaluation, alongside StreamDC with ShapeCond and its standard CNN student. Configuration details are in Appendix E.5 (Table 19).

DatasetUniTSC SPC1StreamDC SPC1UniTSC SPC5StreamDC SPC5
HAR59.4486.3547.8793.92
FacesUCR14.6693.8014.6695.40
Electric26.3264.2925.8774.63
Sleep42.7447.8643.1068.07
Insect12.0129.9613.0456.94

We also tested UniTSC's condensed data with three common student architectures on HAR/Sleep, following the submitted Table 4's evaluation protocol. Results are mean ± standard deviation (%) (Table 20). UniTSC uses its native labels; StreamDC retains its backend's teacher supervision.

DatasetStudentUniTSC SPC1StreamDC SPC1UniTSC SPC5StreamDC SPC5
HARCNNLN48.27 ± 2.4358.84 ± 9.6537.67 ± 5.5780.97 ± 3.62
HARTCN16.18 ± 1.6353.52 ± 10.2616.18 ± 1.6381.67 ± 7.73
HARGRU15.64 ± 6.2334.30 ± 6.7816.36 ± 6.4957.44 ± 13.91
SleepCNNLN27.92 ± 2.2738.97 ± 6.7128.11 ± 1.9352.82 ± 5.66
SleepTCN22.76 ± 14.6423.89 ± 7.0122.76 ± 14.6438.87 ± 5.13
SleepGRU23.73 ± 11.3623.15 ± 5.9124.02 ± 10.5837.69 ± 12.01

In our reproduction, StreamDC has higher reported accuracy in all ten native-configuration comparisons and 11/12 shared-architecture comparisons, including all six at SPC5. These numbers describe our reproduction under the datasets, budgets, and settings in Appendix E.5. Together with the original transfer results, they support StreamDC's usefulness across convolutional and recurrent students.

A.4. Why does the additional joint-synthesis benefit decrease with SPC?

The HAR–ShapeCond sweep compares joint synthesis with post-hoc hybrid decoding (Section 5.4; Appendix E.6, Table 21). Accuracy (%) is mean ± standard deviation; differences use the displayed means.

SPCJoint synthesisPost-hocDifference (points)
186.35 ± 2.1082.98 ± 2.81+3.37
291.75 ± 1.4190.54 ± 1.32+1.21
593.92 ± 0.6593.41 ± 0.53+0.51
1094.17 ± 0.5494.11 ± 0.59+0.06

The additional benefit decreases monotonically across these four measured budgets, with nearly identical accuracy at SPC10. This is consistent with the motivation: adapting limited stored signal to multiple contexts is particularly valuable at small budgets, whereas larger class segments already contain more within-class contexts. HAR also has less accuracy headroom at larger budgets. We added this interpretation alongside the broader controls, which show that the joint-synthesis increment varies across datasets and backends. Post-hoc decoding supplies the larger share of the overall gain. The complete synthesis-and-decoding procedure remains our main method.

A.5. Why occupancy targets for synthesis and teacher targets for students?

The targets serve complementary purposes: directing the memory toward intended class content, then aligning student supervision with each actual decoded view.

During synthesis, the teacher is fixed and the signal is optimized through Eq. (7). Occupancy prescribes the class composition: pure windows are optimized toward their class, and cross-class windows toward the proportions assigned by the layout. This supplies a class-directed objective, rather than reproducing the teacher's current prediction on a still-evolving synthetic input. Occupancy specifies intended synthetic composition, not an observed transition probability.

After synthesis, the memory is fixed and a fresh student learns from densely decoded windows. Class evidence depends on the motifs present at each offset and need not be proportional to segment duration: a short discriminative motif can be more informative than a longer portion. Teacher predictions give each view a soft target matching its learned content and conveying relationships among classes. Thus, occupancy directs what the memory should learn, while relabeling matches student targets to what the window presents. We clarified these roles alongside Eqs. (7)–(8) (Sections 3.2–3.3).

We hope the added diagnostics and comparisons address your concerns and strengthen the evidence for the framework's practical value.

I thank the authors for the detailed rebuttal. The rebuttal addresses my concerns. I will be happy to increase my score to 3.

Thank you very much for your encouraging response! We are very glad that our rebuttal has addressed your concerns, and we sincerely appreciate your positive reassessment of our work. Thank you for your time and constructive feedback throughout the review process.

Reviewer_K7qpOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses the challenges of tight storage budgets and temporal positional variation in time-series dataset condensation for classification. It proposes concatenating synthesized sequences from different classes into a shared temporal memory with the same signal storage budget. During the synthesis phase, this method extracts global linear windows and class-local cyclic windows, then integrates them with an existing condensation backend to jointly optimize the shared memory. In the downstream training phase, overlapping windows are decoded densely with finer strides. The paper also provides analysis and experimental verification regarding gradient accumulation and cyclic coverage.

Critical strengths

1.The pipeline exhibits good overall engineering performance and is easy to integrate. StreamDC, as a plug-and-play solution, can be directly applied to various existing condensation backends and achieves good performance with a fixed number of stored signal values. 2.Experimental details are transparent, providing crucial ablation analysis. The authors report the results of post-hoc concatenation and decoding, window ablations, and sample replication controls in Table 3 and the appendix, and also record the number of windows, runtime, and GPU memory usage in detail.

Critical weaknesses
  1. Limited methodological novelty compared to existing literature; the core components show limited additional gains in control experiments and are dependent on specific settings.

The overall framework of shared representations and overlapping readouts, and adjacency ordering based on class similarity has been proposed in PoDD. Multiple views (IDC) generated along the time axis and multiple local views combined with teacher soft labels to train students (LADD, SRe^2L) also have precedents. The truly novel designs in this paper focus only on class-local cyclic windows, cross-class linear windows, and the joint synthesis of shared memory; engineering increments dominate. However, the data in Table 3 shows that the actual contribution of these new components is limited: First, post-hoc concatenation and decoding accounts for the vast majority of the gains, leaving only a net gain of 0.51% for joint synthesis on HAR (SPC=5); second, cross-class windows show limited gains at SPC=5, and removing cross-class windows actually results in higher mean accuracies on HAR and Sleep than the full method. Furthermore, the improvements shown in Table 1 are derived by combining joint synthesis with dense decoding and distillation, compared to the Indep. baseline without sliding windows. Table 3, on engineering techniques, is only validated on two datasets and one backend, insufficient to demonstrate the broad and stable advantages of the new mechanism.

2.Unclear attribution of downstream performance gains in the experimental setup, the method is limited by task assumptions.

With equivalent signal storage (dKB), StreamDC increases the number of training windows during student training through dense decoding, and uses pre-trained teacher models for soft label generation and distillation. This approach uses more teacher-labeled views than the Indep. baseline, resulting in difficulty isolating the gains attributable to the shared-memory design. Simultaneously, the design of segmenting by discrete classes and cyclic shifting makes this method more suitable for position-insensitive classification tasks and potentially limits its applicability to phase-sensitive tasks. In Table 4 (SPC=1), when the student model is a GRU or TCN that relies on temporal continuity, StreamDC underperforms the Indep. baseline in multiple settings and substantially lags behind TexTSC on Insect (SPC=1) in Table 2. However, the paper lacks in-depth analysis of whether seam artifacts interfere with the learning of time-series models.

3.The theoretical analysis is superficial.

The derivation in Section 4 fails to compensate for the lack of methodological novelty. Equation (9) is merely a multivariate chain rule under differentiable slicing operations. Lemma 1 is merely a property of uniform counting on cyclic grids under stride divisibility and does not guarantee uniform coverage in non-divisible settings. Propositions 1 and 2 only provide conditional interpretations based on triangle inequalities and first-order Taylor expansions, which seem to have little practical effect.

Questions for authors

1.Given that Table 3 shows that post-hoc concatenation and decoding recovers most of the gain at SPC=5, and removing cross-class windows slightly improves accuracy, could the authors supplement the comparison results of post-hoc concatenation and decoding and removing cyclic windows on other backends and other datasets to demonstrate the true net gain of joint synthesis and cyclic windows in a wider range of settings?

2.If the independent samples synthesized from the traditional Indep. baseline are also subjected to the same number of random cyclic shifts and cross-sample one-dimensional CutMix in the downstream student training phase, so that the number of training views generated per epoch reaches N, and the pre-trained teacher model is used to generate soft labels online for distillation, how does its accuracy compare to StreamDC?

3.Please explain why StreamDC shows performance degradation on the GRU and TCN student models in Table 4 (SPC=1), and why it lags substantially behind CondTSC and TexTSC on Insect (SPC=1) in Table 2? Do cyclic concatenation and discontinuities at cross-class boundaries negatively impact recurrent or convolutional models?

4.StreamDC heavily relies on pre-trained teacher models to generate soft labels online across hundreds of densely decoded windows during downstream training. If teacher relabeling is not used during downstream student training, and instead temporal-occupancy soft labels or hard labels are used directly (consistent with the downstream protocol of TexTSC), how does StreamDC perform relative to linear canvases and independent baselines?

4 comments
AuthorsOpenReview ↗

Response to Reviewer K7qp (1/2)

Thank you for your careful review and suggestions for further analysis. We respectfully disagree with the weaknesses concerning novelty, limited effectiveness, and soundness. We first address these judgments against the stated contribution and evidence, then summarize additional tests prompted by your questions.

1. Methodological novelty: the temporal learning construction (Weakness 1).

Section 1 formulates the storage-placement problem: presenting class evidence at different temporal positions using fixed signal storage. At SPC1, a class stores exactly L time steps, admitting only one contained length-L linear window. Finer linear sampling cannot create another pure-class origin. Our insight is to combine class-local cycles, which supply those origins, with global linear windows, which retain cross-class contexts, and optimize shared coordinates under both. An independent decoding stride then controls how densely the learned memory is reused.

The original submission already cites and distinguishes all four works invoked in the review (Section 2; Appendix C.1):

  • PoDD targets compact image storage through a spatial poster, overlapping crops, class ordering, and fixed or learned layout labels. We acknowledge shared storage and ordering as precedents. Our construction couples class-local temporal cycles with global linear contexts under occupancy supervision, addressing the one-origin constraint above.
  • IDC expands compact inputs through partition-and-resize formation, including time-axis formation on speech spectrograms. StreamDC changes temporal origins while preserving sample spacing within each piece, and jointly learns the shared values used across contexts.
  • LADD adds labels for local views of distilled images. StreamDC's readouts also enter memory optimization itself; the learned representation is adapted to those contexts before student training.
  • SRe²L supplies a teacher-guided condensation and relabeling backend. We retain that objective and change the temporal representation and readout family. Backend supervision is an acknowledged foundation, not a claimed invention.

Calling this an engineering combination overlooks the temporal constraint and the construction that resolves it. Precedents for overlap or labels do not by themselves supply that construction.

2. Attribution: post-hoc already includes our readout contribution (Weakness 1).

Table 3 of the original submission already reports "post-hoc concat. + StreamDC decode", including the small HAR-SPC5 increment and window ablations. The proposed hybrid readout is itself part of StreamDC's methodological contribution. The post-hoc control already uses this design, so its gap to the full method measures the added benefit of joint synthesis, not the overall contribution evaluated in Table 1. We now quantify the allocation across all settings: decoding supplies the larger share, while synthesis increments vary (Section 5.4; Appendix E.1).

3. Broad effectiveness and a stated scope (Weaknesses 1–2).

The original results show gains in 30/30 within-backend configurations, averaging 19.48 percentage points, the highest mean in 9/10 external-method comparisons, and gains in 18/22 transfer settings. These substantial full-method gains cannot be judged solely by a component's residual gap. Broad advantage need not mean winning every setting: TexTSC, Table 2 also reports lower means than DSA on HAR/Sleep at SPC1.

We used established benchmarks, not datasets screened for position invariance: HAR, Electric, Sleep, and Insect are used in CondTSC/TexTSC, and FacesUCR in ShapeCond (Section 5.1; Appendix C.2). They cover activity, electricity, physiology, acoustics, and facial-series classification. These experiments already cover a broad range of time-series classification domains, providing empirical support for the method's usefulness across diverse applications. The original Section 1 already motivated position-varying class evidence, and Section 6 already explicitly discussed phase-sensitive tasks. This stated scope was part of the original method, not an overlooked assumption uncovered by the review.

4. Analysis explains the design; it does not substitute for novelty (Weakness 3).

Section 1 lists the problem formulation, method, and evaluation as contributions. As already stated in the original Section 4, analysis explains the design: how shared gradients couple contexts, cyclic completion adds temporal origins, and decoding changes readout density. Experiments assess practical value beyond the analytical conditions. The review's observations about standard tools and conditional bounds restate explicit features; they identify no incorrect derivation. Their simplicity does not itself establish a soundness defect. This explanatory role should be distinguished from a claim of theoretical novelty.

AuthorsOpenReview ↗

Response to Reviewer K7qp (2/2)

5. Teacher supervision is a shared backend component (Weakness 2).

Relabeling is one of the three defining stages of SRe²L, alongside Squeeze and Recover. It is part of the backend's learning procedure, not a training device introduced by StreamDC. G-VBSM also uses ensemble teacher targets, and ShapeCond replaces synthesized hard labels with teacher soft labels. The original Section 5.1 and Appendix C.2 already specify that StreamDC and Indep. retain the same backend supervision. Teacher dependence shared by both cannot, by itself, explain the measured advantage of one over the other. Removing relabeling changes the training objective; it does not invalidate the gains under the stated backend protocol. The additional views, which do differ between methods, are tested separately by the matched controls below.

Your questions nevertheless motivate useful additional tests. We summarize their findings here; full results and protocols are in the revised paper. They refine the component-level interpretation while retaining the evidence for the complete framework.

A.1. Expanded post-hoc and linear-only controls.

Appendix E.1 (Tables 12–13) covers all five datasets, three backends, and both budgets. Post-hoc uses the same hybrid decoder and backend supervision; linear-only omits cyclic windows at both stages. Post-hoc improves over Indep. in all 30 settings. Joint synthesis gives higher means than post-hoc in 16/30, including HAR at SPC1 for all three backends (+2.47 to +5.51 points). Post-hoc decoding supplies the larger share of the overall gain.

Against linear-only, StreamDC has higher means in 13/15 SPC1 settings, averaging +7.48 points; the SPC5 average difference is +0.11. The separate synthesis/decoding ablation with ShapeCond also gives the full hybrid method the highest SPC1 mean on all five datasets (Appendix E.8, Table 24). These results support the small-budget design motivation. Cross-class windows likewise give comparable HAR/Sleep performance at SPC5 but add 3.84/12.08 points at SPC1, as already reported in Table 3.

A.2. Random shifts, CutMix, and repeated exposure.

We applied random cyclic shifts and cross-sample 1D CutMix separately to independently synthesized samples with backend teacher supervision, matching per-class view counts and epoch budgets. StreamDC has higher means in 21/30 shift comparisons (one tie), 27/30 CutMix comparisons, and 29/30 replication comparisons. On HAR with ShapeCond at SPC1, shifts/CutMix/StreamDC achieve 77.54/73.78/86.35. These controls directly test whether teacher-labeled augmentation or repetition explains the complete method's gains (Appendix E.2, Tables 14–16).

A.3. Transfer, Insect, and boundary contexts.

Table 4 reports four SPC1 differences of only 0.36–1.09 points below Indep., alongside a 30.75-point gain for HAR–TCN and gains in all 11 SPC5 transfer settings. These observations do not support a general incompatibility with recurrent or temporal convolutional students.

Boundary-crossing windows are already supervised as complete views during synthesis: Eq. (7) applies occupancy targets to cross-class views and class targets to cyclic views, while classification and backend matching losses update the joined signal on both sides. These contexts shape the memory before student training (Section 3.2).

On Insect at SPC1, linear-only reaches 17.67; adding cyclic decoding reaches 29.50, close to the full method's 29.96 and above Indep.'s 12.48 (Table 24). Thus, the observed gain from cyclic views does not support blaming them for the gap to CondTSC/TexTSC at this budget. At SPC5, StreamDC reaches 56.94 and exceeds both methods (Table 2).

A.4. Alternative targets and controls with shared teacher access.

On fixed StreamDC views, hard labels take the argmax of occupancy: mixed windows receive the largest-occupancy class; pure/cyclic windows retain their class. For SRe²L-style, teacher/hard-label accuracy is 80.73/75.40 and 93.73/88.92 on HAR at SPC1/5; on Sleep it is 49.39/47.50 and 65.80/44.25. With ShapeCond, teacher/occupancy/hard results are 86.35/80.16/79.37 on HAR-SPC1 and 68.07/54.46/42.08 on Sleep-SPC5. Occupancy soft targets have higher means than hard labels in 9/10 settings, while teacher relabeling is higher in all ten (Appendix E.3, Table 17). Occupancy specifies intended composition; teacher predictions reflect the actual decoded view (Section 3.3).

Giving Random and K-center the same KL supervision and hybrid readout still leaves StreamDC higher in 19/20 comparisons (Appendix E.7, Table 23). Together with the original matched-backend results, this supports the framework's value with those components shared. We respectfully request reassessment of the novelty, effectiveness, and soundness judgments in light of these distinctions and evidence.

Reviewer_K7qpOpenReview ↗

Response to Author's Rebuttal

Thanks to the authors for their response, but my core concerns remain unresolved:

  1. Methodological novelty: The overall framework is an incremental extension of existing shared-canvas and multi-view distillation methods. PoDD (Shul et al., 2025) already distills data into a single shared representation via overlapping readouts, similarity-based class ordering, and layout supervision. Multi-view temporal expansions and pairing local views with teacher soft labels also follow IDC (Kim et al., 2022), SRe2L (Yin et al., 2023), and LADD (Kang et al., 2025). Adapting a 2D canvas to a 1D concatenated sequence with class-local cyclic shifts and CutMix-style occupancy labels (Yun et al., 2019) is a direct adaptation to time series rather than a substantial methodological novelty. Moreover, the theoretical analysis in Section 4 mainly restates standard calculus and grid-counting properties.

  2. Net gains of core components: Extended experiments show that most performance improvements come from post-hoc concatenation and decoding, while joint synthesis does not show consistent gains. Across all 30 expanded settings, joint synthesis outperforms post-hoc decoding in 16/30 cases and yields lower accuracy in the remaining 14/30. For instance, it drops by 10.36% on Electric at SPC=1 (SRe2L-style) and by 6.61% on Insect at SPC=5 (G-VBSM), while its difference on HAR decreases to +0.06% at SPC=10. At SPC=5, cyclic windows also produce an average difference of +0.11% over linear-only readouts. Removing cross-class windows at SPC=5 yields comparable or slightly higher accuracy on HAR (93.95% vs. 93.92%) and Sleep (68.76% vs. 68.07%).

  3. Empirical attribution: Coreset and relabeling experiments under aligned settings indicate that the improvements mainly come from dense window augmentation and teacher distillation, rather than the condensed representation itself. When Random selection uses the same hybrid readout and teacher KL supervision at SPC=5, it achieves comparable accuracy to StreamDC on HAR (92.76% vs. 93.92%) and higher accuracy on Electric (75.11% vs. 74.63%). Conversely, replacing teacher relabeling with hard labels reduces accuracy from 68.07% to 42.08% on Sleep at SPC=5.

Therefore, I will maintain the original score.

References:

  1. Jang-Hyun Kim et al. Dataset condensation via efficient synthetic-data parameterization. In ICML, 2022.
  2. Seoungyoon Kang et al. Label-augmented dataset distillation. In WACV, 2025.
  3. Asaf Shul et al. Distilling datasets into less than one image. TMLR, 2025.
  4. Zeyuan Yin et al. Squeeze, recover and relabel: Dataset condensation at ImageNet scale from a new perspective. In NeurIPS, 2023.
  5. Sangdoo Yun et al. CutMix: Regularization strategy to train strong classifiers with localizable features. In ICCV, 2019.
AuthorsOpenReview ↗

Response to the Reviewer’s Follow-up

Thank you for the follow-up. We respectfully disagree with the conclusions drawn from the cited results: the comparisons address different parts of our method.

1. The methodological design addresses a specific storage-placement constraint.

StreamDC asks how fixed signal storage can expose class evidence at different temporal positions. Under tight storage budgets, short class segments restrict where a full-length window can start; denser linear sampling cannot extend that range.

PoDD, IDC, and LADD share the broader principle of deriving multiple views or richer supervision from compact stored data, but instantiate it differently: overlapping poster crops, partition-and-resize formation, and local-view labels. This shared principle does not make their methodological contributions interchangeable.

Our construction couples class-local cycles, which create additional pure-class temporal origins, with global linear reads, which retain cross-class contexts. Cyclic views retain class targets; cross-class views receive temporal-occupancy targets. Both families jointly optimize the same stored coordinates, adapting content to its multiple placements and contexts. After synthesis, a separate stride controls dense decoding of the frozen memory. The reads preserve sample spacing within each piece. Calling this a "direct adaptation" does not explain why these changes to temporal origins, class contexts, and their joint optimization are merely engineering rather than a methodological response to the storage-placement constraint. As already explained in our first response, Section 4 clarifies shared-coordinate optimization, cyclic coverage, and the roles of synthesis and decoding; experiments assess practical value beyond the analytical conditions.

2. Post-hoc decoding retains a core design of StreamDC.

The hybrid readout is our design whether used during synthesis or decoding. Post-hoc applies it to independently synthesized data. The gap to a control that already uses our readout cannot measure the contribution of that readout. The 16/30 result and negative increments characterize the additional synthesis stage. Revised Section 5.4 explicitly states that post-hoc decoding "supplies the larger share of the overall gain"; Appendix E.1 explains that synthesis increments depend on dataset, backend, and budget. This does not remove the readout from our contribution or negate the full method's 30/30 improvements over Indep. at equal signal storage.

The budget comparisons likewise need their full context:

  • The cited HAR +0.06 at SPC10 is the endpoint of +3.37/+1.21/+0.51/+0.06 at SPC1/2/5/10, documenting diminishing additional synthesis gains as storage increases (Section 5.4; Table 21).
  • The cyclic-window average is +0.11 points at SPC5, but +7.48 at SPC1, with higher means in 13/15 settings (Appendix E.1).
  • Cross-class windows yield similar SPC5 accuracies but add 3.84/12.08 points on HAR/Sleep at SPC1 (Table 3).

Both reviews select the higher-budget near-equalities to characterize the design while leaving its larger gains out of the argument. This selective reading does not substantiate the broader dismissal of the components.

3. The matched coreset evidence does not support the stated attribution.

With the same hybrid readout and student-supervision protocol, StreamDC has higher means in 19/20 comparisons, including every SPC1 comparison. Against the same Random baseline cited in the follow-up, the SPC1 gains are 14.40 points on HAR and 19.87 on Sleep. Across all five datasets and both budgets, the mean gains are 4.85 points over Random and 9.51 over K-center (Table 23).

These controls hold readout and supervision fixed and vary the stored data. The follow-up cites two SPC5 Random comparisons, including the only lower mean among twenty comparisons, without accounting for the remaining evidence. That selection does not support the claim that gains arise "rather than the condensed representation itself." The attribution must account for the complete matched results, not just their least favorable entries.

Our first response already explained that teacher relabeling, a defining stage of SRe²L (Yin et al., 2023), is part of the backend shared by StreamDC and Indep. (original Section 5.1; Appendix C.2). Removing it changes the backend itself. The resulting drop is not a valid test of StreamDC's contribution under the common backend protocol. The follow-up repeats that drop without addressing our explanation or the matched comparisons with supervision held fixed.

Finally, all five cited works were already discussed in the original submission (Section 2; Appendix C.1). Those acknowledged foundations should be assessed alongside our specific temporal construction. We respectfully ask that the assessment be based on the contribution actually claimed and the complete controlled evidence, rather than on selected component outcomes.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Responses

We thank all reviewers for their careful reading and constructive suggestions. We have done our best to address their concerns in the individual responses and to improve the manuscript through additional experiments and clearer explanations. Added and revised text is highlighted in blue for ease of review.

StreamDC addresses the storage–placement problem in time-series condensation: presenting class evidence at different temporal positions under fixed signal storage. It jointly learns a shared memory through global linear and class-local cyclic windows, then decodes that memory densely. The proposed readout and joint synthesis together constitute the method. Across five established datasets and three backends, the complete framework improves mean accuracy over independent synthesis in all 30 main configurations.

The main clarifications and additions are:

  • Methodological contribution and design rationale. We expanded the comparison with prior shared-storage and multi-view methods, emphasizing how class-local cycles provide additional pure-class temporal origins that linear overlap cannot supply at the smallest budget. We also clarified the analysis's explanatory role: it explains the design choices, while experiments assess practical value beyond its assumptions (Sections 2 and 4; Appendix C.1).

  • Contributions of synthesis and decoding. Expanded post-hoc, linear-only, stage-specific, and budget controls clarify their respective roles. Post-hoc decoding supplies the larger share of the overall gain and already incorporates our proposed hybrid readout; joint synthesis gives higher means in 16/30 settings. The complete synthesis-and-decoding framework remains our main method (Section 5.4; Appendices E.1, E.6, E.8).

  • Comparisons and supervision. We added random-shift, CutMix, replication, and coreset controls. With the same teacher supervision and hybrid readout, StreamDC has higher means than Random/K-center in 19/20 comparisons. We clarified the complementary roles of occupancy and teacher targets and added alternative-label diagnostics. Following the reviewers' recommendation, we also cited and reproduced UniTSC with its native backbone and shared student architectures (Sections 3.2–3.3 and 5.3; Appendices E.2–E.3, E.5, E.7).

  • Information retained in the learned data. New diagnostics use an independent convolutional neural network trained on the full real training set. StreamDC readouts carry recognizable class information and have closer synthetic-to-real class centroids in all four evaluated settings. Complementary spectral analysis explains why spectral similarity alone does not characterize the training utility of temporal reuse. The feature results now appear directly in the main text (Section 5.7, Table 5; Appendix E.4).

The author has responded to everyone's comments. Please reply when it is convenient for you.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.