acceptodds
Under review as a conference paper at ICLR 2027

Learning Event Representations by Localizing Temporal Reversals

Abstract

Self-supervised learning has proved useful for time-series classification and forecasting, but common pretraining objectives do not explicitly require features that make individual events easy to recognize and locate. We introduce Temporal Reversal Discrimination (TRD), which trains an encoder to find a short segment reversed within an otherwise unchanged signal. A single linear auxiliary head learns this self-labeled task alongside the original pretraining objective and is discarded afterward, leaving inference unchanged. We evaluate TS2Vec, PatchTST and PMT on heart sounds and wearable operations using frozen features, small labeled training sets and separate labeled validation data. On CirCor heart sounds, linear classifiers fitted on 5% of training patients gain 14–47 points in mean onset macro-F over pretraining without TRD. The benefit persists across the tested classifiers and labeled subsets. On OpenPack, linear classifiers trained on five labeled minutes per training person gain 2–18 points in mean macro-F, with mixed results at other settings. PatchTST and PMT also improve on seismic and ECG localization. In the CirCor controls, predicting the reversed positions produces better event features than assigning one reversal label to the window, even when the latter task is learned well. Comparisons with block shuffling show that the better manipulation depends on the encoder and dataset. On CirCor, all three encoders fine-tuned from TRD weights outperform their counterparts fine-tuned after original pretraining or trained from scratch. With the same separately labeled validation cohort, TRD fine-tuning on seven patient groups matches or exceeds the mean onset F of training from scratch on 31 groups.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Self-supervised learning, Time-series representation learning, Temporal reversal, Event detection, Event localization, Label-efficient learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_6hAcOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the limitations of existing self-supervised learning methods for time series in event detection and temporal localization tasks. While traditional contrastive learning and masked reconstruction approaches can learn general time series features, they may not effectively capture the occurrence positions, temporal boundaries, and local temporal structures of events. To address this issue, the authors propose Temporal Reversal Discrimination (TRD), which randomly selects a continuous segment from unlabeled time series, reverses it along the time axis, and uses a lightweight linear classification head to predict which time positions belong to the reversed region. This auxiliary task is jointly trained with the original self-supervised objective, encouraging the encoder to learn features related to event patterns, temporal order, and local variations. The additional classification head is removed after pretraining, thus incurring no extra inference cost downstream. The authors integrate TRD into three distinct architectures—TS2Vec, PatchTST, and PMT—and evaluate them on tasks including phonocardiogram recognition, wearable-based action recognition, seismic arrival time detection, and ECG waveform localization. Results show that on the CirCor phonocardiogram dataset, a linear classifier trained with 5% of patient data achieves an improvement of approximately 14–47 percentage points in macro-F1 score for event onset localization; on the OpenPack dataset, the main linear classification setting improves performance by about 2–18 percentage points. Furthermore, in supervised fine-tuning experiments on CirCor, TRD pretraining achieves results comparable to or better than training from scratch, even with fewer labeled samples. The authors also analyze the impact of supervision granularity and temporal perturbation strategies on representation learning by comparing with Dense Localization, Window-level Reversal Classification, and Block Shuffling. Overall, this work demonstrates that a simple local temporal reversal localization task can enhance the ability of existing self-supervised encoders to represent real-world event temporal structures, particularly benefiting event detection and localization scenarios with limited labeled data.

Critical strengths

Strength 1. TRD does not require modifying the original encoder architecture, nor does it rely on additional event annotations. It can enhance the pre-training target merely through local time reversal and lightweight auxiliary classification headers. The author verified the method on three different structures, namely TS2Vec, PatchTST and PMT, indicating that it has certain architectural adaptability. Strength 2. It performs outstandingly in low-labeled data scenarios. In the CirCor experiment, TRD achieved a 14-47 percentage point increase in Linear Probe Macro-F1 under a 5% training patient setting. The supervised fine-tuning experiment further indicates that a small amount of labeled data combined with TRD pre-training can achieve or exceed the results of de novo training with more labeled data, demonstrating strong data efficiency. Strength 3. This work not only tested End-to-end Fine-tuning, but also froze the Encoder, evaluated the feature quality using Linear Probe and Temporal Classifier, and compared it with paired random seeds and different annotation budgets. This helps to distinguish the impact of the quality of pre-trained representations on the learning ability of downstream models.

Critical weaknesses

Weakness 1. The method's innovation is relatively limited. Time reversal, temporal order prediction, and local perturbation detection have all appeared in existing self-supervised learning research. The main distinction of TRD lies in combining local time reversal with token-level dense classification as an auxiliary pretraining objective. Although the experimental results are valuable, the paper still needs to further clarify its unique contributions compared to existing methods such as Temporal Order Prediction, Dense Corruption Detection, and AnomalyBERT. Weakness 2. The model may learn artificial artifacts from perturbations rather than genuine event semantics. TRD constructs training samples by reversing local signals and smoothing boundaries via crossfade, which could lead the model to rely on reversed boundaries, statistical changes in signals, or artificially generated features from crossfade to identify reversed regions—without necessarily understanding the true temporal direction or physiological meaning of events. The Discussion section also acknowledges that the contribution of temporal direction itself has not been fully disentangled from interval insertion and crossfade cues. This is, in my view, the most critical methodological issue. Weakness 3. Performance improvements lack consistency across datasets and models. While performance gains on CirCor are highly significant, certain temporal classifier settings on OpenPack show degradation, and TS2Vec’s average F1 score on LUDB drops from 58.83% to 54.01%. Moreover, block shuffling outperforms temporal reversal in multiple TS2Vec configurations. This suggests that temporal reversal may not universally be the optimal self-supervised task, and the method’s effectiveness might heavily depend on data characteristics and encoder architecture.

Questions for authors

Temporal reversal, temporal-order prediction, and dense corruption detection have been explored in prior self-supervised learning studies. Could the authors clarify the main methodological novelty of TRD compared with existing approaches such as AnomalyBERT, temporal-order prediction, and dense temporal localization methods? In particular, what distinguishes the proposed reversal-localization objective beyond combining existing temporal transformations with token-level binary classification?

1 comment
AuthorsOpenReview ↗

Response to Reviewer 6hAc

Thank you for the careful review. The general response above summarizes the contribution and the revision; below we answer your points directly.

W1 and the question, novelty. One plug-in objective, added unchanged to three existing encoders, makes them much better event encoders, and the approaches you name do not do this in our experiments. The revised introduction now states this claim and places TRD among per-position manipulation-detection objectives.

  • Temporal-order prediction assigns one label to a whole input. Per-position targets beat a pooled label on the same reversal by 14.6 points (TS2Vec) and 17.9 (PMT) in every seed, even where the pooled task is learned well (Table 4a).
  • AnomalyBERT deploys the detector itself as an anomaly score. Here the head is discarded and we score real events in clean signals. A control modelled on its corruption transfers less well than reversal (80.7 against 85.7, Table E.3).
  • Dense localization in video (Dave et al., 2024) is the closest precedent. It motivates frame-level targets by pretext difficulty; we show that the dense advantage also holds when both pretexts are learned.

The new controls also show what the objective needs: content moved out of its temporal order, and no cheaper cue that marks the interval (W2). We are not aware of these results in prior work on reversal or order pretexts for time series.

W2, artefacts or event semantics. The encoder is not relying on artefacts; the gain comes from reordered content. The submission left this open, and new Section 4.3 tests it.

  • Order-only and direction-only reversal, which cut and join the signal at the same places, differ by 15.3 points on test (88.3 against 73.0, every seed; Table 5). The first reverses the order of blocks and plays each forward; the second keeps the order and plays each block backward.
  • Training on the splices alone, which the head learns well, reaches 59.9, below a plain loudness change (69.1).
  • With only the fade bands kept, the trained heads are at chance (AUROC 0.49–0.57, against 0.91–0.99 on real reversals; Appendix I).
  • A cheap cue blocks the benefit when it can solve the task: adding a loudness marker to shuffled intervals lowers PMT from 90.3 to 69.0, and the head then reads the marker.
  • The arms that reorder content cut wrong-type S1/S2 calls to 3–4%, against 12–16% for the others (Table E.9).

The Discussion sentence you quote now reports this result.

W3, consistency and shuffle. The gains are consistent for PMT and PatchTST on all four datasets; the exception is TS2Vec on OpenPack and LUDB. Section 3.2 now names this exception and points to its causes. PMT improves in 42 of 42 paired test comparisons and PatchTST in 38 of 42. The LUDB drop you quote for TS2Vec (58.83 to 54.01) uses classifiers trained for 40 updates; with 400, TS2Vec gains 7.5–11.2 points at every label budget (Section 3.2, Table D.9). On OpenPack, TS2Vec needs longer blocks: order-only reversal gains 0.7 points with 1 s blocks and 21.8 with 4 s blocks (validation; Section 4.3, Table E.10).

Shuffle's lead on TS2Vec fits the account in Section 4.3, and Section 4.2 and the abstract now present it this way. Shuffle is one of our controls, and it reorders content, which Section 4.3 identifies as the source of most of the gain. On TS2Vec, order-only reversal, which reorders blocks and plays each forward, also beats TRD, by 10–13 points with blocks of 100 ms or longer (Table 5). The paper's claim is that the objective improves each encoder's own pretraining, which holds in all 18 primary comparisons.

Reviewer_ZX58OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper is based on the premise that time series encoders have the potential to perform better but are currently held back by poor training strategy. In that context, an interval is randomly chosen which is reversed. A linear head on the encoder’s token-wise features predicts which tokens lie in the reversed regime. Based on this, a binary cross entropy loss can be added to the native loss. This addition is called Temporal Reversal Discrimination loss and the paper’s key emphasis is that this greatly improves the encoder’s feature extraction for downstream tasks. For those, they show results with frozen encoder weights as well as finetuned encoder weights - in both cases, the models trained with TRD loss mostly outperform the native encoder models.

Critical strengths

The most interesting part of the paper to me was the formulation of TRD - discussed in the main text as well . The idea behind adding in crossfade at the edges of the reversed interval was a thoughtful touch. This matters because the method is simple, so it would be easy for others to adopt.

Another strength of the paper was the variety of experiments shown, and under controlled situations which allow a scientific reader to isolate whether a certain benefit came from TRD loss and evaluate its merit.

Critical weaknesses

It is less convincing that the paper does a great job of convincing the reader that it is only temporal reversal as promised in the initial parts, which is the key to improving encoder performance. In a lot of the appendix cases, block shuffling often beats temporal reversal. Furthermore, Appendix I also mentions how reversing changes the covariance and crossfading can introduce a local variance cue, which the model could pick up on. To me, it sounds as if local corruption within a signal helps generally. TRD seems to work but it is not clear if only TRD works.

Another weakness is the claim of generality. Consistent improvements for all three encoders by incorporating TRD loss are confined to the CirCor dataset. TS2Vec is worse on ¾ OpenPack settings and all LUDB. In the LUDB case, beats are pre-cropped which detracts from a real world-scenario. The authors acknowledge in Appendix D.7 that CREW shares ~94% of stations between training and test; this should be stated in the main text. Ideally, results should be reported on unshared stations.

Questions for authors
  1. Can the authors show that reversal’s benefit remains when edge cues are removed?
  2. On TS2Vec/CirCor, shuffle beats reversal. Could the authors add shuffle to Table E.13 to show whether it also identifies S1 vs S2 better, or only finds their timing better?
  3. For the TRD training case with 7 groups (matched up against training from scratch with 31 groups), would results still hold with a smaller validation set?
  4. On LUDB, is TS2Vec still worse with TRD if given a stronger classifier like the patch encoders get?
  5. For CREW, how do results look on the ~6% of test stations not seen in training? Do the fitting traces also share stations with the test set?
  6. How were the span and lambda tuned? Could that explain weaker results for TRD in some cases?
2 comments
AuthorsOpenReview ↗

Response to Reviewer ZX58

Thank you for the specific questions; each led to a change in the paper. The general response above summarizes the contribution and the revision; below we answer your points directly.

W1, is it only reversal? No. The claim is the per-position objective, and reversal is one instance; block shuffling was already in the submission as a control. Section 4.2 now introduces shuffle this way, and new Section 4.3 shows why both work. As you suggest, local corruption helps in general: a per-position target on a loudness change takes PMT from 45.5 to 69.1. The larger part of the gain, to 88–91, requires reordering content, which reversal and shuffling both do (Table 5). The heads do not use the crossfade cue noted in Appendix I: with only the fade bands kept, they detect the interval at chance (AUROC 0.49–0.57, against 0.91–0.99 on real reversals).

W2, generality. For PMT and PatchTST, the consistent gains extend from CirCor to all four datasets: PMT improves in all 42 paired test comparisons across CirCor, OpenPack, CREW and LUDB, and PatchTST in 38 of 42. The ¾ OpenPack settings and the LUDB results you count are TS2Vec's, and Section 3.2 now says so. On OpenPack, TS2Vec needs longer blocks (+21.8 points with 4 s blocks, against +0.7 with 1 s; validation; Section 4.3, Table E.10), and on LUDB it gains once the classifier trains longer (Q4).

Section 3.1 now states the 94% CREW station overlap, and the gains hold on stations never seen in pretraining (Q5). It also states that LUDB windows are placed from the beat annotations, so the LUDB task is within-beat onset localization. LUDB is a supporting study; the main claims rest on CirCor and OpenPack, which use continuous recordings.

Q1, does the benefit remain without edge cues? Yes. Order-only and direction-only reversal, which cut and join the signal at the same places, differ by 15.3 points on test (88.3 against 73.0, Table 5), and training on the splices alone, with nothing reordered or reversed, reaches only 59.9. Section 4.3 now reports both.

Q2, shuffle on Table E.13. Shuffle identifies S1 and S2 better as well as timing them better. Table E.16 adds it: with true intervals supplied, S1/S2 balanced accuracy on TS2Vec is 89.1 native, 95.1 with reversal and 98.2 with shuffle.

Q3, a smaller validation set. The size of the validation set hardly matters: selecting the epoch on 7 groups instead of 31 costs 0.15–0.24 points (Section 3.3, Appendix E.4). In that re-run, scored by validation state F1, seven-group TRD stays above 31-group scratch training on PMT (+1.2 to +1.8), is seed-dependent on PatchTST (−2.1 to +3.8) and is 0.1–0.2 points below on TS2Vec.

Q4, is TS2Vec still worse on LUDB with a stronger classifier? No, it then gains. With classifiers trained for 400 updates instead of 40, TRD improves TS2Vec at every label budget, by 7.5–11.2 points, with 59 of 60 pairs positive. PMT and PatchTST also gain at every budget. Section 3.2 now reports this (Table D.9).

Q5, CREW on unseen stations. The gains hold on stations never seen in pretraining. Section 3.1 now states this and the 94% overlap (Table D.14):

Encoder, fitting tracesAll test tracesUnseen stations
PMT, 74+36.5+38.1
PMT, 147+31.5+30.7
PatchTST, 74+14.0+13.6
PatchTST, 147+6.9+11.7

The slice is small (65–79 of 3,130 traces per seed). The fitting traces do share stations with the test set (34–66 stations), and gains on test traces outside those stations are similar (+37.1 against +36.5 for PMT at 74 traces).

Q6, tuning. Neither was tuned per encoder. The weight (1.5) is shared by all encoders and datasets, and the CirCor span was chosen once on a short PMT validation run and shared (Section 2.1). On PMT/CirCor every tested weight (0.5–4.5) and span (0.2–1.6 s) improves on native training, by 38–46 points (Appendix G.1). Tuning could therefore explain part of the weaker TS2Vec results: the block-length result in W2 shows that TS2Vec is sensitive to the scale of the manipulation.

Thank you for the detailed rebuttal and additional experiments.

My remaining concern is primarily about the framing and generality of the contribution. The new results suggest that temporal reordering and dense localization supervision, rather than reversal specifically, drive much of the improvement. Moreover, some gains remain sensitive to encoder architecture, manipulation scale, and downstream classifier training. The small unseen-station subset on CREW also limits the strength of the generalization claim.

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

This paper proposes Temporal Reversal Discrimination (TRD), a self-supervised auxiliary objective for learning event representations from time series. The method randomly reverses a local temporal segment and trains an encoder to identify the reversed positions using token-level binary classification. TRD is incorporated into three existing self-supervised models and evaluated on several event recognition and localization tasks. The reported results show substantial improvements in some settings, particularly on the CirCor dataset.

Critical strengths

N/A

Critical weaknesses
  1. Limited methodological novelty. The proposed method is essentially an additional binary classification objective applied to locally reversed temporal segments. Previous self-supervised learning studies have extensively explored temporal reversal, permutation-based pretext tasks, and corruption detection. The contribution appears incremental, and the paper does not sufficiently establish what fundamentally new representation-learning mechanism is introduced.

  2. Unclear advantages of temporal reversal over other transformations. The experimental results suggest that temporal reversal is not consistently the most effective pretext task. In particular, the block-shuffling baseline substantially outperforms TRD in several settings, especially with TS2Vec. This raises questions about whether the improvements come specifically from learning temporal reversal or simply from adding dense corruption-based supervision.

  3. Inconsistent improvements across datasets and settings. Although the results on CirCor are impressive, the improvements are considerably less consistent on other datasets. Some OpenPack and LUDB configurations even show performance degradation. These results weaken the claim that TRD is a generally effective auxiliary objective for learning event representations.

  4. Insufficient understanding of what the auxiliary task actually learns. It remains unclear whether the encoder learns meaningful temporal event structures or simply exploits artificial cues introduced by segment reversal and crossfading. Although the paper provides several corruption controls, the contribution of these low-level artifacts has not been fully isolated. Further investigation is needed to support the proposed explanation of why TRD improves downstream event representations.

Questions for authors

N/A

1 comment
AuthorsOpenReview ↗

Response to Reviewer dhzG

Thank you for the review. The general response above summarizes the contribution and the revision; below we answer your points directly.

W1, novelty. To our knowledge, no prior work shows that a per-position manipulation-detection objective improves event identity and timing in time-series encoders, or what such an objective needs in order to work. The revised introduction now states this claim. The forms of temporal supervision the review mentions do not deliver it in our experiments:

  • A whole-window reversal label, as in earlier reversal pretexts, is learned almost perfectly by PMT (99.8% training balanced accuracy) yet transfers 24 points worse than per-position targets (test onset F1 64.7 against 88.6, Table 4a). TS2Vec shows the same (65.9 against 73.8).
  • A corruption modelled on AnomalyBERT reaches 80.7 on TS2Vec, against 85.7 for reversal (Table E.3).
  • Added to PPT, a published patch-permutation pretext, TRD still raises every OpenPack comparison, by 7.5–14.6 points (Table D.2).

The introduction and related work now place TRD among per-position manipulation-detection objectives, including the closest precedents, frame-level out-of-order localization in video (Dave et al., 2024) and ShED (Tamkin et al., 2021), which the revision adds. Neither evaluates event identity and timing in time series.

W2, reversal or dense corruption in general? Both contribute, and the larger part comes from moving content out of its temporal order. New Section 4.3 separates the two. A per-position target on a loudness change, which reorders nothing, takes PMT from 45.5 to 69.1 (Table E.9). Cut the reversed interval into blocks: reversing the block order while playing each block forward (order-only reversal) takes it to 88–91, as TRD does, while keeping the order and playing each 50 ms block backward (direction-only reversal) reaches only 73.0 (Table 5). Block shuffling, one of our own controls, also reorders content, which is why it does well; Section 4.2 and the abstract now say so. We evaluate reversal in depth because it has no block length to set.

W3, consistency. The gains are consistent for PMT and PatchTST on all four datasets; the exception is TS2Vec on OpenPack and LUDB. Section 3.2 now names this exception and points to its causes. PMT improves in 42 of 42 paired test comparisons, PatchTST in 38 of 42, and TS2Vec in all 12 on CirCor. The revision traces the TS2Vec exceptions to two causes. On OpenPack, TS2Vec needs longer blocks: order-only reversal gains 0.7 points with 1 s blocks and 21.8 with 4 s blocks (validation; Section 4.3, Table E.10). On LUDB, it gains 7.5–11.2 points at every label budget once classifiers train for 400 updates instead of 40 (Section 3.2, Table D.9).

W4, event structure or artefacts? The encoder is not exploiting artefacts; the gain comes from reordered content. New Section 4.3 tests this, and the Discussion now reports it.

  • Order-only and direction-only reversal, which cut and join the signal at the same places, differ by 15.3 points on test, in every seed (88.3 against 73.0, Table 5).
  • Training on the splices alone, which the head learns well (0.89 training balanced accuracy), reaches 59.9, below a plain loudness change (69.1).
  • With only the fade bands kept and the interior restored, the trained heads are at chance (AUROC 0.49–0.57, against 0.91–0.99 on real reversals; Appendix I).
  • The arms that reorder content reduce the errors that depend on temporal context. Spurious onset calls fall to 15–20%, against 40–46% for the other arms, and wrong-type S1/S2 calls to 3–4%, against 12–16% (Table E.9).

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response to All Reviewers

We thank the reviewers. One plug-in objective, added unchanged to three existing time-series encoders, makes them much better event encoders, and the well-explored alternatives the reviews name do not do this in our experiments. All 18 primary comparisons improve (14–47 points on CirCor, 2–18 on OpenPack), and on CirCor, fine-tuning from it beats training from scratch in every paired run. A whole-window reversal label, the form used in prior work, gives less than half the gain of per-position targets (Table 4a), and a corruption modelled on AnomalyBERT transfers less well (Table E.3). To our knowledge, no prior work shows that a per-position manipulation-detection objective improves event identity and timing in time-series encoders, or what such an objective needs in order to work. The revised introduction states this claim and places TRD among per-position manipulation-detection objectives, now including ShED (Tamkin et al., 2021).

The reviews share three concerns.

  1. The paper does not claim that reversal is the only or the best manipulation; its claim is the per-position objective. Block shuffling is our own control: the same objective with shuffled blocks in place of a reversed interval. Its strong results therefore support the claim; Section 4.2 and the abstract now present it this way.
  2. The gains are consistent for PMT and PatchTST on all four datasets: PMT improves in 42 of 42 paired test comparisons and PatchTST in 38 of 42. The exception is TS2Vec on OpenPack and LUDB, which the revision traces to two causes: TS2Vec needs longer blocks (Section 4.3) and a longer-trained classifier (Section 3.2).
  3. The encoder is not exploiting artefacts. The submission left this open; new Section 4.3 tests it by splitting TRD's reversal into its two halves. Cut the reversed interval into blocks. Reversing the block order while playing each block forward (order-only reversal) matches TRD: 88–91 against 90.3 with PMT on CirCor (test). Keeping the block order and playing each 50 ms block backward (direction-only reversal) reaches only 73.0. Both cut and join the signal at the same places, so the splices cannot explain the difference. Trained on the splices alone (each block shifted by 5–20 ms, so the joins are real but nothing is reordered or reversed), the encoder reaches 59.9, and the trained heads ignore the crossfade (AUROC about 0.5).

The revised PDF marks changes in green (ZX58’s questions) and red (shared concerns). Section 4.3 and Table 5 are new, and their test scores were read once, for a list of runs fixed in advance. The abstract, introduction and Discussion now state the contribution and its novelty as above, and new appendix results answer the individual questions. Wording changes made only for readability are not coloured.

Dear Reviewers,

We know this is a busy period, so just a short reminder that the discussion closes on 18 November (AoE). Our response and the revised PDF are posted above. We are happy to answer any follow-up questions or clarify anything before then.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.