acceptodds
Under review as a conference paper at ICLR 2027

Learning Representations for Multivariate Time Series Classification via Context Prediction

Abstract

Self-supervised representation learning for multivariate time series classification is dominated by contrastive learning and masked modeling. However, these paradigms incur substantial computational overhead and often rely on complex data augmentations or reconstruction decoders that introduce domain-specific biases. In this work, we propose **C-Pred**, a lightweight self-supervised framework based on next-patch context prediction. C-Pred formulates self-supervised learning as a -way classification problem: given a reference patch, the model identifies its immediate temporal neighbor from candidate patches within the same instance. This lightweight formulation, requiring only patch-level encoding rather than full-sequence processing, makes **channel-independent encoding** practical: each channel is processed independently by a shared univariate backbone, allowing a single model to accommodate arbitrary channel dimensions without increasing the backbone size. By requiring only a small number of temporal patches and no data augmentation, cross-instance negative sampling, or signal reconstruction, C-Pred substantially reduces the computational cost of self-supervision. Across six large-scale benchmarks (20k instances) and 28 UEA multivariate datasets, C-Pred matches or exceeds state-of-the-art representation quality while achieving up to a **7.5 speedup** in training time over competitive baselines. On the large-scale benchmarks, channel-independent encoding further improves accuracy by 2.99 percentage points over channel mixing, achieving 84.36% average accuracy.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
self-supervised learning, time series classification, pretext tasks, representation 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_Z43VOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The authors propose a self-supervised pretraining method for time-series classification models using a simple pretext task. They evaluate their pretraining framework under a simple evaluation setup with the InceptionTime architecture and compare it against several SSL baselines.

Critical strengths

The proposed method is very compute-efficient during both pretraining and inference, which also enables easy deployment.

Critical weaknesses

1. Outdated Evaluation Setup (Model Architecture, Baselines, Benchmarks)

The proposed pretraining framework is evaluated in an outdated evaluation scheme that does not follow recent literature. Specifically:

  • Model architecture: The proposed model and all baselines are evaluated using the InceptionTime architecture. This architecture does not model temporal interactions but instead uses global average pooling over patch embeddings. This is an extremely restricted setup and ignores modern architectures that include temporal interactions (CNNs and transformers). It is therefore unclear whether the proposed pretraining framework provides any benefits when using an architecture with temporal interactions at inference.
  • Baselines: The authors do not compare against any modern time-series foundation models and also lack common simple time-series baselines. Missing baselines include:
    • UniTS (Gao et al., 2024)
    • MOMENT (Goswami et al., 2024)
    • OTIS (Turgut et al., 2026)
    • TimesNet (Wu et al., 2023)
    • Ti-MAE (Li et al., 2023)
    • PatchTST (Nie et al., 2023)
    • CroSSL (Deldari et al., 2024)
    • TSLANet (Eldele et al., 2024)
    • InvConvNet (Germain et al., 2025)
  • Benchmarks:
    • The authors only consider the 28 UEA multivariate archive (Bagnall et al., 2018) datasets while missing the commonly used 121 UCR (Dau et al., 2019) archive datasets
    • The used ECG dataset only contains 5 subjects for testing. Authors should instead use larger datasets for benchmarking, like PTB-XL (Wagner et al., 2020) or MIMIC-IV-ECG (Gow et al., 2023)
    • Most other benchmarks also contain only a few subjects (all less than 30). Authors should consider including benchmarks with more subjects, where available.

Overall, the evaluation setup is very limited and outdated, especially using an architecture without temporal interactions and lacking modern baselines, while using only benchmarks with few subjects. Therefore, the work cannot be judged in comparison to existing literature.

2. Limited Contribution and Motivation

The major contribution of the proposed work is the context-prediction task for pretraining. While the pretraining objective seems simple and effective within the authors’ evaluation setup, it relies on the properties of the InceptionTime architecture, which does not have temporal interactions. It is unclear how the method could generalize to models with temporal interactions. Therefore, the contribution may only be relevant within this extremely restricted evaluation scenario (without temporal interactions) and it is unclear why this scenario would be relevant.

While channel-independent pretraining seems interesting, the benefits are also not clearly motivated. Importantly, no channel interactions are possible during inference in this scenario. Furthermore, transformer-based time-series foundation models like OTIS (Turgut et al., 2026) already enable processing arbitrary channel dimensions, even without retraining the full model.

Overall, it is unclear what the proposed method contributes to the field, especially considering the large number of time-series foundation models and pretraining methods developed in the recent years. Authors would need to clearly motivate their work and evaluation setup considering recent progress and literature.

Questions for authors
  • Why focus on the InceptionTime architecture?
4 comments
AuthorsOpenReview ↗

Official Response by the Authors to Reviewer Z43V

We appreciate the time you took to review our work and are glad you found the method compute-efficient and easy to deploy.


Model architecture

It is important to clarify that InceptionTime is a 1D convolutional network with multi-scale kernels (up to length 40) and stacked residual blocks. It models temporal interactions through its receptive field, and only the final aggregation is global average pooling. The same holds for the original dilated-CNN encoder of TS2Vec and the simple 1D CNN of CaTT.

We chose a single common backbone deliberately, so that differences reflect the pretext objective rather than the architecture. Table 3 also evaluates FCN and ResNet encoders.


Baselines

Many of the listed methods are backbone architectures that focus on encoder design rather than on self-supervised learning (TimesNet, TSLANet, InvConvNet, OTIS), forecasting models (PatchTST), or foundation models pretrained on large external corpora (UniTS, MOMENT). These address a different setting from ours, which is a self-supervised pretext task on unlabeled data under a fixed compute budget.


Benchmarks

  • UCR. Our paper targets multivariate classification, and the multivariate UEA archive is the standard benchmark for this task. The UCR archive is univariate and, by design, does not require the channel-independence strategy that we aim to highlight in our experiments.
  • ECG. We followed the TNC and CaTT literature, which also use this dataset. We agree that larger-scale ECG benchmarks are valuable. However, the test set contains 16,645 samples, which is already substantial compared with the datasets in the UEA archive. Note also that ECG is only 1 of the 34 diverse datasets (6 large-scale + 28 UEA) in our experiments.

Channel independence and contribution

We agree that channel independence (CI) cannot model cross-channel interactions at inference. This is consistent with our UEA results, where CI underperforms on datasets in which cross-variable dependencies matter, such as StandWalkJump.

As noted in our response to Reviewer BLfS, despite the lower average accuracy of the CI variant on UEA (70.67% vs. 72.05%), it achieves the best result on more individual datasets than channel mixing (8 wins vs. 4 wins). This suggests that explicit channel interactions are not needed to obtain good performance in several domains.

Our contribution is a lightweight, augmentation-free SSL objective for practitioners who pretrain directly on their own unlabeled data with limited compute. This complements, rather than competes with, foundation models.

Thank you for the detailed comments, and we hope our responses fully address your concerns.

Thank you for your response. However, this does not resolve my concerns.

Model Architecture

CNN-based models are more and more being replaced by transformer-based models for time-series, as CNNs cannot capture long-range dependencies well. Considering that, I think transformer-based architectures need to be considered as well. With that in mind: does your representation learning approach also work with transformers and how well? Also, can it handle longer-range dependencies? Because anything longer than the patch-size is just handled via global-average pooling and in my opinion your proposed method relies in global-average pooling and may not work with other forms of long-range interactions.

Baselines

Even if these focus on architectural aspects, it would be interesting to compare their benchmark results with yours. These benchmark scores would set your results in context. While your method does not need to outperform those scores (because those foundation models may have been trained on way more data), I think it is important to see how far your experimental results are from SOTA results on the used benchmarks

Benchmarks

  • UCR: I understand your argument for exclusion.
  • ECG: if you want to demonstrate performance on ECG, you need reliable benchmarks. Testing on only 5 subjects is not reliable as they will not cover different patient distributions (e.g., age, sex, phenotypes/diseases) well. Even if it contains 16,645 patches/timepoints, these are highly correlated as they stem from the same few patients. Considering that there are larger datasets available for ECG, I do not understand why you chose this small benchmark.
AuthorsOpenReview ↗

Official Response by the Authors to Reviewer Z43V

Thank you for the prompt response and continued interest in our work.

Model architecture

We agree that transformers can model long-range dependencies and have proven effective in time series foundation models. In self-supervised representation learning for time series, however, the backbone is typically trained from scratch on each dataset, and transformers generally need large amounts of training data. As a result, they have seen little adoption in this setting.

Does your representation learning approach also work with transformers, and how well?

We did not adapt our approach to transformer backbones. We follow prior time series representation learning work (TNC, TS2Vec, TF-C, TS-TCC, SoftCLT, CaTT), which uses CNN encoders. TS2Vec's encoder ablation (Table 5) specifically shows a transformer backbone underperforming a CNN on the 128 UCR datasets.

Baselines

We agree that comparing with foundation models puts our results in context. Following your suggestion, we evaluated the open-source MOMENT-1-base model on the six large-scale datasets, using frozen embeddings and the same linear evaluation protocol as our other baselines. MOMENT-1-base model uses channel independence (CI) encoding, so we added C-Pred (CI) variant for comparison. We could not set up a reliable UniTS evaluation within the response period because its prompt tokens are dataset-specific.

ModelECGHARTHPAMAP2SKODASLEEPWISDM2Avg. Acc.
*MOMENT-1-base (CI)78.1193.3573.8299.2476.1468.0581.45
C-Pred76.1593.4370.5499.0885.1563.8781.37
C-Pred (CI)79.8894.4580.9599.6285.1666.0984.36

Again, note that C-Pred and MOMENT represent different pretraining paradigms. C-Pred is a dataset-specific self-supervised learning (SSL) method, pretrained independently on the training split of each benchmark dataset without labels. In contrast, MOMENT is a time-series foundation model (FM), pretrained on a large and diverse corpus of time-series data.

*Edit: Improved and updated MOMENT results by resizing all input sequence lengths to 512 as recommended by the authors.

Benchmarks

Thank you for suggesting more diverse ECG datasets such as PTB-XL and MIMIC-IV-ECG. We will consider them in future work.

Thank you for your answer.

The additional foundation model results provide useful context.

However, the model architecture remains a critical issue. The field is moving to transformer-based architectures, even for smaller-scale scenarios. Therefore, the value of a new pretraining method that solely focuses on CNN-based models (and may not even be extendable to transformer-based models) seems unclear to me. Please explain the motivation why you focus only on CNN-based models and ignore transformer-based ones.

Regarding the ECG benchmarks: I understand that it may not be feasible to study your method on larger ECG benchmarks within the rebuttal period. However, it remains unclear why these have not been included so far as they are publicly available and commonly used datasets. And I repeat that testing on only 5 subjects CANNOT provide a robust analysis of ECG performance.

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

The paper presents a method to learn representations for multivariate time series, by doing next-patch context prediction (C-Pred). Given a reference patch, the model needs to identify immediate temporal neighbor from K candidate patches. The authors claim that compared to prior methods like contrastive learning which needs curation of positive and negative pairs, and masked modeling which relies on reconstruction, C-Pred is lightweight self supervised learning, since it avoids augmentation and cross instance negatives and reconstruction, and that it is cheaper because the encoder sees only K + 1 short patches per step. The number of candidates K control the difficulty of the task. The patch is encoded by Inception Time encoder, with MLP head to score pairs. A channel independent variant encodes each channel separately. Experiments are done on 6 larger benchmarks, and 28 UEA multivariate datasets, and C-Pred performs competitively for lesser training time compared to baselines. The method is not compared with other representation learning methods.

Critical strengths
  1. The adaptation of context prediction pretext task in computer vision where spatial relationships are used as a signal for representation learning is interesting. Reformulating this as K-way classification with one unique correct answer is a reasonable application.
  2. The pretext task itself is simple and easy to to reproduce, without needing augmentation or creation of positive negative pairs
  3. Experimental results are promising when compared with contrastive learning methods
Critical weaknesses
  1. Since the patches are continuous and do not overlap, the closest temporal match would be a patch which starts exactly where the current patch ends, and make the task quite simple. In computer vision, this is avoided by adding gaps and jitter but this is missing from the work, and the model could learn to shortcut the task due to this continuity. The claim that K would increase difficulty thus may not be the case and is not supported
  2. Since the main contribution is a representation learning framework, masked modeling is only represented by SimMTM, and the comparison is weak. There is no comparison with pretrained time series foundation models on classification.
  3. Channel independence is not new, and the contribution of CI is overstated. On UEA, CI is worse than channel mixing and no baseline outside of TNC was run on it
  4. Next patch prediction and formualting it as K way classification is quite close conceptually to contrastive learning. Minor: The greyed out values in Table 1 and 2 are hard to read, and could be formatted better
Questions for authors
  1. Is there a relationship between the K that is picked and the variance in the data? It is treated as a hyperparameter, but the hardness of the task is probably related to how smooth the trajectories are/etc. Do the authors have results on pretext-task accuracy across values of K and across datasets?
  2. What is the positioning of the work against close prior work on temporal pretext tasks?
  3. Could the authors comment on the task being easily solvable by just boundary continuity? (i.e. end of one patch would be exactly same as beginning of other task)
2 comments
AuthorsOpenReview ↗

Official Response by the Authors to Reviewer BLfS

We are grateful for the constructive feedback and pleased that you found the KK-way reformulation reasonable and the method simple and reproducible. We hope the following responses fully address your concerns.


W1 / Q1 / Q3: Boundary continuity shortcut

Thank you for this important point. We agree that adjacent non-overlapping patches could, in principle, admit a boundary-matching shortcut. It is important to state, however, that:

  • Each patch is mapped by the encoder to a single dd-dimensional vector (via global average pooling) before any comparison. The model never directly compares the last time step of the reference with the first time step of a candidate. For a boundary shortcut to work, the boundary information would have to survive pooling and projection into the dd-dimensional space Rd\mathbb{R}^d.
  • A pure boundary-matching shortcut is symmetric, so it would work equally well for previous-patch and next-patch identification. Table 3 shows a consistent preference for next-patch prediction (e.g., 72.05% vs. 70.70% on UEA), which suggests the encoder learns direction-dependent temporal structure beyond edge matching.

That said, we do not claim that this rules out shortcuts. To investigate this empirically, we conducted an additional boundary-discontinuity ablation in which we remove a fixed number of timesteps from the end of every patch. We consider gaps of 1, 2, and 5% of the original patch length (rounded up to at least one timestep).

ModelECGHARTHPAMAP2SKODASLEEPWISDM2Avg.
C-Pred (default)76.15 ± 2.7393.43 ± 1.3770.54 ± 0.5599.08 ± 0.1285.15 ± 0.1763.87 ± 4.2181.37
C-Pred (gap=1)75.88 ± 1.3193.64 ± 1.3570.15 ± 0.4598.93 ± 0.1785.15 ± 0.2061.69 ± 1.2980.91
C-Pred (gap=2)76.49 ± 2.0992.56 ± 1.8870.49 ± 1.1599.02 ± 0.1285.16 ± 0.2063.10 ± 3.2081.14
C-Pred (gap=5%)77.30 ± 1.7992.54 ± 3.4470.15 ± 0.4598.93 ± 0.1785.17 ± 0.1961.69 ± 1.2980.96

The results show no consistent or substantial degradation across the evaluated datasets when boundary continuity is disrupted. This suggests that C-Pred does not rely strongly on immediate boundary continuity to identify the temporally neighbouring patch.


W2: Comparison with pretrained time series foundation models

Our study targets task-specific SSL on the target dataset under a unified backbone and protocol, which is why all baselines share the same InceptionTime encoder. Foundation models are pretrained on large external corpora and therefore answer a different question. We also see our method as complementary to foundation models: C-Pred can be explored as a fast and effective pretext task in the design of time series foundation models.


W3: Channel independence

We agree that channel independence (CI) is not new in time-series modelling. As stated in the introduction, CI is not our contribution. Rather, our observation is that the lightweight patch-level pretext task in C-Pred makes CI practical in self-supervised learning.

To the best of our knowledge, this is the first comprehensive empirical study of channel independence and channel mixing for time-series classification within this SSL setting.

We also acknowledge that CI achieves lower average accuracy on UEA than channel mixing (70.67% vs. 72.05%), as reported in the paper. However, despite the lower average, the CI variant achieves the best result on more individual datasets than channel mixing (8 wins vs. 4 wins), which suggests that some datasets benefit from this style of encoding.


W4 / Q2: Relation to contrastive learning and prior temporal pretext tasks

We agree that KK-way softmax cross-entropy is related to InfoNCE. C-Pred can be viewed as a within-instance, relative-position variant that uses a learned scoring head rather than a similarity function. Unlike TNC and CPC (Appendix D), it requires no neighbourhood-width estimation, PU correction, context aggregator, or horizon-specific parameters.

Minor: We will reformat the gray values in Tables 1 and 2 to improve readability.

Once again, thank you for the thoughtful comments and questions.

Thank you for your response.

  • Regarding boundary continuity - my concern is about continuity in general. Properties like trend, etc survive pooling so that does not necessarily prevent the shortcut. The gaps of 1/2 are too small to prove the point that the shortcut will not be used. A more meaningful experiment would be to do these experiments over larger gaps.
  • The gap between next and previous patches (72.05% vs. 70.70% on UEA) shows that they are not symmetric, but it does not prove that it will not be used as a shortcut
  • Regarding baselines, I understand the difference between task specific SSL and foundation models. However since the pipeline is motivated by low cost, comparison with pretrained models is a reasonable comparison. The response also does not address the concern on SimMTM being the only baseline for masked modeling
  • Thank you for clarifying that CI is not claimed as a contribution. However the claim on the first to study CI and channel mixing in SSL setting is not supported when no other baseline other than one other is run with it
  • Q1 is not addressed

Overall, the pretext task is simple and the results are reasonable, but my main concerns about the boundary shortcut, task difficulty, and baselines remain.

Reviewer_7q6WOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes C-Pred, a lightweight self-supervised framework for multivariate time-series classification based on next-patch context prediction. Given a reference patch, the model identifies its immediate temporal neighbour from K candidate patches within the same instance using a K-way classification objective. The framework avoids data augmentation, cross-instance negative sampling, and reconstruction, and is evaluated on six large-scale datasets and 28 UEA datasets with both channel-mixing and channel-independent encoding.

Critical strengths
  1. The paper studies a practically relevant problem: reducing the computational and design overhead of self-supervised learning for time series while retaining competitive representation quality.
  2. The experimental evaluation is relatively broad, covering six large-scale datasets and 28 UEA datasets, with a unified InceptionTime backbone for the compared methods and explicit reporting of training time.
  3. The analysis examines several design choices, including the predictive versus contrastive objective, K, patch count, prediction head, encoder architecture, batch size, and channel encoding strategy, which provides useful empirical evidence about the behavior of the proposed framework.
Critical weaknesses
  1. The methodological novelty is limited relative to prior temporal-context SSL methods. C-Pred shares the core premise of CPC and TNC that temporal proximity provides self-supervision, and mainly reformulates this signal as K-way next-patch identification. The paper does not provide a controlled comparison against CPC-style predictive objectives or other matched temporal-context formulations to establish that the proposed reformulation itself provides a substantive methodological advantage.
  2. The reported benefit of channel-independent encoding is confounded by a substantially larger downstream representation. The standard C-Pred representation has dimension (d), whereas the CI variant concatenates per-channel representations into a (Cd)-dimensional vector; for example, (C=52) and (d=256) yields 13,312 dimensions. Therefore, the reported 2.99-point improvement cannot be cleanly attributed to channel independence without a dimension-controlled comparison.
  3. The hyperparameter-selection protocol is insufficiently specified. The paper adopts (K=2) for the large-scale datasets and (K=5) for UEA based on their observed performance, and similarly selects the prediction head, patch count, and encoder architecture through empirical comparisons. However, it is unclear whether these choices were made using a separate validation set or the reported evaluation results, making the fairness of model selection difficult to assess.
  4. The efficiency comparison does not fully control for training compute. C-Pred processes only ((K+1)T/P) time steps per update—(0.3T) under the default (P=10,K=2)—whereas several baselines process full sequences or multiple views. The reported wall-clock speedups are therefore partly induced by substantially less encoder input per optimization step. Matched-FLOP, matched-token, or convergence-based comparisons would be needed to support a stronger claim of intrinsic computational efficiency.
Questions for authors

Please refer to Weaknesses.

1 comment
AuthorsOpenReview ↗

Official Response by the Authors to Reviewer 7q6W

W1: Novelty relative to CPC/TNC

Thank you for raising this important point. We agree that C-Pred shares with CPC and TNC the premise that temporal proximity provides self-supervision (Appendix D). Our contribution is not this premise, but how it is turned into a pretext task that is:

  • Strictly intra-instance, so instance-level cues (amplitude, drift, subject identity) carry no discriminative information
  • Patch-level, so the encoder sees only (K+1)T/P(K + 1)T /P time steps per update
  • Free of a context aggregator, horizon-specific parameters, neighbourhood-width estimation (ADF in TNC).

Table 3 already isolates the objective formulation: the contrastive variant uses the same patch pairs with NT-Xent. We acknowledge that the contrastive variant performs better on some large-scale datasets, whereas K-way classification performs better on UEA (72.05% vs. 69.56%). The efficiency gain comes from the patch-level, intra-instance design shared by both variants.


W2: Channel independence is confounded by representation dimension

This is a fair concern. The existing results provide some initial evidence:

  • The randomly initialised CI baseline (Table 1) also produces a CDCD-dimensional representation, but improves over random channel mixing by only 0.52 percentage points (79.39% vs. 78.87%).
  • Pretraining improves accuracy by 4.97 points under CI (84.36% vs. 79.39%), compared with 2.50 points under channel mixing (81.37% vs. 78.87%).

However, we agree that without controlling representation dimensionality, the reported improvement cannot be attributed solely to channel independence. We therefore conducted an additional dimension-controlled ablation with three configurations:

  • C-Pred (mixed-DD): The standard mixed channel-processing configuration, with output dimension DD.
  • C-Pred (mixed-CDCD): The same mixed channel-processing formulation, but with output dimension increased to CDCD to match the CI representation.
  • C-Pred (CI-CDCD): The channel-independent configuration used in the main experiments, also with output dimension CDCD.

The results are shown below. We will include this table in the revised version of our paper.

MethodECGHARTHPAMAP2SKODASLEEPWISDM2Time (h)Avg. Acc.
C-Pred (mixed-D)76.15 ± 2.7393.43 ± 1.3770.54 ± 0.5599.08 ± 0.1285.15 ± 0.1763.87 ± 4.210.1981.37
C-Pred (mixed-CD)77.01 ± 5.7693.28 ± 1.6474.48 ± 2.5699.53 ± 0.0885.39 ± 0.2264.20 ± 1.0915.4482.32
C-Pred (CI-CD)79.88 ± 2.6394.45 ± 0.7880.95 ± 0.5099.62 ± 0.0685.16 ± 0.3166.09 ± 1.6910.7684.36

Classification accuracy (%) is averaged over five evaluation seeds. Best and second-best results are shown in bold and underlined, respectively.

Matching the representation dimension improves mixed training from 81.37% to 82.32%. However, CI-CDCD achieves 84.36%, a further 2.04-point improvement over mixed-CDCD. This suggests that the CI improvement is not explained by representation dimensionality alone.


W3: Hyperparameter selection protocol

Thank you for pointing this out. For UEA, we follow the TS2Vec evaluation protocol (fixed seed, no validation split). The hyperparameters were fixed rather than selected using a separate validation split. We retain K=2K=2 for PAMAP despite the sensitivity analysis showing better results with K=5K=5. For the large-scale datasets, we chose K=2K=2 because of the added advantage of reduced computation. Baselines use their official default hyperparameters.

Importantly, our conclusions do not depend on the UEA choice of KK. Using the large-scale default, K=2K=2, C-Pred still achieves 70.58% on UEA, outperforming the best baseline, TNC (70.38%).


W4: Efficiency and matched compute

We agree that part of the speedup comes from processing less input per training step; this is intentional. C-Pred trains the encoder using only the reference patch and its K candidate patches, reducing the amount of input processed while maintaining competitive or better performance.

All methods use the same backbone, 1,500 training iterations, and the same batch size, making the comparison iteration-matched. Full-sequence objectives process the entire sequence at each step, whereas C-Pred's patch-level objective enables reduced per-step computation. We therefore view the efficiency gain as a consequence of the proposed formulation, rather than as a compute-matched advantage independent of input size.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.