acceptodds
Under review as a conference paper at ICLR 2027

RePs: Learning Representations from Known Response Laws

Abstract

Existing augmentation-based self-supervised learning uses transformations to create new views, but often leaves their predictable effects on measurable signal properties unused. Our key insight is that when these effects follow known response laws, augmentation can generate not only new views but also new supervision: an existing measurement can be analytically transformed into the expected measurement of each augmented view. We introduce RePs (Response-law Privileged Supervision), which learns measurement-relevant representations by predicting these targets from augmented signals alone, without additional annotation. We test RePs across three distinct response structures and signal domains: temporal scaling in physiological signals, pitch translation in music, and spatial rotation in multichannel audio. RePs outperforms measurement-free baselines in frozen-feature response prediction across all three domains, reducing spatial angular error from to . With only 10% downstream labels, its ECG representation leads four released encoders on four of five clinical tasks. Controlled ablations further show that direct response supervision yields lower nonlinear readout error than separately predicting source measurements and transformations, supporting known response laws as a source of supervision for representation learning.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
representation learning, self-supervised learning, data augmentation, transformation-aware learning, privileged information, time series

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_4BPEOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes RePs, which utilizes the known response patterns to learn and measure the relevant signal representations. The core approach is to simultaneously transform the signals and the existing measurement values, to enhance the supervisory target for view generation. The model only receives the enhanced signals, without receiving the original measurement values or transformation parameters. This supervision can be used alone or combined with contrastive learning. It requires source samples with existing measurements, but does not need to add annotations for the enhanced view. This idea is applied to electrocardiogram time scaling, music pitch translation, and spatial audio rotation. The main contributions include a systematic comparison of the supervisory targets and the usage of the model. In the main experiments of freezing the encoder and re-training the nonlinear reader, directly predicting the transformed measurements is superior to predicting the original measurements and the transformation parameters separately. However, in the music pitch task, reusing the prediction head during training causes the ranking to reverse. This indicates that a supervision target suitable for learning the representational features that can be re-read is not necessarily the most suitable for directly reusing the trained predictor. The validation covers the response prediction of three types of signals and five clinical transfer tasks. RePs outperforms the baseline that does not use measurement supervision in all three response predictions. The spatial angle error is reduced from approximately 69.0° to 20.2°. In the clinical experiment using only 10% downstream training labels, its model achieves higher scores in four out of the five tasks than the average of four public encoders. However, simple supervision using the existing measurements also achieves similar results. Therefore, the experiments fully support the value of measurement supervision, but the additional benefits brought by the response patterns themselves still depend on the specific task.

Critical strengths
  1. It provides valuable empirical insights into the relationship between the supervision target and the model usage method. In the control group using the same source measurement, direct response supervision outperformed separate prediction of the original measurement and transformation parameters in the main nonlinear readout experiments in three domains, and six comparisons passed the Holm correction. However, on NSynth, when re-fitting the reader, RePs outperformed Factorized; when retaining the training prediction head, the ranking reversed. This indicates that "suitable learning of the representational features that can be re-read" and "suitable training of the predictor that can be directly reused" are not the same goal.
  2. The matching control shows the actual transfer value of measurement-assisted supervision. The clinical experiment matched pre-trained data, backbone networks, initialization seeds, and training budgets for each internal method, reducing the interference of resource differences on the results. When using only 10% of downstream training labels, RePs-SSL outperformed the matched SimCLR in five clinical tasks. PTB-XL AUROC increased from 0.662 to 0.836, and Chapman F1 increased from 0.621 to 0.873. This indicates that the benefits are not limited to enhancing the response prediction of the signal, but can also be transferred to downstream clinical tasks.
Critical weaknesses

The core design's independent contribution compared to ordinary measurement supervision has not been fully established. The paper has acknowledged in related work that the synchronous transformation of signals and labels belong to existing supervised enhancement. Therefore, the key is not only whether performance can be improved after adding measurement supervision, but also what additional value the design of responding patterns provides.

Table 19's matched clinical experiments show that merely predicting the original measurement's Identity has achieved a close mean. Its PTB-XL AUROC is 0.8329, and RePs-SSL is 0.8359; its Chapman F1 is 0.8774, which is higher than RePs-SSL's 0.8730. This table reports the paired confidence intervals relative to SimCLR, but does not fully quantify the incremental effect relative to these measurement supervision controls and its uncertainty. In Table 8, RePs does have advantages in ECG and spatial audio response prediction, but NSynth's Static error is 7.491 semitones, which is close to RePs' 7.502 semitones. The benefits have obvious task dependence.

These results do not deny the effectiveness of the method, but the most significant application benefits more strongly prove that "measurement supervision is useful", and are not sufficient to establish the independent value of the "response pattern objective". With the basic operation having a history, the paper has not clearly defined the applicable conditions for its additional benefits, nor has it fully extracted the empirical conclusions that are sufficient to carry the main innovation. This is the decisive consideration for my decision to reject the paper, rather than considering the method to be wrong or lacking ablation studies.

Questions for authors

In the clinical migration process, how much greater is the independent benefit of the response rule target compared to the ordinary measurement supervision?

Based on the results of the existing five random seeds, please provide the paired differences and 95% confidence intervals of RePs-SSL relative to Identity and Static. Can these results support that correctly responding to the target has additional value, rather than merely proving that adding measurement supervision is effective?

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

The paper proposes a representation learning framework that exploits known response laws linking existing measurements (e.g., ECG intervals, musical pitch, spatial direction) to how those measurements change under signal transformations (e.g., temporal dilation, pitch transposition, angular rotation). Given a source signal and its measurement, RePs transforms the measurement according to the applied augmentation and trains an encoder to predict this transformed target from the transformed signal alone. Across three domains (ECG timing, NSynth pitch, STARSS22 spatial audio), RePs improves frozen-feature response prediction over measurement-free baselines and, in a separate large-scale ECG pretraining, achieves strong linear-probe performance on four of five clinical tasks with 10% labels. Ablations disentangle direct response supervision from factorized source/action prediction, showing that direct response targets generally favor refitted nonlinear readouts, while reusing trained predictors can favor factorized/composed supervision in some settings.

Critical strengths
  1. The framework unifies three canonical response structures (scaling, translation, rotation) across disparate signal modalities, suggesting broad applicability wherever a reliable transformation law of measurements is known.
  2. Comprehensive evaluation across three domains with different response structures, using standardized backbones, common training budgets, and unseen action grids for evaluation.
Critical weaknesses
  1. The approach relies on the availability of response laws. Its utility may be constrained in domains where the response under augmentation is only approximately known or varies across conditions.
  2. Comparisons to measurement-free objectives highlight the value of privileged supervision, but they are not strictly apples-to-apples in terms of supervision budget (RePs uses additional labels).
Questions for authors
  1. How sensitive are the results to approximate or imperfect response laws? For ECG, how do remeasured on-view intervals compare with targets from the digital dilation rule?
  2. Can you quantify how often transformed targets fall outside plausible ranges, and whether excluding or clipping such cases changes the results?
  3. For clinical transfer, can the authors better control for pretraining/evaluation overlap across the released encoders?
  4. In Table 8, Static performs almost identically to RePs on NSynth. How much of the gain comes from measurement supervision itself rather than the response law?
Reviewer_tK4COpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Augmentations with known effects on existing measurements (time dilation → interval scaling, pitch shift → +Δ semitones, rotation → rotated direction) turn one measurement into regression targets for each augmented view. An encoder + response head predicts the target from the view alone, optionally with NT-Xent. Evaluated by frozen-feature readout on unseen actions in three domains, 10%-label linear probes against four released ECG encoders, and ablations of direct vs. factorised supervision with Identity/Static/Wrong controls.

Critical strengths

One genuinely useful finding: on NSynth the ranking of direct vs. factorised supervision flips between refitted readers and retained predictors (Table 4), i.e. the right target depends on how the model will be reused. Anonymised code provided.

Critical weaknesses
  1. The headline comparison (Fig. 2, 69.0°→20.2°) pits a measurement-supervised regressor against measurement-free SSL on the measurement task itself; winning is expected and says nothing about response laws. The controls that do test the claim show small or non-significant effects: Static matches RePs on NSynth (7.491 vs 7.502, Table 8), Identity is within 0.012 on ECG, only one of six Wrong-RePs contrasts survives Holm (Table 9), and Identity/Static/RePs-SSL are indistinguishable clinically (Table 19). The evidence supports “measurement supervision helps”, not the stated contribution.
  2. The released-encoder comparison (Table 2) is not controlled: heterogeneous input adapters, a fixed 1,000-update probe with no early stopping across feature widths 128–768, ECGFM-KED at 0.055 AUPRC on CODE-15 (near chance), and HuBERT-ECG pretrained on all five downstream sources. Since the pretraining targets (RR/QRS/QT) are tightly coupled to the downstream rhythm labels, beating SimCLR is unsurprising and does not isolate the mechanism.
Questions for authors
  1. What is the intended takeaway of Fig. 2 beyond “supervision on the target helps the target”? Could the Identity/Static/Wrong controls (Table 8) become the main comparison?
  2. Is there any clinical task where RePs-SSL significantly beats Identity or Static (Table 19)? If not, should the abstract attribute the clinical gain to measurement supervision?
Reviewer_vx5iOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a signal representation learning method called RePs (Response-law Privileged Supervision). Using known response laws, it analytically transforms existing measurements mm into supervision targets for augmented views, r=Ag(m)r=A_g(m), under a transformation gg. The encoder is trained to predict this target solely from the transformed signal v=Tg(x)v=T_g(x), enabling the learning of measurement-relevant signal representations without requiring additional annotations for augmented views. The method can be trained alone or combined with an NT-Xent contrastive loss. The experiments cover three types of signal transformations: ECG temporal scaling, NSynth pitch translation, and STARSS22 spatial rotation. For clinical transfer, RePs-SSL is pretrained on MIMIC-IV-ECG and compared with four released encoders using 10% labels and linear probes. In addition, the authors conduct multiple ablation studies to compare different prediction schemes, examine the differences between refitted Readers and retained original Predictors, and further investigate remeasured targets and synthetic control experiments.

Critical strengths
  1. The method demonstrates a degree of generality and methodological novelty. By leveraging known measurement response laws, it transforms the measurements of source signals into supervision targets for augmented views, enabling the learning of measurement-relevant representations without requiring additional annotations for the augmented views.
  2. The experiments cover three distinct signal domains. The paper conducts experiments on ECG, musical pitch, and spatial audio, and further investigates its transferability to downstream clinical ECG tasks.
  3. The experimental evaluation and ablation analyses are comprehensive and thorough. In addition to comparisons with several existing methods,the paper also examines different supervision targets, prediction paths, and readout architectures.
Critical weaknesses
  1. The consistency between the theoretical response targets and the actual measurements of augmented signals requires further validation. An important assumption of the paper is that, after applying the known response law to the measurements of the source signal, the transformed measurements should correspond to those obtained by re-measuring the augmented signal. However, it remains unclear whether these two are consistent in practice. The authors are encouraged to further clarify the conditions under which this response law holds.
  2. The idea of the paper is quite similar to conventional label-transforming data augmentation. The authors are advised to further explain what unique learning advantages are provided by modeling based on known response laws, compared with conventional label-transforming data augmentation.
  3. In clinical ECG pretraining, unlike the earlier experiments that use standardized MSE for response-value supervision, the authors adopt Huber loss instead of MSE. However, no controlled experiment directly comparing MSE and Huber Loss under the same clinical pretraining conditions is provided. The necessity of this change is therefore not sufficiently explained.
  4. In clinical ECG pretraining, unlike Shared-base RePs, which uses two independent augmented views, the authors use three independent augmented views. However, similar to the issue raised in Weakness 3, no direct comparison between the two-view and three-view configurations is provided, nor is the necessity of this design choice sufficiently explained.
Questions for authors
  1. Could the authors provide a further mathematical explanation of the conditions under which applying a known response law to the measurement of an original signal yields the same result as directly re-measuring the augmented signal? In addition, in Appendix C.4, only a subset of samples can be successfully re-measured. Does this affect the representativeness of the relevant validation results?
  2. The basic idea of the proposed RePs is highly similar to conventional label-transforming data augmentation. What are the key differences between RePs and conventional approaches? What additional advantages does RePs offer?
  3. Could the authors explain the rationale for using Huber loss and a three-view configuration in clinical ECG pretraining? Were controlled experiments conducted to compare MSE with Huber loss and two views with three views?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.