Catching the Imposter: Self-Supervised Learning of Physical Coherence with Cross-Entity Feature Permutations

Aleksei Rozanov; Arvind Renganathan; Vipin Kumar

Source-linked AI summary

Catching the Imposter: Self-Supervised Learning of Physical Coherence with Cross-Entity Feature Permutations

Aleksei Rozanov, Arvind Renganathan, Vipin Kumar

arXiv:2608.14372v1Aug 14, 2026cs.LG

TL;DR

Scientific representation learning often overlooks the physical coherence linking an entity’s features. This paper introduces imposter, which detects feature swaps between real entities, and finds that objective choice depends on the downstream task while imposter adds complementary information.

Problem
Scientific data contain physically coupled features whose cross-feature coherence carries information, but common self-supervised objectives primarily exploit spatial and temporal structure.
Method
Imposter replaces subsets of an entity’s features with real observations from another entity and trains an encoder to identify the physically incoherent features.
Results
No self-supervised objective consistently dominates across downstream applications; each objective has distinct strengths associated with its inductive bias.
Takeaways & Limitations
Selecting or combining self-supervised objectives should be guided by intended downstream applications, with physical coherence providing complementary information.
Takeaways & Limitations
The comparison is limited to baselines using the same encoder architecture, optimizer settings, and pre-training schedule as imposter.

Abstract

from arXiv · show

Scientific data often describe entities whose features are jointly governed by the laws of physics, yet existing self-supervised learning (SSL) objectives largely ignore this physical coherence. We introduce imposter, a discriminative pretext task that replaces subsets of an entity's features with real observations donated by another entity and trains the encoder to identify the swapped features. Because every donated value is individually plausible, the task can only be solved by learning cross-feature physical dependencies. We evaluate the proposed objectives on global ERA5-Land reanalysis data using 21 environmental variables and assess the learned representations on seven downstream tasks spanning climate classification, carbon flux estimation, and streamflow prediction. Our study includes, to our knowledge, the first systematic comparison of self-supervised objectives for land-surface modeling under a shared architecture and pre-training budget. We find that the most effective pretext task depends on the downstream task family rather than any single objective's superiority, and that imposter provides complementary information when combined with existing SSL objectives. These results suggest that physical coherence is a valuable new source of self-supervision for scientific foundation models.

1 Introduction

Scientific observations contain physically coupled features, but existing self-supervised objectives largely exploit spatial and temporal structure while overlooking this coherence. The paper introduces imposter detection, evaluates its complementarity with established objectives on global ERA5-Land data, and finds that effective pretext tasks depend on downstream task family.

Motivation: Scientific entities have features jointly governed by physical laws, and this coherence carries substantial information in multivariate observations.Examples include temperature, radiation, and water variables in ecosystems, as well as related clinical and engine measurements.
Motivation: Existing self-supervised methods leverage spatial and temporal regularity but remain largely agnostic to cross-feature physical coherence.This gap distinguishes scientific systems from common language and vision settings.
Method: Imposter detection randomly replaces a subset of one entity’s features with corresponding values donated by another entity and trains per-feature binary classification to identify incoherent inputs.Because donated values are individually plausible, detection requires modeling dependencies among the receiving entity’s features.
Study design: The study hypothesizes that imposter’s cross-feature coherence signal complements masking, forecasting, and contrastive objectives that primarily exploit spatial and temporal structure.The paper examines combinations through concatenated embeddings and joint training with five widely used SSL objectives.
Study design: The evaluation pre-trains encoders on 21 ERA5-Land features and assesses representations across seven downstream tasks covering climate classification, carbon flux estimation, and streamflow-related prediction.The tasks include Köppen classification, GPP, RECO, NEE, streamflow, and basin identity prediction.
Findings: The choice of pretext task depends on the downstream task family rather than a universally superior objective, while imposter provides complementary information alongside existing objectives.The objectives are benchmarked under a shared architecture and pre-training budget.

2 Related Works

Related SSL work centers on masking and reconstruction, contrastive discrimination, and other pretext tasks, while corruption-based discrimination offers a distinct alternative. Despite progress in vision and language, scientific applications remain limited because scientific data pose distinct challenges.

Masking and reconstruction: Masked autoencoding masks inputs and reconstructs missing tokens or regions from context, becoming a dominant SSL paradigm across modalities and domains.BERT and related methods apply masking to NLP, while vision MAEs build on earlier image-inpainting formulations.
Contrastive learning: Contrastive learning learns representations by discriminating between similar and dissimilar pairs rather than reconstructing corrupted inputs.SimCLR showed that augmented views of the same image can serve as positive pairs within a batch for downstream representation learning.
Other pretext tasks and the AI4Science gap: Other pretext tasks include jigsaw puzzles, future prediction, denoising, and adversarial generation, but their adoption in scientific domains remains limited.These methods have driven progress in vision and language, while scientific data pose distinct challenges.
Corruption-based discrimination: Corruption-based discrimination replaces input elements with plausible alternatives and trains an encoder to detect the replacements, forming a discriminative alternative to generative objectives.ELECTRA substitutes natural-language tokens with outputs from a small generator network and trains the main encoder as a classifier.

3 Methodology

The methodology defines temporal entities and trains an encoder through imposter, which creates physically incoherent samples by donor-swapping features and detects the swapped features. Randomized swap scheduling and multi-imposter address shortcuts, while distinguishing imposter from reconstruction-based SSL by emphasizing cross-feature dependencies.

Problem setup: The dataset contains unlabeled temporal entities observed across T time steps and F features, with an encoder trained to produce representations for downstream tasks under limited supervision.Each entity is represented as xi ∈ R^(T×F).
Imposter pretext task: Imposter replaces feature subsets with corresponding features from randomly assigned donor entities and trains per-feature binary classification to identify the replacements.Donated values remain individually plausible, so detection requires recognizing their incoherence with the host entity’s remaining features.
Imposter pretext task: The discriminator implicitly learns joint-distribution differences between coherent entities and donor-swapped entities, providing supervision for p(x1, x2, . . . , xn) > p(˜x1, x2, . . . , xn).The swapped feature ˜xi is drawn from another entity.
Implementation: The swap ratio r controls corruption, with the default swap count sampled uniformly from [1, r·F] per sample; this schedule outperforms a fixed rate and prevents constant-rate shortcuts.Random scheduling is identified as the default strategy.
Implementation: Multi-imposter draws each swapped feature from an independent donor to prevent donor memorization at high swap ratios and force feature-wise coherence evaluation with the host.The method decouples corrupted features from one another.
Relation to existing SSL: Unlike MAE, which removes information and reconstructs masked inputs, imposter substitutes realistic alternatives and discriminatively detects them to model cross-feature dependencies.This shifts the encoder’s objective from preserving information for reconstruction to distinguishing coherent from incoherent configurations.

4 Experimental Settings

The experiments isolate pre-training objective effects with a shared patch-transformer setup on globally consistent ERA5-Land data. Representations are assessed through frozen-encoder downstream tasks and complementary embedding-geometry metrics.

Model and data: All methods use the same patch-transformer encoder, optimizer settings, and pre-training schedule to isolate the effect of the pre-training objective.Inputs are divided into temporal patches and encoded with positional embeddings.
Model and data: The study uses 21 co-evolving variables from globally consistent daily ERA5-Land fields at 0.1° resolution, treating each grid cell as a separate entity.The variables span near-surface atmospheric and soil components of the water and energy cycles across 2001–2024.
Model and data: A 30-day temporal window with patch length 2 is fixed across all pre-training and downstream experiments as an empirical balance between compute cost and task coverage.Latitude, longitude, and sine–cosine day-of-year features provide context and are not corrupted or counted among the 21 imposter features.
Pretext tasks: Imposter is compared with five corruption-based, forecasting, or contrastive pretext tasks, including NTXent, NTP, Feat-MAE, and Temp-MAE.Feat-MAE masks 75% of features across all time steps and reconstructs them with MSE, making it the direct generative counterpart to imposter.
Evaluation: Representations are evaluated with frozen-encoder supervised tasks covering Köppen-Geiger classification, CAMELS streamflow and basin identity, and CarbonFluxBench GPP, RECO, and NEE prediction.Reported metrics include balanced accuracy, site-level median R2, RMSE, and site-level median R2 per carbon-flux target.
Evaluation: Unsupervised representation quality is characterized with RankMe and αReQ, computed on a fixed held-out set unseen during training.RankMe measures effective rank, while αReQ reports the spectral-decay slope α.

5 Results

Results show that no self-supervised objective is universally best: downstream performance depends on alignment between an objective’s inductive bias and task structure. Imposter provides complementary representations, especially through concatenation, at computational cost comparable to reconstruction and forecasting objectives.

Individual objectives: No single objective dominates across downstream tasks; each performs best when its learned structural information aligns with the task.The results indicate complementary rather than universally superior representations.
Individual objectives: FT-MAE achieves the best streamflow result (R2 = 0.384), followed by Feat-MAE (R2 = 0.378), while Multi-Imposter achieves the highest single-objective NEE score (0.223).Feat-MAE achieves the best GPP performance (0.499), and NTP achieves the best RECO performance (0.447).
Representation geometry: Imposter produces the highest effective rank (RankMe = 80.4) and lowest αReQ value (1.442), but richer embedding geometry does not ensure uniformly better downstream performance.Both imposter variants perform considerably worse on basin attribute prediction, where physical coherence does not capture all stable entity-level information emphasized by contrastive learning.
Concatenated Embeddings: Concatenating imposter representations yields gains up to +18.5% on LP-30 with FT-MAE and +54.8% on NEE with Contrastive.Multi-Imposter generally improves baselines more than Imposter, while basin attribute concatenation increases RMSE by 13.5–27.0%.
Joint Training: Joint optimization produces smaller average gains than representation concatenation, indicating that independently trained encoders preserve more complementary information.Nevertheless, Multi-Imposter combined with Contrastive reduces basin RMSE by 1.8% relative to Contrastive alone and achieves the best basin attribute score.
Computational cost: End-to-end, Imposter, NTP, and MAE variants fall within 3% of one another in total FLOPs, whereas Contrastive requires approximately 2.05× the total FLOPs.Imposter therefore provides complementary representations at a cost comparable to reconstruction and forecasting methods and substantially below Contrastive.

6 Discussion

The discussion concludes that no self-supervised objective consistently dominates: reconstruction, contrastive learning, and imposter each target different structural properties and downstream needs. Imposter adds complementary physical-coherence information, and combining it with existing objectives consistently improves performance.

Objective-specific inductive biases: No self-supervised objective consistently dominates; reconstruction favors temporal information, contrastive learning entity discrimination, and imposter cross-feature physical coherence.The discussion recommends choosing pretext tasks according to downstream-task requirements rather than seeking one universally superior objective.
Imposter as complementary supervision: Multi-imposter is the second most consistent standalone method across downstream tasks, with compute comparable to NTP and substantially lower than Contrastive.Its central contribution is introducing physical coherence as a complementary source of self-supervision.
Imposter as complementary supervision: Combining imposter with existing objectives consistently improves performance, indicating that physical coherence provides information largely absent from current self-supervised objectives.This supports using imposter alongside, rather than only instead of, existing objectives.
A broader perspective: Self-supervised objectives should exploit different structural properties of scientific data rather than compete to identify a single best objective.Reconstruction emphasizes temporal and spatial regularities, contrastive learning instance discrimination, and imposter cross-feature physical coherence.

7 Conclusion · A Swap Scheduling

Imposter uses physically plausible feature swaps to learn cross-feature dependencies in scientific time series. Across seven downstream tasks, it is competitive and complementary to other objectives, while scheduling experiments compare four swap-ratio strategies.

7 Conclusion: Imposter and multi-imposter replace subsets of an entity’s features with plausible observations from other entities to identify incoherent features.This objective exploits physical coherence in multivariate scientific time series.
7 Conclusion: The proposed task encourages representations that capture cross-feature dependencies governing scientific systems.
7 Conclusion: Across ERA5-Land pre-training and seven downstream tasks, imposter demonstrates competitive standalone performance while providing complementary information to existing self-supervised objectives.The tasks span climate classification, hydrology, carbon flux estimation, and representation analysis.
7 Conclusion: No single pretext task is universally superior because objectives capture different data structures, inductive biases, and downstream strengths.
7 Conclusion: Selecting or combining self-supervised objectives should be guided by intended downstream applications rather than a single benchmark.
7 Conclusion: Scientific structure such as physical coherence, conservation laws, symmetries, and causal relationships can provide sources of self-supervision beyond objectives adapted from language and vision.
A Swap Scheduling: Table 4 compares Constant, Linear, Random, and Random+Linear scheduling strategies across swap ratios.The experiments evaluate different swap schedules and swap ratios.
A Swap Scheduling: Random+Linear samples each instance’s swap ratio uniformly between zero and a linearly increasing upper bound, rmax,t = rmax · t/T.Here, t is the current epoch, T the total number of epochs, and rmax the run’s maximum swap ratio.

B Model and Training Configuration

All pretext objectives use the same patch-transformer backbone and shared optimization settings for a fair comparison. Only the objective-specific decoder varies between methods.

Shared architecture: All pretext objectives use an identical patch-transformer backbone.This shared backbone supports a fair comparison across methods.
Shared architecture: Only the objective-specific decoder differs between methods.The backbone remains fixed while the decoder changes with the pretext objective.
Optimization: Architecture and optimization settings are shared across all runs.Table 7 lists these common architecture and optimization hyperparameters.

C Dataset

The pre-training dataset uses 21 ERA5-Land variables covering the principal components of the land-surface energy and water cycles, organized by physical process group.

Variables: 21 variables span the principal components of the land-surface energy and water cycles.The variables are listed in Table 8.
Variables: The ERA5-Land variables are grouped by physical process.Table 8 presents the grouping used for pre-training.

D Concatenated embeddings

Concatenating imposter representations with baseline representations produced near-universal improvements across all evaluated tasks. Table 5 reports this result for each baseline method.

D Concatenated embeddings: Table 5 evaluates concatenated representations combining an imposter encoder with each baseline method.The evaluated representation pairs an imposter encoder with each baseline method.
D Concatenated embeddings: Near-universal improvements were observed across all tasks.The passage characterizes the improvements as near-universal rather than limited to selected tasks.
D Concatenated embeddings: The results indicate that imposter embeddings add broadly useful information when concatenated with baseline embeddings.This summarizes the reported task-wide improvements from combining imposter and baseline representations.

E Joint Training · F Computational Efficiency

Jointly training both objectives from scratch yields smaller average gains than concatenating their representations, indicating that co-optimization loses part of the complementary signal. The computational-efficiency analysis separates objective-specific decoder overhead from the shared encoder’s full forward-pass cost.

E Joint Training: Joint training from scratch produces smaller average gains than representation concatenation.The paper attributes this reduction to co-optimization discarding part of the complementary signal.
E Joint Training: Table 6 evaluates encoders trained jointly on both objectives from scratch.The evaluation reports improvements across the paper’s downstream tasks, though no individual values are provided here.
E Joint Training: The joint-training comparison indicates that co-optimization can reduce complementary information relative to concatenated representations.This conclusion is based on the smaller average gains reported for joint training.
F Computational Efficiency: The efficiency analysis separates objective-specific decoder cost from the shared encoder’s full forward pass.This distinction enables comparison of head-level overhead with end-to-end computation.
F Computational Efficiency: Table 9 reports task-specific computational overhead in FLOPs for head/decoder components and full-model end-to-end computation.The full-model measure includes the shared encoder, quantified as 23,784,960 FLOPs.
F Computational Efficiency: Overhead metrics in Table 9 are reported relative to imposter.The table compares task-specific computational costs against the imposter objective as the reference.
E Joint Training: The concatenated-representation evaluation uses underlining for improvements over base methods and bolding for the best performance in each column.An asterisk denotes the best score across all objectives and experiments.
E Joint Training: The reported joint-objective results use underlining for improvements over corresponding base methods and bolding for column-best performance.An asterisk marks the best score achieved across all objectives and experiments.