acceptodds
Under review as a conference paper at ICLR 2027

EnRep: Energy-Induced Representations for Classification without Backpropagation

Abstract

Modern classification relies heavily on learned representations, raising the question of whether discriminative features must always emerge through data-driven learning. We introduce EnRep, a framework that constructs energy-induced representations from prescribed energy functionals and extracts spectral features. Unlike learned feature extractors, it does not use backpropagation or iterative learning for feature construction, and the resulting features are coupled with a simple closed-form classifier. Using a single prescribed energy functional, we evaluate EnRep on synthetic spatial pattern data and two image benchmarks spanning handwritten digits and colorectal histology, with comparisons against a raw-input and a trained neural baseline. Across the tested tasks, EnRep outperforms raw-input classification and, on the synthetic benchmarks, markedly exceeds the accuracy of the trained neural baseline, while its closed-form readout avoids iterative optimization once the representation is constructed. These results demonstrate that energy structure can provide informative and computationally efficient representations without learning the feature extractor from data. More broadly, these findings suggest that energy can serve not only as an objective to be optimized, but also as a source of representations for classification.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Energy-based representation, Spectral features, Representation learning, Feature extraction, Image classification

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Li2xOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

this paper proposes a framework that constructs energy-induced representations from prescribed energy functionals and extracts spectral features. results demonstrate that energy structure can provide informative and computationally efficient representations without learning the feature extract or from data.

Critical strengths
  1. results on structured classification task looks good
  2. The paper explores the use of prescribed energy functionals as a source of discriminative representations. Separating feature construction from supervised classifier training is an interesting direction, particularly for settings in which useful domain-specific structure is available.
Critical weaknesses

The main experiments primarily compare EnRep against raw-pixel ridge regression and a shallow CNN. While these comparisons demonstrate that the proposed representation can be useful, they do not establish whether the GL energy is responsible for the performance improvements. The paper would benefit from comparisons against stronger deterministic representations, including gradient-based texture descriptors, wavelet scattering, and appropriately matched spectral or polynomial features. Dimension-matched comparisons would be particularly helpful. Without these controls, it remains difficult to determine whether the gains arise specifically from the prescribed GL energy or more generally from introducing suitable nonlinear features and spatial inductive biases.

Questions for authors
  1. Can the authors provide component-wise ablations that isolate the contributions of the gradient-energy term, potential-energy term, and spectral aggregation? How would the results change with a different prescribed potential?
  2. Given that the raw spectral baseline substantially outperforms EnRep in the smallest MNIST training regime, under what conditions should one expect the energy transformation to improve rather than degrade discriminative information?
1 comment

Dear Reviewer Li2x,

Thank you for your detailed and constructive suggestions. We agree that the original experiments did not sufficiently distinguish the contributions of energy-derived transformations, spectral organization, and feature dimensionality. We have expanded the ablation studies in the revised manuscript to address these concerns.

  1. Component-wise and dimension-matched ablations

Appendix H (Table 13) now evaluates the gradient and potential components of GL-EnRep separately under both Energy-Density (ED) and Energy-Decomposition (EDec) modes, alongside Ki-EnRep and dimension-matched raw spectral features.

For the 150×150 colorectal histology benchmark, the results at the same 197-dimensional feature size are:

  • Raw Spectral: 26.09%
  • GL gradient-only EDec: 40.06%
  • GL potential-only EDec: 49.93%
  • Ki-EDec: 51.61%

For 28×28 histology, the corresponding accuracies are 25.48%, 41.08%, 43.93%, and 44.32%, respectively.

These comparisons indicate that feature dimensionality alone does not explain the improvements observed on histology. They also show that informative representations can arise from different prescribed energy functionals.

Nevertheless, we do not claim that these results uniquely identify the physical interpretation of GL energy as the cause of improved classification. The nonlinear transformations and their associated inductive biases may also contribute.

The original concatenated GL-EDec representation achieves 54.44% and 58.46% on the two histology benchmarks. Because it contains 393 features rather than 197, its advantage over individual components cannot be attributed solely to their complementarity. We explicitly acknowledge this limitation in Appendix H.

  1. When does the energy transformation help?

We agree that the benefit of energy-induced representations is task-dependent.

The revised experiments use per-image min-max normalization without feature-wise standardization. Section 3.5 explains that standardizing individual spectral coordinates would remove the Laplacian eigenvalue weighting in the GL gradient-decomposition features, thereby altering their intended frequency-dependent scaling.

At the smallest MNIST split (200 training / 50 test), GL-ED achieves 76.40%, compared with 73.80% for Raw Spectral. However, at the 2,000/500 split, Raw Spectral achieves 84.22%, while GL-ED achieves 84.08%, and the other evaluated energy-derived representations perform worse.

The dimension-matched ablations also show that improvements are not uniform across datasets. For example, on the Turing 36-class benchmark, the differences between raw spectral and individual GL components are relatively modest and vary by construction.

Thus, our results do not establish that energy transformations universally improve discriminative information. The strongest improvements appear on the histology benchmarks, where spatial morphology and texture may be particularly relevant to energy-derived features. We have revised our interpretation accordingly.

  1. Alternative energy functional

To examine whether the framework is restricted to GL energy, we have introduced Ki-EnRep, based on a quadratic deviation from the pixelwise training-set mean image.

The reference image is computed without class labels and remains fixed when processing test inputs. Thus, Ki-EnRep does not require a learned feature extractor, although its construction uses an unlabeled training-set statistic.

We emphasize that Ki-EnRep provides a second prescribed energy functional rather than a direct replacement of the GL double-well potential.

Appendix B provides raw spectral comparisons, Appendix C examines mode-count sensitivity, and Appendix H evaluates individual energy components under matched feature-vector dimensions.

We appreciate your suggestions, which have helped us strengthen the experimental analysis and clarify both the contributions and limitations of EnRep.

Reviewer_Jej2OpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
3
Soundness
2(+1)
Read review
Paper summary

Authors propose EnRep, a framework a framework that constructs energy induced representations from prescribed energy functionals and extracts spectral features for image classification. The proposed method is evaluated on a single synthetic datasets, as well as on two image datasets (mnist and medical images). Obtained results suggest that proposed energy based approach can be used to generate representations for classification.

Critical strengths
  • the main advantage of the proposed approach is that it is simple and its computational requirments are low (no backpropagation).

  • reported results suggest that the generated features can provide useful features for the classification.

Critical weaknesses
  • Submitted paper is very similar to [Shin et al. "Exploring potential of Turing pattern classification through convolution maps." Sci Rep], undermining the novelty of the submitted work. After reading the work of Shin et al, I believe that the overlap with the submitted work is not well described in the Related Work. Both papers investigate methods based on Lengyel–Epstein models (similar ones). Fig. 2 in Shin et al and Fig. 2 in the submitted paper use the same parameter grid for illustrating pattern formation (images differ as the models differs). Also, the morphology classes seems to be the same, to point a few similarities.

  • Moreover, the comparison with Shin et al. is incomplete. The proposed method utilizes a single layer CNN with six filters while Shin et al also uses multi-layer networks with convolutional features. It would be good to fully align the evaluation protocols.

  • The proposed method is evaluated on toy datasets, there is no evaluation including a large scale dataset.

Questions for authors

My recommendation is reject. I believe that the overlap with the work of Shin et al is too substancial, undermining the novelty of the submitted paper.

My main suggestions to the authors are as follow:

  • Better highlight the novelty of the proposed methods compared to Shin et al.
  • Compare EnRep using the same classifier and evaluation protocol. It would be good to clarify if the proposed energy based method offers an improvement over the formulation from Shin et al.
2 comments

Dear Reviewer Jej2,

Thank you for carefully examining the relationship between our work and Shin et al. (2026). We take your concern regarding novelty seriously and agree that the overlap with prior Turing pattern studies should have been communicated more clearly.

  1. Relationship to prior work

We acknowledge that the Lengyel–Epstein model, parameter configurations, and morphology labels are closely related to the existing Turing pattern literature. We do not claim these benchmark choices or the underlying GL-inspired quantities as novel.

Section 2.2 clarifies the relationship to Oh and Lee (2020), who investigated engineered GL-related features, and Shin et al. (2026), who studied learned convolutional features and network architectures. We have also moved the detailed Turing pattern generation and parameter grid to Appendix E.

  1. Methodological distinction and novelty

The intended contribution of EnRep is not a new Turing pattern benchmark, a new GL energy functional, or a new convolutional architecture. Rather, it is a framework for systematically constructing compact spectral representations from prescribed energy functionals.

Specifically, EnRep introduces two complementary constructions:

Energy-Density (ED) projects spatial energy-density fields onto a spectral basis, retaining information about their spatial organization.

Energy-Decomposition (EDec) constructs features directly from mode-wise energy contributions after the required spectral transformations.

For GL-EDec, these contributions include a Laplacian-weighted gradient term and a potential term involving the spectral coefficients of both u and u³ (Section 3.3).

The distinction from learned convolutional representations is therefore not limited to the absence of backpropagation. EnRep explicitly organizes energy-related quantities into spectral coordinates or mode-wise energy contributions, rather than learning convolutional filters from classification data.

Furthermore, the revised manuscript introduces Ki-EnRep using a training-mean-centered quadratic energy (Section 3.1.2), demonstrating that the construction is not restricted to the GL functional.

  1. Classifier comparison and additional evidence

We agree that our six-filter CNN does not constitute a fully matched comparison with the deeper architectures investigated by Shin et al.

Our main experiments use the same ridge readout for EnRep and raw-input representations. Appendix H further introduces dimension-matched comparisons against raw spectral features, using the same ridge classifier.

For example, on the 150×150 colorectal histology benchmark, the dimension-matched raw spectral representation achieves 26.09% accuracy, whereas GL potential-only EDec achieves 49.93%, both using 197 features (Appendix H, Table 13).

This provides evidence that the energy-derived construction can contribute discriminative information beyond retaining the corresponding number of raw spectral coefficients, although it does not establish superiority over other deterministic representations or the deeper CNNs of Shin et al.

Appendix G additionally evaluates RBF kernel ridge regression and MLP classifiers while keeping the EnRep features fixed. These experiments demonstrate readout flexibility but are not presented as reproductions of Shin et al.'s deeper CNN architectures.

  1. Evaluation beyond Turing patterns

The revised manuscript includes MNIST experiments with up to 48,000 training samples (Appendix F) and colorectal histology benchmarks at two image resolutions.

These experiments demonstrate that EnRep is not restricted to synthetic Turing patterns, although they do not constitute a comprehensive large-scale natural-image evaluation.

We hope these revisions clarify that the contribution of EnRep lies in its energy-induced spectral representation constructions, rather than the shared Turing benchmark or a claim of superiority over modern CNNs.

Thank you again for your valuable feedback.

I thank the authors for providing a detailed response to my review. However, in my opinion, the submitted manuscript is still too similar to the work of Shin et al. (2026) when it comes to the motivation and conception.

Nevertheless, I will increase my score by one point to weak reject, as I believe rebuttal has improved the submitted paper.

Reviewer_HH26OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

EnRep is a feature representation framework that disentangles spectral features of an energy functional (Ginzburg-Landau, eq. 1) from a downstream classifier. In the context of this paper, a single sample E(u) is not considered as a feature, but rather, the representation Phi_E is derived from the form of the energy functional E by decomposing it to energy density and energy decomposition modes, using three components in order:

  • Q: extracts structural quantities. i.e: The energy density uses an "edge map" e_g and "potential map" e_p (eq. 2) that measure how far a data point is from values +1 and -1.
  • S: converts to a spectral coordinate system. i.e. the energy density mode projects to a real orthonormal Laplacian eigenbasis with DCT-II or FFT (eq 3).
  • T: retains / collects which coordinates are picked. i.e. keeping m x m low frequency DCT block and appending the image mean (Section 3.5).

Similarly, the energy decomposition constructs features directly in the spectral domain. Specifically, Phi_E has no trainable parameters.

The authors validate their framework on image benchmarks of handwritten digits, and colorectal histology, in addition to a synthetic benchmark of Turing patterns (according to Lengyel-Epstein Model).

Critical strengths

Since EnRep has no learned parameters, the computational cost of the optimization itself is lower (Section 5.4 shows the ridge based pipelines finish under 0.13 seconds, v.s. minutes for the main CNN baseline).

Critical weaknesses

I find it hard pointing out where GL-EnRep shines against other representation methods (in particular learned ones).

For validation, I am unfamiliar with Turing patterns (or at least, this is the first time I encounter it in the context of ML). It seems the authors have generated 720 images for validation here, which is not a large benchmark, and I suspect this structure was chosen to fit the suggested features better than textures prevalent in natural images.

I'll therefore focus on the image datasets: EnRep underperforms a weak CNN on MNIST classification (Table 2), and while it outperforms the CNN of colorectal histology classification (Table 3), the main baseline is a weak form of single-layer CNN of 6 filters, which can hardly be considered a practical or fair comparison.

If the contribution of this paper should be practical, perhaps EnRep could have been useful for domains like medical images, where patterns, rather than natural patches are more common. However the authors have made no such claim, and haven't done at which scale a modern day CNN would achieve parity with GL-EnRep.

Minor: The paper does not explain why the energy density and energy decomposition encodings were chosen the way they are. It would have been more convincing to obtain intuition to why the authors made these choices.

Minor: While the representation does not use learned coefficients, it does require manually setting some hyperparameters per domain (i.e. m=10, d=201 for mnist, m=14, d=393 for histology). It's unclear how sensitive the method is to these choices, or how to make them.

Questions for authors

What advantage does this specific form of handcrafted features brings over a learned representation, which may be tuned over a variety of tasks?

Corrections -

line 124 formualtes the representation pipeline as T ∘ S ∘ Q. This fits the energy-density mode well, but the way I read it, the energy decomposition should be T ∘ Q ∘ S as it performs spectral transform first (that would also match Figure 1 that clearly separates Energy -> Spectrum and Spectrum -> Energy).

1 comment

Dear Reviewer HH26,

Thank you for your constructive feedback and for recognizing the computational advantages of fixed energy-induced representations. We have revised the manuscript to clarify the mathematical motivation, experimental scope, and limitations of EnRep.

  1. Operator order and mathematical motivation for ED and EDec

Your observation regarding the operator order was correct. Section 3.1 and Figure 1 now explicitly distinguish the two constructions.

Energy-Density (ED) first constructs spatial energy-density fields and then projects them onto a spectral basis. Energy-Decomposition (EDec) instead obtains the spectral coefficients required by the energy functional and constructs mode-wise energy contributions.

For example, GL-ED projects the gradient and potential energy densities onto the Laplacian eigenbasis. In contrast, GL-EDec directly constructs mode-wise contributions, including the Laplacian-weighted gradient term and the potential term involving spectral coefficients of both u and u³ (Section 3.3).

Thus, the distinction is not merely computational: ED characterizes the spatial organization of local energy densities through their spectral coefficients, whereas EDec describes how individual spectral modes contribute to the energy functional.

We have also introduced Ki-EnRep (Sections 3.1.2–3.4), based on a quadratic deviation from the training-set mean image, to demonstrate both constructions using a second prescribed energy functional.

  1. CNN comparison and practical advantages

We agree that a six-filter, single-layer CNN is a limited learned-feature baseline and does not represent modern CNN performance.

Our contribution is not intended to establish superiority over state-of-the-art neural architectures. Rather, we investigate whether prescribed energy functionals can provide informative representations without optimizing a feature extractor.

The main experiments use closed-form ridge regression, while Appendix G additionally evaluates RBF kernel ridge regression and an MLP readout. In the MLP experiments, only the classifier is optimized through backpropagation; the EnRep representation remains fixed.

The timing results in Section 5.4 apply to the specified NVIDIA RTX A6000 setup and 10,000-epoch CNN protocol. We do not interpret the observed differences as universal speedup factors.

  1. Benchmark scope and mode-count sensitivity

Turing patterns provide a physics-aligned benchmark, while MNIST and colorectal histology evaluate the representations beyond this setting. Appendix F extends the MNIST experiments to 48,000 training samples, and Appendix C reports mode-count sensitivity across the image benchmarks.

We use compact spectral cutoffs of m = 10 for MNIST and m = 14 for histology. These are fixed experimental choices rather than validation-selected or universally optimal values. The sensitivity analyses also show that retaining additional spectral modes does not necessarily improve classification performance.

We hope these revisions clarify the mathematical motivation, practical scope, and limitations of EnRep. Thank you again for your helpful suggestions.

Comments on the paper

4 comments

Dear Reviewers HH26, Jej2, and Li2x,

We sincerely thank all three reviewers for their careful evaluations and constructive feedback. Before addressing the specific concerns, we would like to clarify the intended contribution of EnRep, particularly its relationship to Shin et al. (2026).

  1. Distinction from Shin et al. (2026)

We acknowledge that both studies consider Turing pattern classification and share related mathematical motivations. In particular, the nonlinear features discussed by Shin et al. are closely related to quantities arising from the Ginzburg-Landau energy functional.

However, the central objectives and methodologies differ. Shin et al. primarily investigate the effectiveness of learned convolutional features for Turing pattern classification, including the effects of network depth and data augmentation. In contrast, EnRep constructs features directly from prescribed energy functionals by expressing energy-related quantities in fixed spectral bases, such as the discrete cosine and Fourier bases, and retaining or aggregating selected spectral coefficients or mode-wise energy contributions. This feature extraction does not require learned convolutional filters or backpropagation.

Specifically, EnRep introduces two complementary constructions: Energy-Density Mode, which extracts spectral features from spatial energy-density fields, and Energy-Decomposition Mode, which constructs features from the spectral contributions of the energy functional.

The resulting representations are not restricted to a particular classifier and can be used with different downstream methods, such as linear ridge regression, RBF kernel ridge regression, or multilayer perceptrons (MLPs). In the current submission, linear ridge regression serves as the primary readout, with an additional RBF kernel ridge experiment presented in Appendix G. We also plan to include MLP-based classification results in the appendix of the revised manuscript to further demonstrate the flexibility of the proposed representations.

Thus, while the Turing pattern setting provides a common experimental context, the proposed representation mechanism and the underlying research objectives are distinct. Moreover, EnRep is not restricted to Turing patterns, as illustrated by our experiments on MNIST and colorectal histology images.

  1. Clarification of the Intended Contribution and Further Evaluation

We also acknowledge the reviewers' concerns regarding the experimental evaluation and the need for further validation.

Our intention is not to claim that EnRep universally outperforms modern learned representations or to develop a state-of-the-art classifier. As noted in the ICLR reviewer guidelines, the significance of a contribution does not necessarily depend on achieving state-of-the-art results. Rather, our goal is to investigate whether prescribed energy functionals can provide informative spectral representations without training a feature extractor, while maintaining a computationally lightweight classification pipeline. In this context, we use a shallow CNN as a compact learned-feature reference, while recognizing that comparisons with stronger architectures could provide additional perspective.

We would also like to highlight that the current submission already includes mode-count sensitivity analyses (Appendix C). We plan to conduct additional experiments to further address the reviewers' remaining concerns.

We appreciate the reviewers' comments and recognize that these methodological distinctions and the scope of our claims should have been communicated more clearly in the original submission.

Thank you again for your valuable feedback.

Dear authors,

I acknowledge reading your rebuttal. Since the paper is leaning towards negative scores, I wanted to leave some advice.

As a reviewer, I would have been more convinced if there was one practical application or benchmark where your method shines. As the paper stands now, the evaluation is not strong enough to support the motivation of the paper. I humbly suggest you think about these parts some more, since for me as a reviewer, it was not immediately clear from the paper what the benefits of the method are.

Dear Reviewer HH26,

Thank you for reading our response and for your thoughtful advice. We understand your concern that the practical benefits of EnRep are not yet sufficiently demonstrated, and appreciate the opportunity to clarify our experimental evidence.

The Turing pattern benchmarks provide a physics-aligned setting where GL energy is closely related to spatial morphology. On the 4-class task, GL-EDec achieves 99.26% accuracy, compared with 37.41% for raw-pixel ridge regression and 82.41% for the six-filter CNN. On the 36-class task, GL-ED achieves 72.78%, compared with 10.56% and 22.99%, respectively (Table 1). These results demonstrate the effectiveness of fixed energy-induced representations when the energy structure aligns with the task.

Beyond synthetic patterns, colorectal histology provides evidence of usefulness on real images. On the 28×28 benchmark, GL-EDec achieves 54.44%, compared with 25.50% for raw-pixel ridge regression and 49.54% for the six-filter CNN (Table 3). On 150×150 histology, GL potential-only EDec achieves 49.93%, compared with 26.09% for Raw Spectral, using the same 197 features (Appendix H, Table 13). Thus, feature dimensionality alone does not explain the improvement.

EnRep also offers computational advantages through fixed feature extraction and closed-form ridge regression. On 28×28 histology, GL-EDec takes 0.0052 s, compared with 4598.9 s for the CNN trained for 10,000 epochs (Table 4). This comparison is specific to our experimental protocol and does not imply a universal speedup.

We acknowledge that these results do not establish superiority over modern CNNs or effectiveness in deployed applications. Nevertheless, they support the potential of EnRep to provide informative, compact representations without training a feature extractor, particularly for spatially structured data. We appreciate your constructive guidance and agree that stronger practical validation would further strengthen our work.

Dear Reviewer HH26,

We sincerely appreciate your valuable observation that the practical benefits of EnRep were not immediately clear from the paper.

In particular, EnRep is motivated by small-data, low-compute, and fast-refitting scenarios, where fixed energy-induced representations can be reused with lightweight downstream classifiers without retraining the feature extractor. Our intention is to provide a computationally efficient alternative to learned feature extraction in such settings, rather than to claim universal superiority over modern neural architectures.

Thank you again for your constructive guidance, which has helped us better clarify the intended contribution of our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.