acceptodds
Under review as a conference paper at ICLR 2027

Blind Recovery of Signal Domains via Unsupervised Symmetry Discovery

Abstract

A signal-processing perspective interprets data as discretized signals over geometric domains. When the underlying domain is inaccessible, we lose geometric priors, hindering interpretability and generalization. We propose an unsupervised framework for recovering signal domains by discovering symmetries of the data distribution. Our framework treats data samples as unknown linear measurements of continuous latent signals, and learns to recover them through a linear group-convolutional network. Assuming latent signals form a stationary Markov random field, we train the model by enforcing stationarity, minimizing total correlation, and maximizing information. We theoretically show that the proposed approach transforms an otherwise unidentifiable inverse problem into blind deconvolution, and we test the model on stochastic processes, Ising models, scrambled images, and biological neural recordings, revealing latent signals. Our results suggest a promising direction for unsupervised structure learning and blind inverse problems.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
symmetry discovery, blind inverse problems, geometric deep learning, unsupervised structure learning

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 2 reviews, above 27% of the others.
Reviewer_6q9vOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper presents a new practical and theoretical framework for identifying underlying grid (group) structure in measurements that have been scrambled in an unknown way. The proposed method relies on a combination of a recovery network and a loss estimation network together with a group-theoretic parameterization that captures underlying symmetries (in this case, the underlying grid) of latent signals/images. The method is demonstrated in a series of small-scale experiments, including generalized shot noise, Ising model estimation, MNIST pixel shuffling/bit scrambling, and learning the rotational symmetry from mouse visual cortex signals. Comparisons with LieGAN in the shot noise experiments show that the proposed method is more robust to varying configurations.

Critical strengths
  • The paper tackles the highly challenging problem of recovering hidden symmetries (or, in this case, the underlying grid) from unstructured data, which hasn't received significant attention from the machine learning community
  • Experiments, despite the small scale and high compute cost, are promising and can pave the way for new methods.
Critical weaknesses

My main concern with the paper is that it is quite hard to read/understand:

  • The recovery and loss networks are not well defined - what are the learnable parameters of these networks? What is θdiv\theta_{\text{div}} and θent\theta_{\text{ent}} in Figure 2? Similarly, the dataset is not defined anywhere, I guess it is some set of xix_i?
  • The experimental figures (3,4) can be significantly improved: what are the different MNIST images in 3.b - the output of which network? At what moment of training? The axes of subfigures in figure 4 need labeling - what are the different axes values?
  • The abstract and conclusion mention a clear link with blind deconvolution, but the expression "blind deconvolution" is not present in the main sections 3, 4, and 5, and it's unclear what this link is?
  • The experiments mention multiple times "signal recovery", but this concept is not explained anywhere in the main paper - its only mentioned in the introduction and experiments.
  • I believe the notation is overly complicated, and many symbols can be removed, which would improve understanding. For example, are all the intermediate zz and yy representations crucial for understanding? what is the matrix CC? Similarly, the different mathcal, lower/uppercase and bold notations should be simplified, for example figure 2.b alone uses yi\textbf{y}_i, YB\mathcal{Y}_B, YnY_n, Y\textbf{Y} and Y(ℓ)\textbf{Y}^{(\ell)}.
  • It is quite hard to understand how the data is defined on each experiment and how it fits with the model definition. For example, it's unclear what the data is and what the goal is in the generalized shot noise experiment.
  • some mathematical symbols are not explained: what is hh in the total correlation and infomax losses? what is DJSD_{JS}?
  • some acronyms are not explained, such as "JS"
  • In section 6.2, why zero-padding if a central crop is used?
Questions for authors

Does eq. 1 assume that the mixing matrix AA is equivariant? If it has an unknown form, I wouldn't expect it to be necessarily equivariant

1 comment

Dear Reviewer,

Thank you for your time, your constructive criticism, and for acknowledging the strengths of our methodology. Since the necessary revisions are substantial, we have decided to prepare our manuscript for another venue.

Based on your feedback, we have:

  • Defined previously undefined parameters, such as θdiv\theta_{div} and θent\theta_{ent}, which are the parameters of the Jensen–Shannon (JS) divergence and entropy estimators, respectively.
  • Added a subsection on the connection with blind deconvolution to Section 5, which also makes the link to signal recovery much clearer.
  • Simplified the notation by adopting that of Goodfellow et al. (with the exception of random variables).
  • Added definitions of the terms hh and DJSD_{JS}.

In addition, we will:

  • Ensure that the figures are self-explanatory.
  • Make sure that the link between each experiment's dataset and the problem setting, as well as our motivation, is clear.

Regarding your questions:

In section 6.2, why zero-padding if a central crop is used?

Actually, we use random crops. In Figure 3, to focus on the digit itself, we tested the model's response on the center cropped digits, which could be the source of misunderstanding. We will make that clear in the caption.

Does eq. 1 assume that the mixing matrix is equivariant? If it has an unknown form, I wouldn't expect it to be necessarily equivariant.

Here, ρX\rho_X denotes the true symmetry representation induced on the sample space which is also unknown, not the one estimated by the model, which is denoted by ρ^X\hat{\rho}_X. When the mixing matrix is invertible, it guarantees the existence of a translation symmetry representation ρX\rho_X on the sample space. Our method works by discovering this representation, thereby adapting to the unknown mixing matrix. In summary, the only assumption here is the invertibility of the mixing matrix; there are no further constraining assumptions.

Best regards,

The Authors

Reviewer_5bpHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The authors propose recovering the latent space of the signal by discovering the symmetry of the data. They state that this is possible by assuming that the latent signal is part of a stationary Markov random field, constraining the model to be stationary, minimizing total correlation, and maximizing information. On the theoretical side, they establish that without these constraints the inverse problem admits multiple solutions, whereas with them it reduces to blind deconvolution. The proposed model consists of a recovery network, in charge of mapping observations to a symmetry-based representation, and a loss estimation network, which provides the information-theoretic quantities needed to train the recovery network. The framework was tested on stochastic processes, Ising models, scrambled images, and biological neural recordings, revealing the latent signals in each case. The experiments carried out in this work cover an interesting set of approaches for comparing the discovered symmetry and the recovered signal against state-of-the-art methods that are relevant to the proposed method.

Critical strengths
  1. The formulation of the two networks that the authors propose in the framework is well structured, defining the parameters that are learned during training for both networks.

  2. In addition, the authors specify how the loss function terms are configured for training: stationarity, total correlation minimization, and information maximization.

  3. This work presents a set of experiments comparing the proposed method with one method related to data symmetry and others in signal recovery, showing interesting results.

Critical weaknesses
  1. All signal-recovery baselines (CM, GLASSO-SPEQ/FAQ, UMAP, ISOMAP) are second-order or manifold-learning methods, so they cannot resolve the kernel phase, and the comparison mainly shows that second-order statistics are insufficient. The most natural competitor is missing: ICA followed by ordering. Appendix F notes that for zeroth-order MRFs the TC objective reduces to the ICA principle, so a pipeline of ICA for unmixing and then GLASSO-SPEQ or CM for ordering should be included, at least for the random-mixing settings (Ising random, DST-I). In addition, LieGAN is compared only at d = 15; the higher-dimensional experiments have no symmetry-discovery baseline.

  2. The paper presents the reduction of blind linear inversion to blind deconvolution as a central contribution, but the argument is spread across Sec. 5 and Appendices A and F. The resulting problem, Y = KF with circulant K, appears only in Appendix A.1.4. As I understand it, stationarity makes the recovery kernel translation-invariant (Thm. D.8, Lemma F.6). InfoMax then fixes its magnitude spectrum (Prop. C.13), so the remaining ambiguity is an all-pass filter. Total correlation minimization finally resolves this phase ambiguity via the compact-support argument (Prop. F.9–F.11). If this reading is correct, it should be stated explicitly in the main text. It also connects to the classical result that second-order statistics cannot identify filter phase, while non-Gaussianity can (e.g., minimum entropy deconvolution, Shalvi–Weinstein). That literature is not discussed, although it would explain why the second-order baselines (CM, GLASSO-SPEQ) fail. The reduction is also never tested against existing blind deconvolution methods, since all three losses are optimized jointly. I suggest an oracle experiment: fix the learned representation to the true symmetry and compare the TC criterion with classical higher-order or sparsity-based deconvolution, such as minimum entropy deconvolution, ICA, or multichannel sparse blind deconvolution (Qu et al., 2019; Shi & Chi, 2021). A two-stage variant would also help: train with stationarity and InfoMax only, freeze the representation, and then apply these methods. The current ablation (Table 4) shows that removing TC causes failure, but not whether TC is preferable to established deconvolution criteria.

Qu, Q., Li, X., & Zhu, Z. (2019). A nonconvex approach for exact and efficient multichannel sparse blind deconvolution. In Advances in Neural Information Processing Systems 32 (NeurIPS 2019).

Shi, L., & Chi, Y. (2021). Manifold gradient descent solves multi-channel sparse blind deconvolution provably and efficiently. IEEE Transactions on Information Theory, 67(7), 4784–4811.

Questions for authors

The largest images are 27×27 MNIST crops (d = 729), where a single run takes ~132 h on an A100 and performance already drops relative to 15×15 (r: 0.993 → 0.947; S₀.₇₅: ≈0.90 → ≈0.72). Given the cubic cost per update (Sec. 6.4), how is the method expected to scale to high-dimensional images (e.g., 64×64, 128x128,etc)? Is the dimensionality-reducing conditioning matrix of the bit-scrambling experiment a general solution, and at what cost in recovery quality?

1 comment

Dear Reviewer,

Thank you for your time, your constructive criticism, and for acknowledging the strengths of our methodology. Since the necessary revisions are substantial, we have decided to prepare our manuscript for another venue. In response to the weaknesses you raised, we have conducted preliminary experiments.

The most natural competitor is missing: ICA followed by ordering. [...] at least for the random-mixing settings (Ising random, DST-I).

Thank you for this sharp observation. We ran ICA (total correlation minimization) followed by perfect ordering:

DatasetddTransformICA (perfect ordering)Ours
GSN-G63DST-I0.783±0.0070.783 \pm 0.0070.931±0.015\mathbf{0.931 \pm 0.015}
GSN-L63DST-I0.449±0.0060.449 \pm 0.0060.992±0.001\mathbf{0.992 \pm 0.001}
Ising33Random0.307±0.0010.307 \pm 0.0010.992±0.004\mathbf{0.992 \pm 0.004}

LieGAN is compared only at d = 15; the higher-dimensional experiments have no symmetry-discovery baseline.

We ran experiments at higher dimensions for continuous translations. LieGAN fails in most settings; however, the results should be read within the fact that we force LieGAN beyond its operating assumptions, since we model the latent translations as infinite-dimensional to match the natural setting.

DatasetddTransformLieGAN S0.75S_{0.75}Ours S0.75S_{0.75}
GSN-G63Identity0.693±0.1850.693 \pm 0.1850.996±0.002\mathbf{0.996 \pm 0.002}
GSN-L63Identity0.909±0.0030.909 \pm 0.0030.910±0.022\mathbf{0.910 \pm 0.022}
GSN-G63DST-I0.454±0.1530.454 \pm 0.1530.994±0.002\mathbf{0.994 \pm 0.002}
GSN-L63DST-I0.911±0.0280.911 \pm 0.0280.976±0.007\mathbf{0.976 \pm 0.007}
Ising33Permuted0.304±0.0180.304 \pm 0.0180.999±0.000\mathbf{0.999 \pm 0.000}
Ising33Random0.294±0.0250.294 \pm 0.0250.955±0.026\mathbf{0.955 \pm 0.026}

[...] the remaining ambiguity is an all-pass filter. Total correlation minimization finally resolves this phase ambiguity [...]. If this reading is correct, it should be stated explicitly in the main text.

Your reading is correct, and we now state it explicitly in Section 5. Thank you for your in-depth reading and the suggestion.

[...] second-order statistics cannot identify filter phase, while non-Gaussianity can (e.g., minimum entropy deconvolution, Shalvi–Weinstein). That literature is not discussed [...]

Thank you for pointing this out; we will add a discussion of this literature.

I suggest an oracle experiment: fix the learned representation to the true symmetry and compare the TC criterion with classical higher-order or sparsity-based deconvolution [...]

Thank you for the suggestion; we will add this experiment.

A two-stage variant would also help: train with stationarity and InfoMax only, freeze the representation, and then apply these methods.

The ablation study (Table 4) shows that removing TC leads to failure not only in signal recovery but also in symmetry discovery. This may stem from a coupling between the symmetry and the local dependency structure, such that the symmetry operation can also be effective at resolving it. However, the two-stage variant may succeed in lower dimensions such as d=15d = 15, where distributional invariance alone may suffice for symmetry discovery; this is worth testing.

Regarding your question:

[...] how is the method expected to scale to high-dimensional images (e.g., 64×64, 128×128)? Is the dimensionality-reducing conditioning matrix of the bit-scrambling experiment a general solution, and at what cost in recovery quality?

If the method is run directly on D×DD \times D image patches, the computational cost scales as O(D6)O(D^6). There are two possible general workarounds:

  • Using a dimensionality-reducing conditioning matrix (tested in the bit-scrambling experiment).
  • Applying PCA prior to training (not tested).

The performance degradation would depend on the effective rank of the dataset. Given our adaptive rank estimation algorithm, the variance explained by the top dZd_Z components could be a reasonable heuristic.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.