acceptodds
Under review as a conference paper at ICLR 2027

Group-Orbit Modern Hopfield Networks

Abstract

Modern Hopfield networks retrieve over individual vectors, even when rotational symmetry generates infinitely many equivalent observations of each semantic object. We introduce Group-Orbit Modern Hopfield Networks (), together with a theory and complete -step update algorithm for aggregating transformation evidence within each quotient orbit and retrieving over semantic orbit classes. The update admits finite-sampled and continuous-Haar evaluation under the same orbit- level retrieval semantics. We prove equivariance of the update, orbit invariance of its energy, and separation-dependent retrieval and capacity certificates for finite banks and smooth positive-dimensional orbits. On official splits, the continuous posterior mean reduces hidden-atom by 10.8–22.8% relative to nearest-pose Procrustes retrieval and wins all 40 split-condition comparisons. These results show that orbit memory can improve symmetry-aware retrieval without treating equivalent poses as separate semantic memories.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Hopfield, Group, Orbit-Pooling, Symmetry-compatible dynamics, Group-orbit associative memory

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Wx1iOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces Group-Orbit Modern Hopfield Networks (G-MHNs), which represent semantic memories as transformation orbits rather than individual vectors. The method aggregates pose evidence within each orbit and uses the resulting orbit-level posterior to update the retrieved state. It supports both finite-sampled evaluation and continuous integration, with Fourier/Bessel and matrix-Fisher implementations for SO(2) and fixed-correspondence SO(3), respectively. The paper establishes equivariance, energy invariance, and conditional retrieval and capacity guarantees. Experiments investigate symmetry-aware retrieval on ModelNet40 and molecular coordinate completion on rMD17, together with numerical audits and boundary controls.

Critical strengths
  1. The paper distinguishes an object's semantic identity from its pose and represents equivalent observations within a single orbit. This provides a coherent framework for combining pose-invariant object probabilities with equivariant retrieved states.
  2. Both implementations follow the same orbit-level retrieval rule. The continuous constructions avoid explicit pose grids in the supported settings, while the theoretical analysis accounts for geometric separation and approximation errors.
  3. The experiments include matched comparisons, numerical-accuracy audits, and adverse controls. The paper also distinguishes improvements in conditional-mean reconstruction from the effects of recurrent updates, helping clarify what the reported gains demonstrate.
Critical weaknesses
  1. Practical guidance on backend selection could be more concise. The paper carefully documents numerical accuracy and computational trade-offs, but a consolidated explanation of when to prefer sampled retrieval over continuous evaluation would improve usability.
  2. The distinction between reconstruction and recurrent retrieval could be emphasized earlier. The experiments acknowledge that recurrent updates can improve proximity to stored modes while increasing coordinate reconstruction error. Making this distinction more prominent would help readers interpret the role of the update budget.
Questions for authors

Could the authors provide practical guidance on choosing the retrieval backend and whether to use direct readout or recurrent updates?

Reviewer_5zVZOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Group-Orbit Modern Hopfield Networks (G-MHN), in which each stored memory is the orbit of a prototype under a group action, rather than a single vector. The update aggregates softmax evidence over all poses of each prototype. The authors prove invariance of the object posterior and equivariance of the retrieved state for the continuous (Haar) version, derive an orbit-invariant energy, give closed-form Haar statistics for SO(2) and SO(3), and give contraction, noise-robustness and capacity guarantees whose constants are computed numerically for a given bank. Experiments cover hidden-atom completion on two rMD17 molecules, a masked-feature task on ModelNet40, synthetic checks of the theory, and two negative controls.

Critical strengths

The continuous (Haar) formulation is clean and correct: invariance of the object posterior, equivariance of the retrieved state, and an orbit-invariant energy hold exactly, and the SO(2)/SO(3) evaluations are implemented carefully with certified numerics. This matters because, if a setting is found where orbit memory does something a weighted average of aligned items cannot, the machinery to exploit it is already in place.

Critical weaknesses
  1. The central empirical result does not demonstrate orbit memory. The authors' own Sec. 5.3 shows that the continuous posterior mean coincides with a soft (softmax-weighted) Procrustes average to within 0.0022 A, and that the best hidden-atom RMSD is reached at T = 0, with each recurrent step making it worse. The reported 11-23% gain over nearest-pose Procrustes is therefore the usual gain of an average over a single nearest neighbour. It needs neither the orbit construction nor the Hopfield dynamics, and the baseline that would show this (soft Procrustes) is missing from Table 2.

  2. The theory is never connected to the experiments. A large part of the paper consists of guarantees (Theorems 2–4) that retrieval succeeds when the memory bank satisfies certain conditions. These guarantees are tested only on a synthetic bank of 12 random objects, not on molecules, and even there the bound is loose: it predicts a failure probability of 26% where the observed rate is 0.1%, and in the same audit strict retrieval success drops to 0 for three-dimensional orbits, which is the molecular case. For the actual rMD17 banks none of the quantities entering the theorems is computed, and the authors acknowledge that no certificate exists for them. On ModelNet40 the stored vectors are encoder outputs on which rotations do not act by any matrix, so the theory does not apply to that experiment at all.

Questions for authors
  1. Can the authors include, as a competitor in the molecular experiment, a simple weighted average of Procrustes-aligned stored conformations, and show results for memory banks substantially larger than the one used?

  2. In the ModelNet40 experiment, how much of the reported gain remains when the same number of update steps is applied to an ordinary MHN without rotated copies?

  3. Do the quantities that enter the retrieval and capacity guarantees take non-trivial values on the actual memory banks used in the experiments?

Reviewer_iwyDOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper find a essential operation that is one of the weak point of associative memories (Hopfield networks), which is rotations. The proposed model, Group-Orbit Modern Hopfield Network make sure that (1) same object with different rotation still matches the same object and (2) rotating the query also rotate the retrieval by the same amount. The paper formalize a memory object as a group orbit, and averages possible transformation within one orbit. Furthermore, the paper developed a energy-based interpretation, efficient continuous integration methods for 2-D and 3-D rotations, and sufficient condition for memory convergence and memory capacity.

Critical strengths
  1. Good motivation and important problem solved. Rotation is one of the most common operation that can happen in real world scenario, and using group operation to remove the impact of rotation is elegant, as it do not require a large memory bank containing copies of certain object in different angles.
  2. The framework considers invariant rotational energy, equivariant update/retrieval, and the unified theoretical framework have good interpretability, and it can accounts for why G-MHN works on rotations.
  3. Honest disclosure on failure on insufficiently separated representations are appreciated.
Critical weaknesses
  1. In the full table (Table 23) of the headline result (Table 2), it seems like the baseline soft (Soft Procrustes) is better than G-MHN (column continuous)? Can the authors give us insights why soft works so well?
  2. The experiments in the main part (excluding the appendix) is not convincing. It seems like you have conducted experiments in rotated image/point cloud restoration/retrieval in Appendix G and H. But can you provide me quantitative results? It would be better if you report the accuracy (occurrence when the retrieved object matches the ground truth, which is the object that generates the query) of retrieval instead of just showing examples in Fig. 4 and Fig. 5. And compare with the retrieval accuracy of the baselines (incl. soft Procrustes), and report the results in the main part.
  3. Can you briefly discuss that is it possible to use group operations to formalize other operation such as scaling/displacement? It would be great if G-MHN can handle all kinds of operations, not just rotation.
  4. There is a weird spacing problem at the beginning of the first line in Introduction (line 28).
Questions for authors

See the weakness part.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.