acceptodds
Under review as a conference paper at ICLR 2027

A Local Associative Memory that Knows Where It Holds

Abstract

An exemplar memory must both retrieve corrupted versions of stored items and reject cues absent from the store. We cast these as two outcomes of one local deci- sion: whether a cue lies in a region where a stored memory has grounds to answer. The escape and retrieval field (ERF) assigns each memory such a region using the spherical support of a truncated Gaussian well, together with a shared outward radial drift. Its static rule accepts cues inside the union of the supports and re- jects cues outside. Its dynamics provide conditional geometric guarantees for any chosen radii: every isolated support is forward invariant, while a cue whose ray from the reference point meets no support escapes geometrically. We construct the regions in two ways. Geometric radii, obtained from the store’s angular spacing, yield a closed-form tolerance to missing coordinates filled with the reference point and capacity bounds for random memory directions. Fitted radii instead adapt a metric and one region per memory using only stored items and their corrupted views. Across two frozen encoders, three stores and six corruption families, the fitted metric identifies the source memory with 87 % accuracy, compared with 64 % for modern Hopfield retrieval in the encoder’s original space. Because both methods select the same identity in the fitted metric, this gain comes from metric adaptation; the fitted radii determine whether to answer. Rejection performance depends on the store’s separation: above an exploratory threshold, ERF has the highest balanced accuracy in 83 % of conditions and on average in every corrup- tion family; below it, ERF trails the strongest comparator on average. With the threshold fixed, a subsequent test predicts the sign of ERF’s mean advantage in all 10 new store–encoder settings, but only a fraction 0.68 of individual conditions. ERF therefore unifies retrieval and rejection through the local regions in which its memories hold.

Primary area:
learning on time series and dynamical systems
Keywords:
hopfield network, associate memory

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_XvRoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes an escape and retrieval field (ERF). This is a memory model that attracts queries near stored examples and pushes them away elsewhere. This approach allows for geometric results on retention, escape, and capacity. The image experiments use a fitted static rule, with a learned representation and a separate radius for each memory

Critical strengths
  1. The rejection mechanism is a real strength. Its geometry makes it clear where a query is supported. The local guarantees give this interpretation a mathematical basis.
  2. Comparisons of individual versus global radii help show what the fitted rule contributes. The supplementary code also makes several implementation choices clear.
Critical weaknesses
  1. The experiments leave a question about what ERF adds. They use static acceptance and identification matches the comparator for the fixed representation. A static method is useful, but the paper needs to clarify its added value
  2. There are some gaps in theory and experimental record. For Proposition 1, c=0,  m=0.05,  R=1,  κ=0.1c = 0,\; m = 0.05,\; R = 1,\; \kappa = 0.1 gives two stationary points, x1≈−0.052446x_1 \approx -0.052446 and x2≈0.152446x_2 \approx 0.152446. This looks fixable. However the package lacks the promised notebooks, main run records, and preregistration documents and does not fully control minibatch randomness. These issues make it difficult to verify important claims
Questions for authors
  1. For the fixed representation, which result best shows what the method adds? Which guarantees still hold when fitted supports overlap?
  2. How would you correct the assumptions of Proposition 1? Does this change any later result?
  3. Could you provide the promised files and explain how minibatch sampling was seeded? From the existing main runs, could you also report accuracy among accepted queries, coverage, and false acceptance of unknown queries?
Reviewer_p3ApOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies two tasks in instance-level associative memory: retrieving stored samples from corrupted inputs, and rejecting inputs that do not belong to the stored set. The authors propose the Escape and Retrieval Field (ERF): each memory is represented by a truncated Gaussian potential well and its spherical support region, plus a shared radial extrapolation term centered at a reference point c. The static rule accepts when the input lies within any support region and rejects otherwise; the dynamics are attracted by the potential well inside the support regions and escape radially outside all support regions.

The paper gives two types of radii. The geometric radius is determined by the angular separation, yielding an exact self-support-region criterion when missing coordinates are filled with the mean center, as well as a tolerance–capacity relationship under a random-direction model. The fitted radius learns a 128-dimensional linear metric on frozen DINOv2 and ResNet encoders, and is set according to the 0.9 quantile of each memory's own synthesized-view distances and the nearest distances of other memories' views.

Experiments include static recognition and dynamics checks under the geometric radius, as well as static detection and identity retrieval under the fitted radius. The fitted experiments cover two frozen encoders, three storage scales on CIFAR-100 and DISC21, and six types of perturbations; the paper reports an identity accuracy of 0.870 for the fitted metric, compared with 0.638 for modern Hopfield in the original encoding space. It also explores the influence of sample-set separability on the accuracy of the ER method.

Critical strengths
  1. The problem design and definition are clear, summarizing how to refuse to answer as whether a clear sample can be found as evidence. A clear dynamical method is defined by using potential fields.

  2. The paper's proofs are detailed in length and include the proof content required.

  3. The paper's reporting is transparent, giving clear conclusions about when ERF fails.

Critical weaknesses
  1. The numerical calculations of the paper's conclusions are unclear and need more detailed explanation. For example, the values 0.87 and 0.638 in the text, as well as the proportion of ERF failures, all directly depend on the 200 conditions explained in the supplementary material. The main text does not clearly explain the calculation process for these numbers, causing confusion.

  2. Intuitively, the current work's ERF itself relies on unclear foundations, and it is not yet apparent what help the introduction of the dynamical field provides compared with static methods. Theorems 1 and 7 in the paper rely on the isolated-support assumption, requiring all pieces of evidence not to overlap with one another. Once this condition is satisfied, the meaning of the dynamical field itself basically disappears.

  3. The design of the fitted radius directly breaks the isolation requirement, and at the same time, because of insufficient separability, it still performs worse than the other methods mentioned in the paper. This reduces the credibility of the method on multi-sample multi-class settings.

  4. The baselines cited in the paper and the cited papers themselves are insufficient; the sources of the baselines are not clearly stated, and there is no more detailed explanation of related work.

  5. The paper is too long, and it repeatedly mentions explanations of theoretical inapplicability; the writing itself is difficult to read. The reproducibility of the method is also weak, and the setting of W in the calculation of the fitted radius in the paper is unclear.

Questions for authors

see weakness

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

This paper proposes the Escape and Retrieval Field (ERF), an associative memory framework that jointly addresses retrieval and rejection through local support regions around stored memories. The authors derive geometric guarantees for isolated supports and evaluate fitted support radii on image retrieval and novelty detection tasks.

Critical strengths

S1. The paper addresses an important limitation of associative memory systems: determining when a query should be rejected rather than retrieving an unrelated memory. The local-support formulation is intuitive and provides a principled geometric perspective.

S2. The theoretical analysis provides useful conditional guarantees, and the experiments explicitly distinguish metric adaptation from support-radius calibration. The authors also acknowledge settings where their method underperforms, which improves the credibility of the evaluation.

Critical weaknesses

W1. The practical contribution of the proposed energy field is unclear. In the fitted setting, the main experiments use a static support-membership test rather than the proposed dynamics. Furthermore, the improvement in identity retrieval comes from metric adaptation, and modern Hopfield retrieval produces identical identities under the same fitted metric. This raises questions about what the energy-field construction contributes beyond conventional metric learning and distance thresholding.

W2. The theoretical guarantees are weakly connected to the main empirical setting. The principal invariance results assume isolated support regions, whereas fitted supports may overlap. Moreover, the static acceptance rule can accept a query because it lies within one memory's support while returning a different nearest memory outside its own support. This weakens the interpretation that each returned memory is supported by its local region.

Questions for authors

How much does ERF improve over a simple nearest-neighbor classifier with the same learned metric and independently fitted per-memory thresholds? Can the authors report joint acceptance-and-correct-identity accuracy?

Can the authors establish theoretical guarantees applicable to overlapping fitted supports, or clarify which aspects of the energy-field construction are essential to the reported empirical improvements?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.