acceptodds
Under review as a conference paper at ICLR 2027

The Readout Decides Which Hopfield Networks Have a Constant-Metric Energy

Abstract

A modern Hopfield network retrieves a memory by iterating a map that scores the stored patterns against a query, weights them by a softmax, and forms the next state with a readout. Its iteration is backed by an energy that never increases from one step to the next. However, the frameworks that generalise its score and weighting keep the readout a weighted average of the stored patterns, which we call the pattern readout. We show that the readout decides which scores make each step a concave-convex step on an energy in a constant metric. We call a map with this property admissible. In particular, we consider scores whose gradient splits into a fixed per-pattern part and a part shared by all patterns. Among them, and for patterns in general position, the pattern readout is admissible in the Euclidean metric essentially only for the dot product. In contrast, the gradient readout takes the gradient of a potential of the scores. It is admissible for every score that is convex up to an added quadratic. Among 32 published retrieval maps, 8 are shown not to be admissible. For 5 of them, one reason is that their readout is not the gradient of their score's potential. One of them is the adaptive Hopfield network, which can enter cycles of period two as a multiple-instance pooling layer. We keep its score on squared distances, switch to the gradient readout, and constrain its learned weights to be non-negative and non-increasing. These changes give an exact energy that is bounded below throughout training. In our multiple-instance experiments, non-negative weights alone already stop the cycling under the pattern readout, so the energy adds a guarantee that no step raises it. Overall, our results identify the readout as the component that decides which scores give an admissible map, and they show how to test and design admissible retrieval maps.

Primary area:
learning theory
Keywords:
Hopfield networks, Associative memory, Energy function, Retrieval dynamics, Convex analysis

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_tyjUOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The authors propose a theoretical framework for a unified stability analysis of a particular class of associative memory models. The models in question are formulated in the style of Universal Hopfield Networks https://arxiv.org/abs/2202.04557. These are maps consisting of a scoring rule, a separation rule, and a readout. For such models, the authors offer a way to analyse stability. The construction is applied to many published models of associative memory.

Critical strengths

The most appealing part of the analysis is a strive for generality. It would be great to have stability results for a large class of memory models, and certainty about their stability properties without needing to repeat a tedious stability analysis for each of them separately. The authors partially succeed in this by defining GG-admissible models and studying when the model is GG-admissible.

Another strong side is rigour: most of the statements made in the article are in the form of lemmas and theorems. Such a structure makes it easier to understand the content of the paper and formally check the conclusions of the authors. I read the proofs selectively and found them convincing.

Critical weaknesses

There are a few critical problems with the results of the paper.

The article is poorly written and organised. Many verbal explanations are unnatural; it is hard to decipher them and hard to understand their meaning. Certain concepts and setups are not explained or explained indirectly, which also makes the paper harder to navigate.

The authors' goal was, presumably, to analyse a large class of memory models and draw some useful conclusions based on this analysis. The intended generality is reflected in many places in the article: (i) the title is a general statement about Hopfield networks; (ii) the authors analyse 32 published designs of Hopfield networks; (iii) the theory itself is written for networks with three general components. While the generality of analysis is a noble goal, in my view, the analysis of the authors is not general and does not allow drawing broad conclusions on the stability of memory models. This is my main technical objection, so I explain it in detail below.

Questions for authors

Technical contribution and implication for stability

On a technical level, the contribution offers a stability analysis. The approach is based on the well-known Lyapunov theory: in place of studying trajectories of the dynamical system, try to find a scalar function (energy) that is non-increasing (better, decreasing) on trajectories; if such a function exists under certain mild conditions, nothing but fixed points can be found in the domain when the conditions on the energy function hold.

The authors define their energy function in Definition 1, and call the map "GG-admissible" if, under their definition, a particular choice of energy parameters ensures that the energy is non-decreasing on the trajectory.

Crucially, if the conditions stated in Definition 1 fail, it does not mean that there is no Lyapunov energy function for a given dynamics. Yet, the authors make it sound as if being "non-admissible" is something bad. Note that "G" is often dropped, e.g., in the introduction: "We apply these results to 32 published retrieval maps. Eight of them are shown not to be admissible." Relatedly, in Table 1, the authors state "no constant metric and no convex potential make the map admissible".

Given this assessment and context, here are my questions:

a. What is the reason the authors decide to call this particular property "G-admissible" and why do they find it is fine to switch to simply "admissible" and "not admissible", which is a much stronger property, normally meaning "can not be allowed"?

b. Suppose my map T(x)T(x) is "inadmissible" according to the authors' criteria. So what? It doesn't mean the dynamics is bad, no fixed points exist, no other Lyapunov function exists, etc. It only means the particular construction of the authors does not work. Then why should the readers care that the map is not "admissible"?

c. In my view, the title of the paper, being too general, is misleading. Can the authors explain how their results support the assertion made in the title?

Setups and assumptions

I want to clarify several assumptions and preferences the authors made without a proper discussion:

  1. Why is stability important at all? Is it assumed that the map T(x)T(x) is iterated many times until convergence? Note that there are some contributions where it is suggested that the map is applied once and appropriate results are shown, e.g., in https://arxiv.org/abs/2008.02217.
  2. After Definition 1, the authors write: "The definition guarantees only descent, with no lower bound on EE and no memories at the fixed points." Is it assumed that ξμ\xi_{\mu} should directly correspond to memories? If so, why? Is it appropriate in the authors' view that there are fixed points that are distinct from ξμ\xi_{\mu} but are defined by them?
  3. Often, an additional requirement for Hopfield networks is biological plausibility, e.g., https://arxiv.org/abs/2008.06996. Do the authors believe it is an important property, or are they ready to abandon it? Why not focus on simple systems such as gradient flows, e.g., https://arxiv.org/abs/1910.02720?

Organisation of the paper

I personally find many phrases to be unnatural; below is an incomplete list. I explain why a few of them sound weird, and the rest are listed with no explanation. In general, it is often possible to infer what the meaning of a particular construction is, but it is often unnecessarily hard.

  1. line 260 "A map without an energy need not cycle."

    There are limit cycles in dynamical systems, so the trajectory (states) forms a cycle under a given map. So while it is possible to infer what "map cycles" means, the phrasing is unnatural.

    Another problem is that "without an energy" is undefined. As I explained above, the results of the authors do not imply the absence of energy, at best an absence of a particular energy.

    My translation of this phrase is: "Even when the conditions listed in Definition 1 do not hold, or, more generally, when a discrete dynamical system admits no global Lyapunov function, its trajectory does not need to contain limit cycles."

  2. line 325 "An adaptive score that keeps its energy"

    What is the meaning of this phrase precisely? What does it mean to "keep someone's energy"?

    My translation is: "In this section, we show how to construct an update rule T(x)T(x) based on a particular score function such that it admits a global energy function from Definition 1."

  3. line 164 "What the pattern readout admits"

  4. line 263 "To see which maps can cycle"

  5. line 324 "A monotone gate gives the gradient readout an exact energy"

  6. line 368 "The energy holds throughout training"

  7. line 400 "Where the construction stops."

  8. line 475 "Our map pays for its energy in iterations."

  9. line 478 "We expect slow contraction to cause both the gap and the cost."

In general, I would insist on improving the organisation of the paper.

For example, a few phrases at the beginning of the introduction explain the setup with a weird mixture of equations and words. It is unclear why patterns and TT are specified explicitly and never used in this section, while simultaneously scores, the separation function, and queries are explained only in words.

After such an introduction, Section 2, where the technical setup is really covered, does not go into much detail. Why not write here that we denote memory patterns by ξ\xi, that we assume a particular structure of the map, that we iterate the resulting map, etc.? Why not define the stability problem and back up the claim from the introduction that "in both frameworks the energy is proved for a different map from the one that the model iterates"? Sure enough, all this information is available or inferable from the context, but it is scattered in various places in the article, so it becomes harder for the reader to understand what is going on, unless they already have all the necessary context.

5 comments
AuthorsOpenReview ↗

Response to Reviewer tyjU (1/2): the name, what a cross means, and the title (a, b, c)

We thank the reviewer for the careful reading, and for the kind words about the generality and the rigour of the analysis. Line numbers refer to the revised PDF, where added and reworded passages are in blue.

(a) "why do they find it is fine to switch to simply "admissible" and "not admissible", which is a much stronger property"

We agree. Dropping the GG was a mistake, and "not admissible" reads as a verdict that our results do not support. We chose the name because the condition admits a map into the energy argument of the modern Hopfield network. Each step becomes a concave-convex step (Yuille & Rangarajan, 2003) on E=12x⊤Gx−hE = \tfrac12\mathbf{x}^\top G\mathbf{x} - h, the form of the energy of Ramsauer et al. (2021) with G=IG = I. The main text no longer writes "admissible" without a metric. It keeps "GG-admissible" in Definition 1, the results and the table entries, and elsewhere says that a map has, or has no, constant-metric energy, which it defines through Definition 1 (line 86). For example:

WhereSubmissionRevision
Introduction"Eight of them are shown not to be admissible.""find eight that have no constant-metric energy at a tested configuration and no energy proved for them in their own papers" (line 101)
Table 1Column "Admissible". A cross means that "no constant metric and no convex potential make the map admissible".Column "Constant-metric energy". A cross means a non-real Jacobian eigenvalue at a tested configuration, and "The map may still converge and have an energy of another kind."

The appendix keeps the short form and states this convention in one sentence (Appendix A).

(b) "So what? It doesn't mean the dynamics is bad, no fixed points exist, no other Lyapunov function exists, etc."

We agree with this reading, and the revision now states it after Definition 1 (line 168): a map without a constant-metric energy "may still converge, have fixed points and have a Lyapunov function of another form, so the lack of such an energy does not mean that the map is unstable." Our 32 maps include each of these cases:

Maps without a constant-metric energyWhat we find
tanh memory at M<dM < d, and two others (Table 8)A proven energy outside Definition 1.
4 of the 6 maps from rows with a cross that we iteratedThey converge in every bag: the Manhattan model, one-step attention with a value readout, the nonparametric model and a normalised universal Hopfield network (Appendix H.7). The other two have period-two bags: the adaptive network and a modified universal Hopfield map.
Adaptive Hopfield networkA period-two bag in 250 of 500 runs of its squared-distance form, trained as a multiple-instance pooling layer through five unrolled steps at learning rate 0.1 (Table 3).

A cross matters for two reasons. First, the energy argument behind the convergence proof of Ramsauer et al. does not apply to the map as published, and none of the six papers behind the eight crosses proves an energy for these maps. Second, only a map without a constant-metric energy can have a cycle. Proposition 3 proves that once the softmax saturates, the pattern readout has an attracting cycle of period two near every pair of patterns that are each other's best match. A map with such a cycle has no function of any form that decreases at every step that moves.

(c) "Can the authors explain how their results support the assertion made in the title?"

We agree that the title is broader than the results. "The readout" is supported. Under the pattern readout and for patterns in general position, generic splitting scores have such an energy in the Euclidean metric essentially only for the dot product (Theorem 1), and under the gradient readout every semi-convex score has one (Proposition 1). "Decides" overstates. Necessity under the pattern readout is proved only for splitting scores, and three of the eight crosses come from the score or the separation. "Hopfield networks" is also too broad, since Definition 1 covers only discrete-time maps in a constant metric. The revised abstract and introduction therefore say "restricts" in place of "decides". To avoid any confusion during the review process, we have kept the current title. However, if the paper is accepted, we will revise it accordingly, for example to "The Readout Restricts Which Scores Give a Modern Hopfield Update a Constant-Metric Energy", if the program chairs permit the change.

AuthorsOpenReview ↗

Response to Reviewer tyjU (2/2): setup, assumptions and writing

Setup 1: "Why is stability important at all? Is it assumed that the map T(x) is iterated many times until convergence?"

No. Definition 1 bounds every single step, E(x)−E(T(x))≥12∥T(x)−x∥G2E(\mathbf{x}) - E(T(\mathbf{x})) \ge \tfrac12\|T(\mathbf{x}) - \mathbf{x}\|_G^2, so a map with a constant-metric energy never raises EE, even when it is applied once, as in the one-step retrieval of Ramsauer et al. (2021). The revised Section 4 adds that our verdicts say nothing about whether one step retrieves the right pattern (line 312). Stability matters when the map is iterated. The adaptive network's paper states an energy that decreases monotonically and converges (its Theorem 2, quoted in Appendix E.4), although its released code applies the map once (line 152). When we iterate it as a multiple-instance pooling layer trained through five unrolled steps at learning rate 0.1, its period-two cycles change the predicted label between steps 999 and 1000 in 9 to 26% of held-out bags on squared distances, and in 9 to 27% with both bases (Table 2).

Setup 2: "Is it assumed that ξ^μ should directly correspond to memories? […] Is it appropriate in the authors' view that there are fixed points that are distinct from ξ^μ but are defined by them?"

No result in the paper assumes that the stored patterns are fixed points. Definition 1 guarantees descent only, and the revision says after it that the definition "neither bounds E below nor ensures that the fixed points are stored patterns" (line 166). We think such fixed points are appropriate. The modern Hopfield network has them too: Ramsauer et al. describe metastable states that average over subsets of patterns. Appendix G.3 shows iterates of our map that stop away from every stored pattern.

Setup 3: "Do the authors believe it is an important property, or are they ready to abandon it? Why not focus on simple systems such as gradient flows"

Biological plausibility is outside our scope, and Section 2 now says so (line 133). Our gradient readout is itself a unit gradient step, x−G−1∇E(x)\mathbf{x} - G^{-1}\nabla E(\mathbf{x}), on EE in the metric GG. Section 2 now cites Bartunov et al. (2020), whose memory model retrieves by gradient descent on a learned energy with learned step sizes. We study the discrete map because the published models apply or iterate one, and a descent proof for continuous-time dynamics does not cover it. The universal Hopfield paper, for example, proves descent for continuous-time dynamics, while its code runs a discrete update that has no constant-metric energy at a tested configuration (Section 2, Table 1).

Writing: "I personally find many phrases to be unnatural […] In general, I would insist on improving the organisation of the paper."

We agree with the reviewer's translations, and we rewrote each phrase:

SubmissionRevision
A map without an energy need not cycle.A map without a constant-metric energy need not have cycles. (line 314)
An adaptive score that keeps its energyAn exact energy for the order-statistic score (Section 5)
What the pattern readout admitsWhen the pattern readout has a constant-metric energy (Section 3.1)
To see which maps can cycleTo see when a pattern readout's iterates enter a cycle (line 318)
A monotone gate gives the gradient readout an exact energyWith a monotone gate, the gradient-readout map has an exact energy (Theorem 2)
The energy holds throughout trainingWith a softmax-parameterised gate, the map has an exact energy at every parameter value (Corollary 1)
Where the construction stops.Limits of the construction. (line 405)
Our map pays for its energy in iterations.The guarantee of Theorem 2 comes at a cost in iterations. (line 515)
We expect slow contraction to cause both the gap and the cost.Our map contracts slowly, which likely causes both the accuracy gap and the cost. (line 520)

We also gathered the setup in one place. The first paragraph of the introduction now defines the score, the separation and the inverse temperature in symbols (lines 38 to 46). Section 2 opens with the memory patterns ξμ\boldsymbol{\xi}_\mu, the map TT and its iteration, points to a notation table (Table 4) and says which results assume which readout. Its new paragraph "The stability problem" (line 145) states our question and backs the claim of the introduction with the universal Hopfield case above and the adaptive network, whose proof of its Theorem 2 uses a score without the sort (Appendix E.4).

We thank the reviewer again for these detailed comments, which made the paper clearer. We would be happy to answer any further questions.

AuthorsOpenReview ↗

Follow-up before the discussion closes

Dear Reviewer tyjU,

Thank you again for your detailed comments. Since the discussion period ends on November 18, we would like to check whether the revised PDF addresses your main points. In particular, the main text no longer writes "admissible" without a metric. After Definition 1, it now states that a map without a constant-metric energy may still converge and have a Lyapunov function of another form. We remain available for any further questions.

Best regards, The authors

Reviewer_okz5OpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper investigates when modern Hopfield retrieval maps admit a constant-metric energy function with guaranteed descent. The authors show that, for a class of splitting similarity scores, the conventional pattern readout admits such an energy essentially only for dot-product-like scores. They further establish that a gradient-based readout can support a broader class of semi-convex similarities.

The paper also analyzes 32 existing retrieval maps and identifies eight that do not satisfy the proposed admissibility condition. Based on these findings, the authors modify the Adaptive Hopfield Network using a constrained monotone gate and gradient readout, providing an exact energy guarantee throughout training. Experiments on synthetic retrieval and multiple-instance learning demonstrate the theoretical properties and show that the modified model avoids the period-two oscillations observed in the original adaptive network.

Critical strengths
  1. Systematic theoretical characterization. The paper provides a rigorous analysis of how the interaction between similarity scores and readout mechanisms determines the existence of a constant-metric energy. The characterization of admissible scores and the Jacobian-based criteria offer useful insights into the dynamics of modern Hopfield networks.

  2. Comprehensive theoretical and empirical validation. The analysis of existing retrieval maps, together with the explicit energy construction for Adaptive Hopfield Networks, is interesting. The experiments and ablations also provide useful evidence for understanding the causes of oscillatory behavior.

Critical weaknesses
  1. Unclear practical significance of the energy guarantee. Modern Hopfield networks are often used as single-step retrieval or attention-like layers rather than relying on iterative dynamical evolution. It is therefore unclear how important a strict constant-metric energy guarantee is for their practical applications. Moreover, energy descent alone does not guarantee successful memory retrieval or convergence to a desired memory. The proposed method does not demonstrate clear improvements in retrieval accuracy or memory capacity, while introducing substantially higher computational costs due to additional iterations. This weakens the practical motivation for the proposed framework.

  2. Limited empirical benefit attributable to the gradient readout. The ablation results in Table 3 show that simply constraining the gate to be non-negative already eliminates the observed oscillations while retaining the original pattern readout. Furthermore, the original adaptive model can avoid cycling at a smaller learning rate, as reported in Appendix H.5. These results suggest that the main empirical stability improvement does not necessarily require the proposed gradient readout. While the latter provides a stronger theoretical guarantee, its additional practical benefits remain insufficiently demonstrated.

Questions for authors
  1. Beyond providing a mathematical energy guarantee, what practical advantages does the gradient readout offer over a non-negative gate with the original pattern readout, particularly considering the additional computational cost?

  2. Given that modern Hopfield layers are frequently applied in a single forward pass, which practical applications would substantially benefit from the proposed iterative energy-based formulation? Can the authors identify benefits that cannot be achieved by simpler stabilization strategies?

6 comments
AuthorsOpenReview ↗

Response to Reviewer okz5 (1/2): practical benefit and cost (Q1, W2)

We thank the reviewer for the careful review and for the kind words about our theoretical characterisation. We would like to highlight that we consider the main contribution of the paper to be the characterisation of which Hopfield maps have a constant-metric energy, together with the audit of 32 published maps in Section 4. The modified adaptive network is a worked example of this theory.

Q1 (and W2): "Beyond providing a mathematical energy guarantee, what practical advantages does the gradient readout offer over a non-negative gate with the original pattern readout, particularly considering the additional computational cost?"

Beyond the energy guarantee, we cannot show a practical advantage of the gradient readout. The non-negative gate that the reviewer describes, which we call the clipped gate, is the fix that Proposition 3 points to. Once the softmax saturates, the pattern readout has an attracting orbit of period two between any two patterns that are each other's best match. On squared distances only a negative gate entry can create such a pair, so the proposition predicts that clipping the gate at zero removes these orbits. Three rows of Table 3 show the effect, against 250 of 500 runs with a period-two bag for the squared-distance form of the adaptive network, whose gate is signed:

GateReadoutRuns with a period-two bagBags with a residual below 10−810^{-8} by step 1000Per-seed median iterations to 10−610^{-6}, over the bags that reach itEnergy guarantee
clipped at zeropattern0 of 50099.7 to 100.0%2.0 to 3.4no
monotone: non-negative, non-increasingpattern0 of 50099.9 to 100.0%4.8 to 13.3no
monotone (ours)gradient0 of 50099.9 to 100.0%13.2 to 148.1yes (Theorem 2)

Ranges are over the five datasets, and each run is one seed, fold and dataset, trained through five unrolled steps at learning rate 0.1 and then iterated for up to 1000 steps. Training drives the clipped gate to zero in 159 of its 500 runs. Under the pattern readout a zero gate makes the map return the mean of the bag, and none of the other 341 runs has a period-two bag either. Our cycle test checks period two only. Appendix H.13 therefore takes fits of these three maps on fifteen further seeds, at four learning rates, and follows every held-out bag that has not settled by step 1000 for up to 200,000 steps. The network's published score also has a dot-product base. With that base added and the gate of each base clipped at zero, 15 of 500 runs have a period-two bag, against 243 of 500 without the clipping (Section 6, Appendix H.4). Our guarantee does not extend to that score either. Theorem 2 is stated for squared distances, and on the dot-product base alone our energy has no lower bound (Proposition 5).

What our map adds is the guarantee, as Section 7 says. With the monotone gate, every step of the gradient readout that moves lowers the energy, at every β\beta and every parameter value, so plain iteration has no periodic orbit other than its fixed points (Theorem 2, Corollary 1). Descent does not imply convergence. The clipped gate has no such guarantee, and Proposition 3 explains its effect only for squared distances and a saturated softmax.

We also agree on the cost, which the table gives at learning rate 0.1. The revision adds a majorise-minimise step on the same energy (Proposition 8). Definition 1 does not certify this step, and item C3 of the general response gives its cost. At learning rate 0.1 the clipped gate remains cheaper, at 2.0 to 3.4 iterations, and we claim no advantage in cost.

AuthorsOpenReview ↗

Response to Reviewer okz5 (2/2): the learning rate, single-pass layers and retrieval (W2, Q2, W1)

W2: "the original adaptive model can avoid cycling at a smaller learning rate, as reported in Appendix H.5"

We thank the reviewer for pointing to Appendix H.5, and we agree. In Table 3 we train the squared-distance form of the adaptive network through five unrolled steps at learning rate 0.1, the rate that the network's paper uses for its gate in retrieval, and this form has period-two bags in 250 of 500 runs. At 10−410^{-4} it has no period-two bag. With the temperature, gate stride, gate initialisation and bases of the network's released code restored, the network at the same training depth has period-two bags in 189 of 500 runs at 0.1 (the last row of Table 3), and at 10−410^{-4} only on MUSK1, in 0.7% of MUSK1's held-out bags (Appendix H.6, which was H.5 in the submission). Theorem 2, in contrast, holds at every parameter value (Corollary 1), so it does not depend on the learning rate.

Q2 (and W1): "Given that modern Hopfield layers are frequently applied in a single forward pass, which practical applications would substantially benefit from the proposed iterative energy-based formulation? Can the authors identify benefits that cannot be achieved by simpler stabilization strategies?"

We agree that modern Hopfield layers are frequently applied in a single forward pass, and the adaptive network's released code also runs its map for one step. We cannot name an application that needs our iterated map rather than a simpler change such as clipping the gate at zero. The theory itself has two uses that do not require iterating our map.

First, Definition 1 bounds every single step, so a check mark in Table 1 also makes one application of the map a descent step. For example, Table 1 classifies these two layers:

LayerConstant-metric energyReason
one-step attention with a value readout (Vaswani et al., 2017)✗The score uses the keys but the readout uses the values.
the Hopfield layer released with Ramsauer et al. (2021)✓Its loop uses the keys in both the score and the readout. The values are applied once, after the loop.

The verdicts say nothing about retrieval, and a cross does not mean that a layer is unstable.

Second, the best-match graph of Proposition 3 identifies, before any iteration, the period-two bags of the adaptive network. For the squared-distance form trained through five unrolled steps at learning rate 0.1 and run at its trained temperature, a walk that follows best matches from the pattern that scores highest at the query ends in a mutual pair for 6174 of the 6176 period-two bags at step 1000. Of the 6397 bags whose walk ends in a pair, 6174 are period-two bags (Section 6).

W1: "energy descent alone does not guarantee successful memory retrieval or convergence to a desired memory"

We agree. Descent ensures neither convergence nor retrieval, and Appendix G.3 shows iterates of our map that stop away from every stored pattern.

W1: "The proposed method does not demonstrate clear improvements in retrieval accuracy or memory capacity, while introducing substantially higher computational costs due to additional iterations"

We agree, and we claim neither improvement. On the capacity grid of Appendix G.3, the readout changes none of the retrieval capacities at the initial gate. Our answer to Q1 above gives the cost.

Items C3 and C5 of the general response also summarise the cost and the single-step reading of Table 1. We hope that these answers help to clarify what the energy guarantee does and does not provide, and we would be happy to answer any further questions.

AuthorsOpenReview ↗

Follow-up before the discussion closes

Dear Reviewer okz5,

Thank you again for your careful reading. With the discussion period ending on November 18, we would like to check whether our reply answers your two questions. They concern the practical advantage over a non-negative gate with the pattern readout, and modern Hopfield layers that run a single forward pass. If anything remains unclear, we are happy to clarify.

Best regards, The authors

Reviewer_34ttOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies when a modern Hopfield retrieval map score → separation → readout has a concave–convex energy in a constant metric (Definition 1: there exist a constant SPD matrix G and a convex function h such that G T(x) ∈ ∂h(x)). The authors argue that, under the Euclidean metric, splitting scores, and patterns in general position, the scores that are admissible for pattern readout (a weighted average of stored patterns) lie on the half-flat of the dot product (including negative squared distance); gradient readout is sufficient for scores that are convex up to a quadratic term. They classify 32 published maps on this basis, and they change the Adaptive Hopfield order-statistic squared-distance score to a monotone nonnegative gate with gradient readout, so that there is an exact energy throughout training. The experiments mainly examine descent, cycling, and gradient fidelity on synthetic retrieval and five MIL datasets; the authors explicitly do not claim an accuracy advantage.

Critical strengths
  1. The theoretical question is clean and falsifiable. Definition 1 reduces “having an energy” to one concave–convex step in a constant metric; Theorem 1 characterizes the half-flat of splitting scores under pattern readout. This is a substantive clarification of the energy arguments in Santos/Millidge/Wang. The theoretical corrections are fairly solid, and this is the part of the paper I find most valuable.
  2. The experiments are clearly self-limited. The authors do not claim SOTA accuracy; Table 3 separately varies the sign of the gate and the readout, and states that a nonnegative gate alone already stops cycling. The preregistration and Holm-corrected experiments are also relatively complete.
Critical weaknesses
  1. The main-text experiments bind Adaptive cycling too tightly to “our energy construction fixed it,” while the key qualifications are left in the appendix. Cycling is driven mainly by a negative gate together with mutual best-match pairs under saturated softmax; pattern readout with the gate clipped at ≥0 already yields 0/500 cycling runs. The 250/500 figure is still at lr=0.1; under the same implementation at 10^{-4}, the squared-distance form of Adaptive can have zero cycling and a higher AUC. So although the theory is fairly clean, it has not so far been translated into MIL stability or retrieval quality in the experiments. Contribution 4 and the main-text narrative make it easy to read an overly strong empirical conclusion first.
  2. The novelty lies mainly in organizing and cleaning up the theory, not in proposing an efficient or beneficial modification. The gradient-readout construction that best matches Definition 1 is worse and slower in one step (on synthetic retrieval with mask 0.7 it trails by about 28 points at @1, and only comes close after about 16×/query). With the initialized gate and β=10 it can even fail to retrieve at all. Santos already noted the restriction to the dot product, and the Krotov tutorial already gave the gradient-update template. As it stands this is more like a solid theoretical correction plus a design variant that did not yield a beneficial translation; I still find the contribution somewhat thin.
Questions for authors
  1. Why does the main protocol use the gate learning rate 0.1 from Adaptive’s retrieval experiments, rather than its MIL recipe or the learning-rate grid of Hu et al. (2023)? Under the same implementation at 10^{-4}, the squared-distance form of Adaptive can have zero cycling and a higher AUC. Please state under which training settings the 250/500 figure still holds;
  2. Please report the distribution of β after MIL training. Proposition 3 requires a large β; currently one only sees that the largest softmax weight at the last step has median 1.000, not β itself.
  3. Please narrow the juxtaposition of cycling in Contribution 4 and the abstract so that the main text matches Table 3: what stops cycling is a nonnegative gate, and the energy adds a descent guarantee; the main text is easy to read as a practical fix for Adaptive, as "they show how to test and design admissible retrieval maps". In Table 3, a gate clipped to be nonnegative, still with pattern readout, already gives 0/500 cycling runs. Relative to this control, what is the incremental claim of gradient readout / an exact energy on MIL? If there is no gain in stability or task performance, the experiments cannot support “one must change the readout.” Or rather, can the authors translate the theory into actuall benefits more successfully? That is what I would more like to see.
4 comments
AuthorsOpenReview ↗

Response to Reviewer 34tt (1/2): the clipped gate, what our map adds, and the learning rate (W1, Q3, Q1)

We thank the reviewer for the careful review, and for finding the theoretical question clean and falsifiable and the theory the most valuable part of the paper. Below, we address the reviewer's questions and concerns. We consider the main contribution of the paper to be the characterisation of which Hopfield maps have a constant-metric energy, together with the audit of 32 published maps in Section 4. The modified adaptive network is a worked example of this theory.

Q3 (and W1): "what stops cycling is a nonnegative gate, and the energy adds a descent guarantee"

We agree. The submitted abstract reported this after our construction. The revised abstract gives the mechanism first, then the fix for squared distances ("clipping these weights at zero removes the period-two cycles of the pattern readout in our experiments", lines 30 to 31), and only then our construction. Contribution 4 now states the same fix, as Proposition 3 predicts for a saturated softmax, and the closing clause of the abstract that the reviewer quoted is deleted.

Q3: "Relative to this control, what is the incremental claim of gradient readout / an exact energy on MIL?"

On the period-two cycles there is none. We agree that our experiments do not show that the readout must be changed. What our map adds is the guarantee: with the monotone gate, every step of the gradient readout that moves lowers the energy, at every β\beta and every parameter value, so plain iteration has no periodic orbit other than its fixed points (Theorem 2, Corollary 1). Section 7 now says so, and states that descent ensures neither convergence nor retrieval. The clipped gate has no such guarantee, and Proposition 3 explains its effect only for squared distances and a saturated softmax. Four rows of Table 3, all on squared distances at learning rate 0.1:

GateReadoutRuns with a period-two bagBags with a residual below 10−810^{-8} by step 1000Energy guarantee
signed (adaptive network)pattern250 of 50048.3 to 61.8%no
clipped at zeropattern0 of 50099.7 to 100.0%no
monotonepattern0 of 50099.9 to 100.0%no
monotone (ours)gradient0 of 50099.9 to 100.0%yes (Theorem 2)

Ranges are over the five datasets, and each run is one seed, fold and dataset, trained through five unrolled steps at learning rate 0.1 and then iterated for up to 1000 steps. Training drives the clipped gate to zero in 159 of its 500 runs. Under the pattern readout a zero gate makes the map return the bag mean, and none of the other 341 runs has a period-two bag either. The network's published score also has a dot-product base. With that base added and the gate of each base clipped at zero, 15 of 500 runs have a period-two bag, against 243 of 500 without the clipping (Section 6, Appendix H.4). For this network's sorted and gated score, part (iv) of Proposition 3 covers the squared-distance base only, and our construction does not cover the score with both bases either. Theorem 2 is stated for squared distances, and on the dot-product base alone our energy has no lower bound (Proposition 5). At this learning rate the clipped gate needs fewer iterations than our map, and we claim no advantage over it in cost (item C3 of the general response).

Q1: "Please state under which training settings the 250/500 figure still holds"

It holds at learning rate 0.1 on the Hopfield parameters, the rate stated in Section 6 and in the captions of Tables 2 and 3, with the training depth of five unrolled steps given above. We used 0.1, the rate that the adaptive network's paper uses for its gate in retrieval, because the released code has no multiple-instance setup. The last paragraph below gives the result under that paper's multiple-instance recipe. In separate runs over five seeds (Table 23), the share of held-out bags of its squared-distance form with period two falls with the rate (ranges over the five datasets):

Learning rateShare of held-out bags with period two
0.10.429 to 0.557
10−210^{-2}0.295 to 0.452
10−310^{-3}0 to 0.110
10−410^{-4}0

With the released code's temperature, gate stride, gate initialisation at zero and bases restored, the network has period-two bags in 189 of 500 runs at 0.1 (Table 3), and at 10−410^{-4} only on MUSK1, in 0.7% of its held-out bags. Under the multiple-instance recipe of the adaptive network's paper, taken from Hu et al. and run in our pipeline, the squared-distance form still has period-two bags in 4 to 24% of held-out bags, and the released configuration only on MUSK1, in 5.4%. This is not a reproduction of their results (Appendix H.6).

AuthorsOpenReview ↗

Response to Reviewer 34tt (2/2): the trained β (Q2), novelty, practical benefit and β = 10 (W2)

Q2: "Please report the distribution of β after MIL training."

Table 27 and Figure 10 (Appendix H.12) now give it, for the 100 fits per dataset trained through five unrolled steps at learning rate 0.1. The median β\beta is 26.6 to 47.6 for the adaptive network.

Proposition 3 holds for every large enough β\beta and states no threshold, so we measure saturation instead. In the adaptive network's fits of row 6 of Table 2, the largest softmax weight at the last step has a median of 1.000 in period-two and settled bags alike (Appendix H.3). The trained β\beta does not separate the fits with a period-two bag: their median is higher than the other fits' on four datasets and lower on UCSB. A walk along best matches identifies the period-two bags before any iteration. For its squared-distance form trained through five unrolled steps at rate 0.1 and run at its trained β\beta, the walk from the pattern that scores highest at the query ends in a mutual pair for 6174 of the 6176 period-two bags at step 1000, and 6174 of the 6397 bags whose walk ends in a pair are period-two bags (Appendix H.3).

W2 (novelty): "Santos already noted the restriction to the dot product, and the Krotov tutorial already gave the gradient-update template."

We agree with both points, and the paper credits both, after Theorem 1 and after Proposition 1. What each result adds:

ResultWhat it adds
Theorem 1The restriction as a theorem: for splitting scores and patterns in general position, only a half-flat of scores that contains the dot product gives the pattern readout a constant-metric energy with G=IG = I at all softmax weights
Proposition 1A sufficient condition on the score, convexity up to an added quadratic, under which the gradient readout has a constant-metric energy
Proposition 2A Jacobian test that also excludes a Fenchel-Young step with a twice-differentiable conjugate. Each cross in Table 1 rests on it and on Lemma 1
Proposition 3Why a saturated pattern readout has period-two cycles: at two patterns that are each other's best match
Theorem 2, Corollary 1, Proposition 6An exact energy for a learned, non-smooth score at every parameter value, and, for patterns in general position, why its gate must be non-increasing

W2, Q3: "can the authors translate the theory into actuall benefits more successfully?"

In part. The theory gives three practical results, none of them an accuracy or speed advantage of our map.

  1. Proposition 3 gives the mechanism behind the fix in the reviewer's control. Once the softmax saturates, the pattern readout has an attracting period-two orbit between any two patterns that are each other's best match, and on squared distances only a negative gate entry can create such a pair. Clipping the gate at zero removes these cycles in our runs on squared distances (Table 3).
  2. Theorem 2 gives an energy at every parameter value. The new Proposition 8 gives a cheaper descent step on it, a step that Definition 1 does not certify, at a per-seed median of 3.7 to 8.5 evaluations to a residual of 10−610^{-6}, in place of 13.2 to 148.1 for plain iteration, on the fits trained at learning rate 0.1 (Table 26). It needs fewer evaluations than plain iteration in 88% of bags on Tiger and in at least 91% on the other datasets.
  3. Tables 1 and 8 audit 32 published maps for a constant-metric energy. We find 8 that have none at a tested configuration and no energy proved for them in their own papers. For four of the 8 the published proof covers a different map. A map without such an energy need not have cycles.

We also agree with the reviewer's reading of the synthetic results in Section 7: plain iteration of our map is slow, and we claim no accuracy or speed advantage there.

W2: "With the initialized gate and β=10 it can even fail to retrieve at all."

We agree, and the same holds under the pattern readout. At β=10\beta = 10 the initial gate retrieves no query under either readout, while the uniform gate retrieves every query with a tenth of the coordinates masked. At β=30\beta = 30 and 100 our map with the initial gate retrieves at least 95% of queries with a tenth masked, for up to 512 stored patterns in dimension 64 (Appendix G.3).

We hope that these answers clarify what the theory and the energy guarantee do and do not provide, and we would be happy to answer any further questions.

I have read the rebuttal and new appendix materials. The authors have addressed my concrete technical points: they report learned β statistics, show how cycling depends on learning rate, tone down over‑strong empirical claims, and properly situate their work against prior literature. It is now clear that simple clipped‑gate pattern readout removes period‑2 cycles at low cost, without gradient readout. I think the G‑admissibility framework, Theorem 1 characterization, and audit of existing Hopfield maps are real theoretical contributions for this community and require a lot of work. These tools can help future work check energy‑related claims for iterative retrieval dynamics. Even so, my score remains 2. My main concern is still the impact justification. The paper surfaces a core trade‑off in this area: gradient‑readout gives the formal constant‑metric energy certificate but is computationally heavy; clipped‑gate works well empirically yet lacks this guarantee. The manuscript does not convincingly argue when the G‑admissible certificate is irreplaceable. The passing mention of implicit training is too brief. Given a cheap heuristic already fixes the observed cycling problem, it remains unclear why practitioners or theorists must care about this formal property, and it might already happen to be the engineering trade-off. This makes the overall contribution feel somewhat thin. For a future advice, strengthen along at least one direction below may be helpful: Add a dedicated discussion section explaining concrete scenarios where the G‑admissible certificate is required for analysis. For example, for implicit differentiation, fixed‑point training, or robustness proofs on discrete Hopfield dynamics: under what conditions do existing arguments strictly need such a Lyapunov function? What analytical barriers appear when you only have an empirical heuristic like clipped‑gate without this formal guarantee? This should be more than a one‑sentence note. Find a small empirical setting showing unique benefits from G‑admissibility. Large MIL accuracy gains are not required. Target measurable advantages that clipped‑gate cannot provide: e.g. better solver success for Anderson/Broyden in fixed‑point training, improved gradient quality under implicit differentiation, or measurable differences in dynamical orbit behavior. Keep all existing revisions: learning‑rate dependence of cycling results, β distribution, and the separation between empirical gate effects versus gradient‑readout theoretical properties. In short, this is theoretically solid work with useful analytical tools, but it needs better justification for why the G‑admissible property matters for follow‑up research.

AuthorsOpenReview ↗

Thank you for the follow-up

We thank the reviewer for reading our rebuttal and the new appendix material, and for recognising the G-admissibility framework, Theorem 1 and the audit of published Hopfield maps as theoretical contributions. We agree that the question of when the G-admissible certificate is needed is an important one. The reviewer suggests two directions: a discussion of the analyses that need such a Lyapunov function, and a small setting that looks for benefits the clipped gate cannot provide. Both are concrete, and they will guide our follow-up work. The revisions the reviewer mentions will remain in the paper. We thank the reviewer again for the time and the constructive feedback.

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

The paper provides a definition for a well-behaved Hopfield map, which they call admissibility. They show the conditions under which a Hopfield map is admissible, which largely depend on the readout. They then use this property to analyze previously published Hopfield networks and show how to use admissibility to create a well behaved Hopfield map.

Critical strengths

Provides useful, specific criteria researchers can use in future work to guarantee well-behaved Hopfield networks.

Provides retrospective analysis for a broad set of papers in the field.

Offers thorough mathematical justification.

Critical weaknesses

Result Their proposed alteration incurs a heavy computational cost, making it potentially impractical.

Writing The presentation is a little hard to follow. Some of the poor writing appears to reflect the style of writing produced by LLMs: untethered disclaimers, mindless listing of results, and an organizational style that is readable only if you have an infinite-length memory cache (or context window).

Sections 6 and 7 are the most egregious so I will focus my critique with that portion. It should be unambiguously clear what the claims are, what the experiments are, how those experiments test the claims, and finally, whether the experimental results support the claims. It is not necessary to recite every last result.

Instead the section starts out with “The experiments test the predictions of Sections 4 and 5, and Table 2 collects them with two further tests.” First, this sentence should be split in two. Second, “them” is ambiguous. Third, this is an example of what I mean by writing for an infinite-length memory cache. It would be very helpful to list out those claims concisely here. It could be fine to say, “We support the claims from sections 4 and 5, which we enumerate in table 2”. This would show the reader exactly where to look, instead of having to hunt through the prior two sections of the paper.

I see the details of the two tests in the setup section, but it is not obvious what claims they test and why they do so effectively.

“All six predictions hold.” This is when I found out there were six predictions. This section reads like a mindless list of results. There are too many untethered disclaimers whose necessity and effect on the validity of your results are unexplained. Details about the results can be left in the tables or moved to the appendix. The question of how your test supports your claims is still unanswered.

“The gate’s sign alone stops the cycling.” Everything after “A non-negative gate” is way too compact and would be better served expanding on in the appendix.

“The cycling depends on the learning rate, not the base”. What is this paragraph for? Why start this paragraph with a disclaimer? This paragraph contains far too many ambiguous instances of “them” and “it”.

Additionally, titles and subtitles of paragraphs and theorems are overly compact, don't make sense out of context, and use unnatural LLMisms like "Its x, not y".

Questions for authors

Q1 How do your experiments support your claims? I think your experimental section should be rewritten and should ideally include the answer to this question.

Q2 “The gate’s sign alone stops the cycling.” Does this statement not contradict your main claim that the readout decides behavior? Same for "The cycling depends on the learning rate, not the base".

Q3 “The cycling depends on the learning rate, not the base.” Why do you think the learning rate matters here?

8 comments
AuthorsOpenReview ↗

Response to Reviewer XnXA (1/2): Section 6 rebuilt around what each experiment tests (Q1, Q2, the writing)

We thank the reviewer for the detailed comments, and for finding our criteria useful for future work and our mathematical justification thorough. We rewrote Sections 6 and 7 along these comments (pages 8 to 10 of the revised PDF, in blue).

Q1: "How do your experiments support your claims?"

Table 2 pairs each prediction with the statistic that tests it, and Section 6 now opens by saying what each row tests. Rows 1 and 2 test Theorem 2 on our trained maps: if its hypotheses hold, no step raises the energy and no bag has a cycle. Rows 3 and 4 test a prediction that our proofs imply only in part and only for large β\beta: a period-two cycle of the adaptive network lies at a mutual best-match pair, two patterns that are each other's best match. On squared distances, only a negative gate entry can create such a pair. Rows 5 and 6 measure two consequences of a cycle.

RowStatisticBaselineOurs
1Synthetic regimes where the energy rises at some step24 of 270 of 27
2Training runs with a held-out period-two bag250 of 5000 of 500
3Runs with a non-negative trained gate that have a period-two bag0 of 142n/a
4Period-two bags whose orbit lies at a mutual best-match pair97 to 100%n/a

The baseline is our score and gate under the pattern readout in row 1, and in rows 2 to 4 the adaptive network, trained through five unrolled steps at learning rate 0.1. The non-zero baseline counts in rows 1 and 2 show that both tests can detect a failure. The trained β\beta is finite, so rows 3 and 4 agree with Proposition 3 but do not verify its hypotheses.

The writing. We revised each passage that the reviewer quoted:

  • "All six predictions hold" is now three paragraphs, one per pair of rows of Table 2, and each opens with its verdict.
  • The passage after "A non-negative gate" is now "Does the gate or the readout remove the period-two orbits?" (line 473), and its first sentence gives the answer.
  • "The cycling depends on the learning rate, not the base" is now "Do the adaptive network's cycles come from our setup?" (line 482), conclusion first.
  • Titles: 22 in Sections 2 to 7 are reworded, replaced or removed, and none has the form "x, not y".
  • In Section 6, in the five paragraphs after the setup, each caveat now follows the verdict that it limits.
  • Each section of Appendices G and H now opens with the statement of the main text that it supports.

Q2: "'The gate's sign alone stops the cycling.' Does this statement not contradict your main claim that the readout decides behavior? Same for 'The cycling depends on the learning rate, not the base'."

No. The main claim is a guarantee: the readout restricts which scores give a map a constant-metric energy, which no step raises (Theorem 1, Corollary 2, Proposition 1). The two quoted sentences were observations about our runs, and our old wording ran the two together. Both titles are gone, and the abstract and introduction now say "restricts" in place of "decides". Under the pattern readout on squared distances, the sign of the gate is what matters in our runs: both non-negative gates remove the period-two cycles, as Proposition 3 predicts for a saturated softmax. At learning rate 0.1, training drives the clipped gate to zero in 159 of its 500 runs. Under the pattern readout a zero gate makes the map return the bag mean, and none of the other 341 runs has a period-two bag. Changing only the readout makes no map better behaved in our runs (Table 3). With our monotone gate, the gradient readout adds the guarantee: of the maps in Table 3, only ours has a proved constant-metric energy.

AuthorsOpenReview ↗

Response to Reviewer XnXA (2/2): the learning rate (Q3) and the cost

Q3: "The cycling depends on the learning rate, not the base." Why do you think the learning rate matters here?

We think that the rate acts through the sign of the trained gate, although our runs do not show that a negative entry causes a period-two bag. For the squared-distance form of the adaptive network trained through five unrolled steps, of 125 runs per rate:

Learning rateRuns that end with a negative gate entryRuns with a period-two bag
0.19467
0.019453
10−310^{-3}678
10−410^{-4}10

Every run with a period-two bag has a negative gate entry, but at 10−310^{-3}, 59 of the 67 runs with a negative entry have none. The rate also moves the temperature, so these counts do not isolate the sign.

Rows 2 to 6 of Table 2 use 0.1, the rate that the adaptive network's paper uses for its gate in retrieval. In our pipeline, with the released code's temperature, gate stride, gate initialisation at zero and bases restored, the network trained through five unrolled steps has period-two bags in 189 of 500 runs at 0.1 (Table 3), and at 10−410^{-4} only on MUSK1, in 0.7% of its held-out bags (Appendix H.6).

Weakness: "Their proposed alteration incurs a heavy computational cost, making it potentially impractical."

We agree that plain iteration of our map is costly, and Section 7 opens with this cost. In our runs on squared distances, clipping the gate at zero removes the period-two cycles more cheaply. Unlike the clipped gate, our map has a guarantee: every step that moves lowers the energy.

The revision adds a cheaper majorise-minimise (MM) step on the same energy (Proposition 8, Appendix C.6). With our gate, every MM step that moves lowers the energy, so the step has no periodic orbit other than its fixed points. Its bound uses a metric that varies with the state, so Definition 1 does not certify it. We apply it at inference to maps trained through five unrolled steps of plain iteration. The table gives the evaluations of the map until the residual falls below 10−610^{-6}, as a per-seed median over the bags that reach it, with ranges over the five datasets at learning rate 0.1 (Table 3 and Appendix H.10):

MapEvaluationsBags reaching the tolerance
Adaptive network, squared-distance form2.0 to 3.148 to 62%
Clipped gate, pattern readout2.0 to 3.4at least 99.7%
Ours, plain iteration13.2 to 148.1100%
Ours, MM step3.7 to 8.5100%

The MM step needs fewer evaluations than plain iteration in 88% of bags on Tiger and in 91 to 99.9% on the other four datasets. The clipped gate needs fewer evaluations than the MM step, and we claim no advantage over it in cost.

We hope that the rewritten Sections 6 and 7 are now easier to follow, and we would gladly revise any passage that remains unclear. We would be happy to answer any further questions.

AuthorsOpenReview ↗

Follow-up before the discussion closes

Dear Reviewer XnXA,

Thank you again for your review. As the discussion period ends on November 18, we would like to check whether our response and the revised PDF address your main concerns. In particular, Section 6 now opens by saying what each experiment tests, and it no longer uses the two paragraph titles you quoted in Q2. The abstract and introduction now say "restricts" in place of "decides". We would be glad to answer any further questions before the discussion closes.

Best regards, The authors

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response: shared concerns and the list of changes

We thank the four reviewers for careful and specific reviews, and for finding the theoretical question "clean and falsifiable" (34tt) and the analysis rigorous (okz5, tyjU). A revised PDF is uploaded, with added and reworded passages in blue. The submitted results keep their mathematical content, and one result is new (Proposition 8).

ConcernRaised byShort answer
C1Terms and titletyjUThe main text now says that a map has, or has no, constant-metric energy, and that a map without one may still converge. The abstract says "restricts" in place of "decides". We keep the title during the review, and our reply to tyjU proposes a narrower one
C2What stops the period-two cycles, and what the energy addsXnXA, 34tt, okz5Under the pattern readout on squared distances, a non-negative gate does in our runs (the adaptive network's gate clipped at zero), as Proposition 3 predicts for a saturated softmax. Our map adds a guarantee that every step that moves lowers the energy
C3Cost of our mapXnXA, 34tt, okz5We agree. A new step on the same energy needs fewer evaluations, and the clipped gate stays cheaper
C4PresentationXnXA, 34tt, tyjUSection 2 now states the setup and the stability problem, and Section 6 says what each experiment tests
C5Layers that apply the map onceokz5, tyjUDefinition 1 bounds every single step, so a check mark in Table 1 also makes one application a descent step

C3. Cost of our map (XnXA, 34tt, okz5). We agree that plain iteration of our map is slow. The revision adds a majorise-minimise (MM) step on the same energy (Proposition 8, Appendix C.6), applied at inference to maps trained through five unrolled steps of plain iteration. Its bound uses a metric that varies with the state, so Definition 1 does not certify it. The table gives the evaluations until the residual falls below 10−610^{-6}, as a per-seed median over the bags that reach it, with ranges over the five datasets. Every map is trained through five unrolled steps at learning rate 0.1 (Tables 3 and 26):

MapEvaluationsNote
Adaptive network, squared-distance form2.0 to 3.1only 48 to 62% of its bags converge
Clipped gate, pattern readout2.0 to 3.499.7 to 100% converge
Ours, plain iteration13.2 to 148.1
Ours, MM step3.7 to 8.5every bag reaches the tolerance

The MM step needs fewer evaluations than plain iteration in 88% of bags on Tiger and in at least 91% on the other datasets. The clipped gate stays cheaper, and we claim no advantage over it in cost. Training drives the clipped gate to zero in 159 of its 500 runs, where the map returns the bag mean, and none of the other 341 runs has a period-two bag either (Table 3).

Main changes.

  • Main text: the abstract, the introduction, Contribution 4, Section 2 (setup, stability problem), what a cross in Table 1 means, Section 6 (what each row of Table 2 tests, a new row in Table 3) and Section 7 (cost, the MM step, what the guarantee adds over the clipped gate). Titles now say what their part shows, and the side label of the middle panel of Figure 1 is reworded.
  • Table 2: row 5 now covers Fox, Tiger and Elephant (35 to 43%, was 35 to 47% with MUSK1). Row 6 now gives the squared-distance form (9 to 26%), and the caption gives 9 to 27% for both bases.
  • Appendix, new results: C.6 (Proposition 8), H.5 (clipped gate, gradient readout), H.12 (trained β\beta), H.13 (bags followed past step 1000), the best-match walk in H.3, the clipped gate on rows 5 and 6 in H.4, the adaptive network with both bases and the gate of each base clipped at zero in H.4 (15 of 500 runs have a period-two bag, against 243 of 500 without the clipping and 0 of 500 for the clipped gate under the pattern readout on squared distances alone), the AUC at each map's own rate in H.6, the MM step in H.10 and Table 12, and the training time in H.11.
  • Appendix: each section of Appendices G and H now opens with the statement of the main text that it supports. We removed material on which no statement of the main text rests (a repair recipe, a run on MNIST bags, the forward-solver section, several tables and figures and the narration of our preregistered protocols), so later tables and sections are renumbered. The Holm-corrected tests remain (Appendix H.1).

We thank the reviewers again, and we would be happy to answer any further questions.

AuthorsOpenReview ↗

Thank you to the reviewers, and the changes we will make in the final version

Dear Area Chair and Reviewers,

We thank the four reviewers for their time and for their detailed comments throughout the discussion. In the PDF that we uploaded on November 12, we revised the abstract, the introduction and Sections 2, 6 and 7, and we added one result (Proposition 8).

What the reviewers helped us clarify

  • Reviewer XnXA asked how our experiments support our claims. We rebuilt Section 6 around what each experiment tests. The reviewer also found that the prose is still hard to read, so on November 18, the reviewer sent detailed feedback on the writing. In reply to Reviewer XnXA's question on Q2, we stated which condition on the readout is necessary for a constant-metric energy, and which two conditions together are necessary and sufficient.
  • Reviewer tyjU found that framing the work as an extension of Ramsauer et al. (2021) clarifies our motivation and the form of the energy. In Section 2 of the revised PDF, we state that the energy of Ramsauer et al. has the form of Definition 1 with G=IG = I.
  • Reviewer 34tt helped us separate what the gate does from what the readout guarantees. Under the pattern readout on squared distances, a non-negative gate stops the period-two cycles in our runs. With our monotone gate, the gradient readout adds a guarantee that every step that moves lowers the energy. The reviewer's main concern is that the paper does not show when this guarantee is needed. Hence, Reviewer 34tt suggested two directions: a discussion of the analyses that require a constant-metric energy, and a small experiment that tests for benefits that the pattern readout with a non-negative gate cannot provide.
  • Reviewers XnXA, 34tt and okz5 pointed to the cost of our map. We added a majorise-minimise step on the same energy (Proposition 8). This step needs fewer evaluations than plain iteration of our map. Reviewer okz5 then asked what makes a constant-metric energy a meaningful property, and what our characterisation provides beyond a general Lyapunov analysis.

Changes for the final version

ChangeRaised by
Revise the prose of the whole paper for the patterns that Reviewer XnXA named, such as ambiguous pronouns and subjects far from their verbsXnXA
Replace the sentences that contain "restricts" in the abstract and the introduction with statements of the form "x is a necessary condition for y"XnXA
Narrow the title so that it matches the proved results, if the camera-ready instructions allow the changetyjU
Add the comparison with general Lyapunov analyses to the discussion of Definition 1okz5

We will also keep the revisions that Reviewer 34tt asked us to keep: the dependence of the cycling on the learning rate, the distribution of the trained β\beta, and the separation of the gate's effect from the readout's guarantee. The two directions that Reviewer 34tt suggested will guide our follow-up work.

In conclusion, the four reviewers encouraged us to make the presentation more readable, so that the work can have more impact. We will prepare the final camera-ready version with readability as the goal.

We thank the Area Chair and the reviewers again for your time and guidance.

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.