acceptodds
Under review as a conference paper at ICLR 2027

Dynamical Memory Retrieval in Attractor Networks as a Denoising Diffusion Process

Abstract

Successful memory retrieval requires the network state to fall into the "attractor basin" of the target memory, which presents a challenge in setting the basin’s size: a basin too broad leads to imprecise retrieval, while a basin too narrow fails to capture the network state. Here, we propose that the neural system can solve this problem by dynamically adjusting the basin’s size, starting wide and gradually narrowing. We use continuous attractor neural networks (CANNs) as the memory model, where the width of the neural activity bump shrinks over time, thereby narrowing the attractor basin. Intriguingly, we find that this dynamical memory retrieval process is equivalent to the reverse dynamics of denoising diffusion probabilistic models (DDPMs), where the decrease in noise variance effectively corresponds to the contraction of the basin’s size in CANNs. We carry out experiments demonstrating the advantages of this dynamical retrieval process; and for hierarchical memories, it gives rise to information retrieval from coarse to fine. We hope this study provides insight into dynamical memory retrieval in the neural system and its link to diffusion generative models in AI.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Continuous attractor neural networks, Denoising diffusion probabilistic models, Memory retrieval, Attractor dynamics, Coarse-to-fine retrieval, Neural dynamics

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_MH4HOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper “Dynamical memory retrieval in attractor networks as a denoising diffusion process” analyzes a memory retrieval process based on the temperature / noise annealing process that is typical of diffusion models (DMs). Despite several studies in the literature have already investigated the formal analogies between modern hopfield networks and DMs (e.g. Hoover, 2023; Raya & Ambrogioni, 2024), no-one had clearly studied the noise reduction schedule that is typical of memorizing DMs in the context of associative memories. This work first establishes a formal equivalence between a class of continuous attractor neural networks (CANNs) and variance exploding DMs, then applies the annealing routine in real experiments.

Critical strengths

The analogy between CANNs and DMs is indeed interesting. This type of associative memory is very much adopted by theoretical neuroscientists to study spatial memory recall, and this study might in practice foster biological implementations of diffusion models. The paper is simply written and instructive, the story line is clear.

Critical weaknesses

My impression as a Reviewer is that, once the analogy between CANNs and DMs is established, the rest is equivalent to memorization in DMs, which is a broadly studied phenomenon that does not necessarily need a new demonstration, as the one in Section 4. I think that the paper would be effectively novel and fundamental if a more biologically plausible model was employed to parametrize the CANN. This would have answered a question sounding similar to “Does the Hippocampus learn a diffusion model?”.

Questions for authors

I will now present my comments and questions with the goal of improving the quality of the presented work. I ask the Authors to address my points extensively.

Q1a: The gaussian mixture expression of the mean bump position x(λ)x(\lambda) depends on the structure of I_ext but also on the fact that J is chosen to be symmetric Gaussian. Then authors decide to choose the width of the interaction a to vary in time. Why so? Isn't it more biologically accurate to assume the interactions among place cells to be quenched and dependent on the data (i.e. aa fixed) and then imagine the memory retrieval to be driven by a time dependent external stimulus that regulates the temperature of the neural dynamics ?

Q1b: what would be the interpretation, in terms of neural activity, of a decreasing bump width? Would this be observed from physiological data, where neural dynamics is usually captured at the equilibrium?

Q1c: The choice of IextI_{ext} is very specific. What is its physical / biological interpretation ?

Q2: Works such as Battista & Monasson (2020) or Rosay & Monasson (2013) are based on graphical models of neurons, as in more classic Hopfield-like associative memories. On the other hand, dense associative memories (Krotov & Hopfield, 2017) do not have a direct interactive counterpart. Can you imagine a mapping between a neural network and the CANN studied in this work? Since we still have not discovered any diffusion model in the brain, if I wanted to actually measure whether a biological neural network is trained in the same way as generative diffusion, how would I do it ? This question is already treated in the conclusions but I would like the authors to expand on this.

Q3: Deep networks that parametrize diffusion models are usually trained by denoising score matching. After training, when the model generates a configuration, even one that coincides with a training point if the model is in the memorization regime as it is in the paper, the training points are not seen again. Conversely, each memory retrieval “annealing scheme” as described in paper relies on presenting all training points to the model, which is costful, and possibly not very biologically plausible. I invite the Authors to discuss this computational issue.

Q4: Since CANNs were involved I asked many questions about biological plausibility, but I realize that Authors have chosen CANNs as a memory model to demonstrate the true efficacy of this annealing retrieval routine. I hence invite them to discuss the novelty of this work with respect to memorization in DMs.

Q5: Authors should definitely cite Lucibello & Mézard (2024). "Exponential capacity of dense associative memories". This seminal paper shows that dense hopfield with spherical constraints have infinite capacity, and this is shown by Raya & Ambrogioni (2024) to be equivalent to a diffusion model in the variance exploding set-up which is also the dynamic setting of the analyzed CANN. More generally I feel like the paper lacks citations both in the matter of modern hopfield networks and the very few biological implementations (see Krotov & Hopfield, 2023; Kafraj et al., 2025). The paper also lacks references regarding memorization in diffusion models, e.g. Biroli & Mézard (2024), Raya & Ambrogioni (2024).

5 comments

We thank the reviewer for the comments and questions. First of all, we would like to clarify our modelling background. Our CANN is modelling the activities of grid cells in MEC, where recurrent connections are considered to be homogeneous. (There are substantial experimental evidence suggesting that grid cell populations in the MEC form an internally maintained continuous manifold [1], where the recurrent connection JJ is required to be homogeneous to form this continuous attractor.) Information about the memories is transmitted from HPC to MEC, and this is the biological interpretation of the external inputs IextI_{ext}. The bump in the MEC travels along the attractor landscape under the attraction of these inputs and carries out the retrieval process.

It is better to interpret the decrease in aa as a descriptive process. Our claim is that MEC may achieve memory retrieval through the modulation of its own network (that is, the variation of the bump width) under a fixed memory configuration, but we also clearly know that there are biological mechanisms underlying this process. In fact, we proposed another model (which is not in this paper) where we gradually increased the firing threshold of the grid cells, and this model also successfully achieved the retrieval dynamics we proposed in our paper. Besides, short-term plasticity may also regulate the recurrent width in order to achieve this. We may observe the width modulation directly from large-scale population recordings during the transition, or indirectly by comparing the bump widths across different stable states, or by finding intercortical feedbacks that could modulate this.

As for Q2, we would like to summarize the mapping between a neural network and our CANN, hoping that you can form a clearer picture in your mind. The CANN units represent MEC grid cells, with x\mathbf{x} corresponding to grid phase on the latent manifold. The recurrent kernel JJ represents approximately homogeneous interactions within MEC, while IextI_{ext} represents memory-related input transmitted from the hippocampus to MEC.

We concluded that CANN can implement the reverse ODE of DDPM. However, we did not indicate that the learning/training process of CANN and DDPM are also the same. From our perspective, in real neural systems, the learning process is carried out through the formation of intercortical synaptic connections between HPC and MEC. After learning a stable connection, it would not be costful for HPC to transmit memory-related information to MEC; the neural activities of HPC would naturally produce proper external inputs for MEC to conduct memory retrieval. Under this picture, IextI_{ext} is not “presenting all training points”. HPC is activated under a certain retrieval context; these neural activities in HPC are then naturally encoded into external inputs to the MEC through HPC-MEC connections.

As for Q4, we would like to clarify that the CANN is not introduced merely as a memory model for demonstrating the efficacy of an already established annealing retrieval routine. Our primary motivation is neuroscientific: we aim to identify a biologically interpretable network mechanism for memory retrieval that provides new computational capabilities and, at the same time, suggests testable neural mechanisms. In particular, dynamically regulating the interaction scale allows the same recurrent network to transition from a regime with a large capture range to one with high retrieval resolution. The connection to diffusion models then provides a theoretical characterization of this neural retrieval process, showing that the resulting coarse-to-fine dynamics can formally implement a denoising diffusion process.

Finally, thank you for pointing out the important related works in Q5. We will incorporate these works into the manuscript shortly and provide appropriate discussion at relevant places to better situate our work within the related literature.

[1] Toroidal Topology of Population Activity in Grid Cells - Gardner et al., 2022

Reviewer_MH4HOpenReview ↗

Rebuttal

I thank the Authors for addressing my questions. However, I believe that answers might go a bit deeper.

  • The answer to Q1 is useful. It is evident that there might be several biological mechanisms giving origin to this form of memory retrieval, and I thank the Authors for providing examples. Nevertheless, it is still not clear to me why internal network modulation (either in terms of STP and neural activity) in MEC should be the favourite candidate, instead of a modulation of IextI_{ext}. If the aim of the work is also motivating future neuro-physiological experiments, the paper should me more clear and explicit about this. The Authors still do not expand around the biological interpretation of IextI_{ext} as the mixture of Gaussians in Eq. (4).

  • Q2: Where are the neuron-neuron interactions in both J and IextI_{ext} ? As in a dense associative memory (Krotov & Hopfield, 2017), pairwise (as in Battista & Monasson (2020) and Rosay & Monasson (2013) ), and even multi-body interactions across neurons (Garnder, "Multiconnected neural network models" (1987)) are not explicit (there is not an effective graph underneath), and this makes the model less biologically plausible. Please expand.

  • Q3: I understand your answer. Still, as for Q2, the paper is missing the "biological" plausibility: you state "the learning process is carried out through the formation of intercortical synaptic connections". Where are the (learned) synaptic connections in the model?

We thank the reviewer for these further questions.

First, we would like to clarify that both the Gaussian recurrent connectivity and the Gaussian external inputs are widely adopted in established CANN models (e.g., Wu et al., 2008; Fung et al., 2010), and they could both be derived from the neuronal level. Specifically, JJ could be explained by Hebbian learning (e.g., Shen and McNaughton, 1996, Page 686). IextI_{ext} can also be derived in the same way through Hebbian learning. For example, let s\mathbf{s} denote the spatial coordinate of MEC neurons and s′\mathbf{s}' the indexing coordinate of HPC neurons. For memory nn, let rH,n(s′)r_{H,n}(\mathbf{s}') denote the associated HPC firing pattern, and we assume the spatial projection of this memory onto MEC as

-\frac{|\mathbf{s}-\mathbf{x}_n|^2}{2b^2} \right).$$ (We approximate $\phi_n$ as a Gaussian because it provides a simple, mathematically convenient description of localized neural activity while capturing the general statistical features of spatially distributed representations.) The Hebbian-connectivity can then be $$W(\mathbf{s},\mathbf{s}')=\eta\sum_{n=1}^{N}\phi_n(\mathbf{s})r_{H,n}(\mathbf{s}').$$ Given the current HPC firing rates $r_H(\mathbf{s}',t)$, the external input received by MEC is $$I_{ext}= \rho_H\int W(\mathbf{s},\mathbf{s}')r_H(\mathbf{s}',t) ,d\mathbf{s}' =\sum_{n=1}^{N} \alpha_n(t) \exp\left( -\frac{|\mathbf{s}-\mathbf{x}_n|^2}{2b^2} \right),$$ where $$\alpha_n(t)= \eta\rho_H\int r_{H,n}(\mathbf{s}')r_H(\mathbf{s}',t) ,d\mathbf{s}'.$$ This recovers the Gaussian mixture input adopted in our model, and this is the "effective graph underneath". Besides, we've always agreed the possibility of the modulation of $I_{ext}$. We wouldn't like to say "memory retrieval is implemented merely through internal network modulations". What our work suggests is **the mechanism, the effectiveness, and the possibility** of such internal modulation in retrieval tasks. (Perhaps internal+external modulation works better, but that's not what we do in this paper.) In this work, we control the external modulation as almost-fixed (except for the slight alteration in $\alpha$) for a better analysis of the dynamical properties of the internal modulation itself, but we're not claiming "internal network modulation in MEC is the favourite candidate". Moreover, such internal modulation doesn't cost significantly more than external modulations.
Reviewer_yCoQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a dynamical memory retrieval mechanism in continuous attractor neural networks (CANNs), where the activity bump width gradually decreases to balance broad memory capture and fine-grained discrimination. The authors establish a mathematical correspondence between the bump-center dynamics of CANNs and the reverse probability-flow ODE of diffusion models. They further analyze how bump contraction reshapes the attractor landscape, enabling coarse-to-fine retrieval. Numerical simulations validate the theoretical correspondence and demonstrate the proposed mechanism on synthetic memory configurations and hierarchical image representations.

Critical strengths
  1. Interesting theoretical connection.

The connection between CANN dynamics and diffusion models is interesting and provides a new perspective on dynamical memory retrieval. The mathematical derivation and stability analysis offer a clear explanation of how changing bump width can reshape the attractor landscape and control retrieval resolution.

  1. Intuitive coarse-to-fine retrieval mechanism.

The proposed mechanism is conceptually appealing. Dynamically decreasing the bump width allows the network to first capture distant memory representations and subsequently distinguish finer ones. The hierarchical retrieval experiment also illustrates how coarse representations can emerge from the collective influence of fine-grained memories without being explicitly stored.

Critical weaknesses
  1. Limited experimental validation and unclear practical utility.

Although the theoretical framework is interesting, the experiments are primarily based on simplified, manually constructed memory configurations, with mostly qualitative results. In particular, Figure 4 demonstrates convergence toward individual attractors, but does not establish whether the correct memory can be recovered from noisy or incomplete cues. The hierarchical image experiment also relies on manually designed feature representations with predefined hierarchical relationships.

Consequently, it remains unclear whether the proposed mechanism provides meaningful improvements in practical associative memory retrieval tasks. Quantitative evaluations on retrieval accuracy, robustness to noise, and generalization to learned representations would substantially strengthen the paper.

  1. Unclear distinctive advantages of the proposed dynamical mechanism.

The resulting retrieval dynamics resemble a bandwidth-annealed mode-seeking process. While the CANN–DDPM correspondence is theoretically interesting, it is unclear what additional computational benefits arise specifically from implementing this process through CANN dynamics, compared with simpler coarse-to-fine optimization methods.

Furthermore, the exact correspondence relies on specific assumptions, including equal-strength and equal-width memory inputs and the state-dependent input strength in Eq. (11). The applicability of these conditions to more general memory retrieval settings is not sufficiently demonstrated.

Questions for authors
  1. Practical retrieval performance: Can the authors provide quantitative comparisons of retrieval accuracy under noisy or incomplete cues? In particular, how does the dynamic-width CANN compare with fixed-width CANNs across different noise levels and memory densities?

  2. Generalization to realistic representations: In the hierarchical image experiment, the feature space is manually constructed. Would similar coarse-to-fine retrieval emerge in representations learned from real data, without explicitly imposing the hierarchical structure?

  3. Computational significance: What are the specific advantages of the proposed CANN dynamics over alternative bandwidth-annealing or coarse-to-fine retrieval methods? Is there any benefit beyond reproducing similar trajectories?

  4. Theoretical assumptions: How sensitive is the CANN–DDPM correspondence to the state-dependent input strength required by Eq. (11)? Does the correspondence remain approximately valid when the input strength is fixed?

4 comments

Part 1

Response to Question 2

We thank the reviewer for raising this important question. In fact, the manually constructed 4D memory space in Section 4.3 was an intentional choice to isolate the CANN retrieval dynamics from the representation-learning process. Since the geometry of a learned feature space depends on the encoder architecture, training algorithm, and loss functions, we adopted a representative memory configuration to investigate the effects of dynamic width modulation independently of these additional factors.

As a result, we carried out additional experiments, and found that the hierarchical structures can naturally emerge in real representation learning. We used DINOv2 ViT-S/14, including self-supervised adaptation on 41,891 real images without category labels, hierarchical supervision, or explicit geometric constraints.

Importantly, we evaluated the exact same 36 real images used in Section 4.3, all excluded from the self-supervised adaptation training. In the original 384-dimensional feature space, their pairwise distances exhibited a clear hierarchical organization, with distances generally increasing from pairs within the same intermediate category to pairs from different intermediate and major categories. The distance matrix showed a significant correlation with the predefined semantic hierarchy (Mantel Spearman ρ = 0.4238, permutation p ≈ 0.0001). Unsupervised spherical KMeans further recovered substantial portions of the hierarchical groupings, achieving adjusted Rand indices (ARI) of 0.477 and 0.582 for the 4- and 12-cluster partitions, respectively.

We further validated these findings on an independent set of 1,800 images with the same 433 heirarchy, with 50 images per fine-grained category. The correlation between class-centroid distances and the semantic hierarchy increased to ρ = 0.5850 (p ≈ 0.0001), while unsupervised clustering achieved ARIs of 0.568, 0.677, and 0.648 for the 4-, 12-, and 36-cluster partitions, respectively.

These results demonstrate that hierarchical geometric relationships resembling those assumed in our simulations can emerge from self-supervised representations of real images without explicitly imposing such structure. Although we have not directly simulated the subsequent retrieval process in the learned high-dimension space (since it requires us to simulate 10^[several hundred] numbers of neurons at the same time, which is impossible), the additional experiments provide empirical support for the geometric assumptions underlying our model, and support the possibility of similar coarse-to-fine retrieval in learned feature spaces.

We will clarify this motivation in Section 4.3 and incorporate these additional results into the manuscript, if necessary.

Response to Question 4

We thank the reviewer for pointing out this issue. As a result, our CANN is robust to the deviation of the input strength from the form of Eq.11. Actually, since we have carried out a great number of simulations, we are empirically certain that CANNs can carry out the same kind of retrieval dynamics as long as α\alpha isn’t way too small or huge relative to the bump size. Eq.11 is only for the purpose of a rigorous mathematical correspondence between CANN and DDPM. Whether it’s state-dependent or not, time-varying or not would not have a significant impact on the retrieval process.

Furthermore, we found that a fixed α\alpha does not change any properties of when, where and how many equilibriums emerge, as well as their stability. Using Eq.38, we have dxdλ∣constant=αconstantα(x,t)dxdλ∣DDPM-like=Z(x,t)dxdλ∣DDPM-like.\left.\frac{dx}{d\lambda}\right|_{\mathrm{constant}} = \frac{\alpha_{\mathrm{constant}}}{\alpha(x,t)}\left.\frac{dx}{d\lambda}\right|_{\mathrm{DDPM\text{-}like}} = Z(x,t)\left.\frac{dx}{d\lambda}\right|_{\mathrm{DDPM\text{-}like}}. This demonstrates that α\alpha only modulates the magnitude of the bump center velocities at different positions, and doesn’t change the instant attractor landscape. We also carried out many experiments using a fixed α\alpha, and they showed great robustness. You can rest assured about this completely.

As for the equal-strength and -width assumptions, we’ve already mentioned this issue in discussions. Our main purpose for these assumptions is to keep the mathematics simple, and highlight “bump width modulation” as a critical variable. It is foreseeable that our retrieval dynamics also work when we adopt different strengths and widths. It only changes the likelihood that each memory would eventually be retrieved, but does not affect the core properties and functions of our retrieval dynamics.

Response to Question 3 and relevant concerns

We thank the reviewer for this question. We would first like to clarify that our primary motivation is neuroscientific. Our goal is to investigate how biological neural networks can perform effective memory retrieval and to identify the underlying neural mechanisms that may provide new computational capabilities. We do not claim that our model quantitatively outperforms all existing bandwidth-annealing or coarse-to-fine algorithms.

However, our model does show several computational advantages.

The first is computational flexibility. The DDPM-matched reverse dynamics represent only one particular case of the retrieval processes that our CANN can implement. As noted in discussions, the width can be modulated through different schedules, including expansion, contraction, and nonmonotonic variations, potentially enabling a broader range of computational functions.

Besides, our model saves training costs from a broader biological perspective. In our HPC–MEC framework, the only connections that need to be learned are the intercortical connections between HPC and MEC. Once these connections have been established, HPC can naturally generate different patterns of neural activity in response to different task contexts. These activity patterns can then be transmitted through the same learned HPC–MEC connections to provide MEC with diverse task-related memory configurations. MEC only needs to flexibly adjust its bump-width modulation schedule according to different task contexts to perform different tasks, without repeated long-term training. This flexibility is more difficult to achieve in many non-biological algorithms, where adapting to different task contexts often requires relearning the entire set of model parameters.

Thirdly, our model is robust to noise and parameter settings, and does not necessarily rely on following strict parameter equations. (As discussed in our response to Question 4)

Response to Question 1

We fully understand the reviewer’s requirement for quantitative results and comparisons. But it is hard to compress our model performance into certain parameters such as retrieval accuracy. The main reason is that the memory configurations, the initial bump state and the width schedules are so variable that one could not establish a “general” experiment for comparison. The retrieval performance is more meaningfully characterized by the underlying dynamics than by a single accuracy value, which can vary substantially with the chosen experimental conditions.

But we are confident to say that our CANN shows robustness to noise, due to its mathematical property of stability. Besides, the main properties under different width settings are stable across different circumstances.

For example, under a single Gaussian memory input, a CANN with a fixed narrow bump width may remain nearly stationary when initialized far from the target, or jump abruptly to the target when the input is pathologically strong. In contrast, both fixed wide-width and dynamic-width CANNs can gradually approach the target from distant initial positions, while dynamic width further improves retrieval resolution, embodied in a smaller variance of center positions.

For multiple memory inputs, we have also conducted repeated simulations under different initial conditions and memory configurations. The results consistently reproduce the retrieval characteristics illustrated in Figure 4, demonstrating that the main dynamical properties of different width settings remain consistent across these experimental conditions.

I appreciate the authors' detailed responses and the additional experiments showing that hierarchical structures can emerge in learned representations. However, these results do not directly validate the proposed retrieval dynamics, and the practical computational advantages remain largely qualitative without quantitative retrieval evaluations. While I find the theoretical connection interesting, my main concerns regarding the empirical significance of the proposed mechanism remain, so I maintain my original assessment.

We thank the reviewer for the further comments and for acknowledging our additional experiments. We would like to address the remaining concerns point by point.

1. Quantitative validation of the proposed retrieval dynamics.

We have revised our manuscript, where we have added three quantitative experiments in Appendix C (Sections C.1–C.3) to provide a more rigorous characterization of the proposed retrieval mechanism.

First, we systematically investigate how bump width affects the capture range under a single external input, finding a clear linear relationship between bump width and capture range. We also compare the retrieval time of fixed- and dynamic-width CANNs under matched initial conditions, finding that dynamic-width CANNs can retrieve faster. Second, we quantify the maximum bump width that supports successful memory-specific retrieval under different memory densities. The results show that this maximum width decreases as memory inputs become increasingly overlapping, and the critical density for resolvability of memories appeared in our measurement. Third, we compare dynamic- and fixed-width CANNs across eight different initial conditions in a multi-memory setting. All eight dynamic-width CANNs successfully converge to individual memories, whereas fixed-wide CANNs only approach the collective center of the memory configuration and fixed-narrow CANNs fail to reach the memory configuration.

Together, these experiments provide a more systematic and quantitative characterization of the complementary roles of broad and narrow bumps, as well as how dynamical width scheduling combines their respective capabilities. We invite the reviewer to refer to the revised manuscript for detailed information.

2. High-dimensional representations and possible extensions.

We agree that our DINOv2 experiments do not directly validate the proposed retrieval dynamics in a high-dimensional learned representation space, since we did not construct a full "image-encoder-CANN-decoder-output" framework. As discussed in our latest response to Reviewer 1, directly simulating a high-dimensional CANN is computationally prohibitive, as the number of neurons grows exponentially with dimensionality.

However, we have considered a possible extension based on coupling multiple low-dimensional CANNs to jointly perform retrieval in a high-dimensional representation space (See Discussion). This modular organization could substantially alleviate the dimensionality-related computational burden while preserving the local retrieval dynamics within each CANN.

Importantly, this approach has both biological and computational precedents. In the medial entorhinal cortex (MEC), grid cells are organized into modules with different spatial periods, whose combined phase representations enable unambiguous spatial encoding over a range far exceeding that of any individual module (Fiete et al., 2008). More directly, Zhang et al. (2020) demonstrated that multiple reciprocally coupled CANNs, each representing an individual stimulus feature, can perform distributed Bayesian inference over the joint posterior of high-dimensional stimulus features. These studies suggest a biologically plausible route for our model to further carry out high-dimensional memory retrieval tasks.

We would also like to clarify the scope of the present work. Our primary objective is to investigate the neural mechanisms underlying flexible memory retrieval, rather than to develop a complete end-to-end image retrieval architecture.

3. On the significance and practical computational advantages of the proposed mechanism.

We would like to further clarify our position regarding the practical computational advantages mentioned in the review. If this concern refers to algorithmic efficiency or performance advantages over existing machine-learning methods, we acknowledge that we have not demonstrated such advantages. In particular, we do not claim that directly simulating a full CANN provides a more computationally efficient implementation than diffusion or other retrieval algorithms.

As mentioned before, our primary scientific objective is neuroscientific. Our submission was made under the primary area "Applications to Neuroscience & Cognitive Science", and our central contribution is to advance the theoretical understanding of how biological neural networks can implement flexible memory retrieval through intrinsic circuit dynamics.

If the concerns regarding practical utility reflect expectations from computer science (just in case, no offense), we respectfully hope that our contributions can also be reassessed from a neuroscience perspective, with due consideration given to their biological significance.

We fully appreciate the reviewer's concerns and hope that the additional quantitative analyses, together with the clarified research objectives and theoretical significance, will help provide a more complete basis for reassessing our contribution.

Reviewer_tM2POpenReview ↗
Rating
2
Confidence
3
Informativeness
2(+1)
Presentation
2
Soundness
3(+1)
Read review
Paper summary

The paper proposes that memory retrieval in attractor networks can be made both broad and precise by shrinking the attractor basin over time. It models this with a continuous attractor neural network (CANN) whose activity bump narrows during retrieval, while memories enter as Gaussian external inputs. Projecting the network dynamics onto the bump’s position, the authors derive a correspondence to the probability-flow ODE of DDPMs, with width playing the role of the noise level. Using the associated energy landscape, they argue that wide bumps settle into coarse collective states and narrowing bumps resolve individual memories, and they derive a minimum resolvable distance for two memories. Simulations on a 2-D torus compare dynamic and fixed bump widths, and a hand-constructed 4-D hierarchical space with a fitted StyleGAN readout illustrates coarse-to-fine retrieval.

Critical strengths
  • Concrete neural mechanism for a diffusion-like process. The paper gives an explicit, analytically derived mapping, via projection onto the bump’s position mode, from a well-studied neural circuit model to the DDPM probability-flow ODE, with the network’s bump width taking the role of the noise level. While the underlying energy-landscape results are known, grounding them in a specific circuit-level variable is an interesting bridge between neuroscience models and diffusion models.
Critical weaknesses
  • Limited novelty and misleading positioning. The main motivating claim in related work that prior studies have not identified the noise variance as regulating attractor basins and retrieval resolution is inaccurate. Prior work already uses the same energy function and interprets the noise variance as a temperature controlling memory separation [1]. It also already shows that the central fixed point loses stability as noise decreases [2], and that the reverse dynamics proceed coarse-to-fine [3, 4]. Therefore, the only new piece in Sec 3. is mapping these existing results onto a CANN. Moreover, the “memory” interpretation is inconsistent with prior associative memory work: the CANN’s recurrent weights store no memories, and the attractor basins belong to the external input field rather than the network. This weakens the claims around their main contribution and the framing makes for a confusing first read.

  • Inconsistent theory and oversimplified experiments. The correspondence requires the input strength α\alpha to depend on the current network state (Eq. 11), which contradicts the constant-input CANN formulation and grows unboundedly far from the memories, violating the weak-input assumption behind the bump solution. Similarly, the connectivity range a\text{a} is treated as fixed in the derivation but must vary over time for the correspondence. The large- and small-width arguments are informal, and the intermediate regime is not analyzed. The experiments are very idealized and simplified verifications, sec. 4.3 is somewhat misleading at first glance since the CANN operates only in a hand-constructed 4D space, and the retrieval itself has nothing to do with the images. Also, most of the key simulation parameters are missing. As a result, it is unclear if there is any practical utility of this connection.

References:

  • [1] Memorization to Generalization: Emergence of Diffusion Models from Associative Memory - Pham et al., 2025
  • [2] Spontaneous Symmetry Breaking in Generative Diffusion Models - Raya & Ambrogioni, 2023
  • [3] Dynamical Regimes of Diffusion Models - Biroli et al., 2024
  • [4] A Phase Transition in Diffusion Models Reveals the Hierarchical Nature of Data - Sclocchi et al., 2024
Questions for authors

Clarifications

  • Is the external input strength α\alpha meant to be constant or state-dependent? The CANN formulation treats it as constant, while the DDPM correspondence (Eq. 11) requires it to depend on the current state xx, which makes the system closed-loop and grows unboundedly far from the memories. How is this reconciled with the weak-input assumption behind the bump solution?
  • Similarly, is the connectivity range aa constant or time-varying? It is treated as a fixed connectivity parameter in the formulation and derivation (Sec. 2.1, Eqs. 3, 17, 27), but must vary over time for the correspondence (Eq. 12).
  • The memories are supplied entirely through the external input, and the recurrent weights store none of them. In what sense is this associative memory retrieval rather than input-driven bump dynamics? Where are the memories stored in this model?
  • How does this work differ from prior results on the energy landscape and noise-dependent stability of diffusion models (e.g., [1–4]), beyond the mapping onto a CANN?
  • Could you provide the parameters used in the experiments (e.g., noise schedule, time constants, u, J, k, ρ\rho, grid resolution, and the noise level added in Sec. 4.2)?

Suggestions:

  • Provide a more rigorous justification for the behavior asserted at large and small widths, which is currently argued informally.
  • Analyze the intermediate regime, where the width is neither large nor small relative to the memory separation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.