acceptodds
Under review as a conference paper at ICLR 2027

Learning Photonic Dynamics for Higher-Order Associative Memory

Abstract

Photonic hardware is usually used to accelerate fixed neural operators. We ask whether a photonic feedback loop can instead be trained as part of the neural model. We study associative memory—recovering a stored pattern from a corrupted cue—with recurrent photonic dynamics. Programmable optical projections and nonlinear interactions repeatedly update an internal state. Higher-order interactions reduce interference among stored memories, and a readout based on the stored memories restores information lost by normalization. At the highest tested load of 128 stored memories, a model trained for four feedback steps reconstructs 5.3 dB better at four steps than after running to convergence. Thus, more convergence can make the task result worse: runtime itself becomes part of what the model learns. Controlled retrieval shows when higher-order interactions help, while a real multi-hop retrieval test shows they can also hurt generalization. Finally, we map the model to a hybrid photonic/electronic circuit and simulate finite precision, loss, drift, and calibration. These hardware results are simulations, not measured device performance.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Photonic AI, Analog computing, Nonlinear dynamics, Associative memory

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_4tiuOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

It is not fully clear to the reviewer what exactly this submission is trying to achieve. The abstract is rather cryptic. It is clear that the authors are trying to include some principles of photonic dynamics into the process of recurrent computations commonly used in some associative memories (it is not clear either from title, abstract or EVEN introduction, which type of associate memory authors are working with). The reviewer does not really understand what the study is trying to do based on its introduction (e.g., "Higher-order associative memories make this question concrete." but what is the question?) and does not find it possible to comment on authors's contributions.

Critical strengths

The starting point of the paper is interesting: whether optical projections can be a useful part of an associative memory with recurrent computations.

Critical weaknesses

Based on both the initial assessment of the reviewer and on the self-assessment provided in “AI Use Statement”, the paper appears to be almost entirely AI-generated, which explains why it is nearly unreadable (not in terms of language but it terms on being a piece of academic writing).

Questions for authors

Based on the text in connection to “AI Documentation For Reviewing” panel, the reviewer's interpretation is that they are softly discouraged to use any AI tools during the review process. When this observation is combined with the earlier conclusion that the paper was created almost entirely by AI, the reviewer does not find it neither necessary nor fair to spend a significant amount of their own time formulating questions to the authors, especially, considering the authors' limited involvement into the process of conducting the research reported in the paper.

2 comments

We appreciate the reviewer's concern about the accessibility of the central idea. We have comprehensively revised the manuscript's abstract, introduction, method overview, and discussion to state the task and contribution before introducing the hardware and auxiliary experiments.

Task. We store image representations in an associative-memory bank and recover the corresponding identity and continuous image reconstruction from a corrupted cue. A recurrent higher-order energy layer compares the evolving state with stored memories, strengthens higher-order compatibility, and updates the state. A memory-conditioned radial readout restores the magnitude required for image decoding.

Research question. How should the recurrent dynamics be learned when inference ends after a prescribed number of feedback steps? The key result compares a four-step-trained memory with its own fixed point:

Original image-memory experimentT=4 executionConverged fixed point
P=48 retrieval / PSNR95.8% / 21.59 dB81.3% / 20.27 dB
P=128 retrieval / PSNR89.6% / 18.44 dB32.3% / 13.13 dB

At P=128, the finite execution horizon provides 5.31 dB higher PSNR. The same model is evaluated at both depths, and state/gradient-convergence diagnostics confirm that the fixed-point computation is being approached as iteration continues. The training-depth controls also show how the preferred execution depth moves with the learning horizon.

Presentation changes. The abstract opens with this task; the introduction states the execution-budget question and quantitative result; Figure 1 identifies the optical projection and electronic recurrence operations; and the detailed mathematical and photonic design choices follow in the method. The conceptual system illustration is placed in the appendix. The related controls are grouped by their explanatory role: interference and basins, reconstruction scale, finite-depth training, and physical sensitivity. We have proofread the full main text and supplementary material for terminology, transitions, and correspondence between each claim and its table or figure.

AI-assisted preparation and provenance. The manuscript retains a detailed AI Use Statement describing brainstorming, implementation, experimental design, drafting, and the two conceptual illustrations. The quantitative figures and image panels are linked to retained model outputs and experiment records; the updated paper identifies the modeling assumptions behind its optical controls. The authors are responsible for validating the computations, experimental records, citations, and final scientific conclusions.

We thank the reviewer for prompting a clearer, more direct presentation of the work.

The reviewer would like to thanks authors for the rebuttal and for having considered the highlighted weaknesses and questions.

Reviewer_jJ4bOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a differentiable recurrent energy layer for higher-order associative memory, together with a conceptual mapping to hybrid photonic-electronic hardware. The authors investigate higher-order Hopfield interactions, learned attraction basins, finite-depth versus equilibrium dynamics, and memory-conditioned latent reconstruction. Experiments are conducted primarily on synthetic associative memory tasks and pretrained image latents, with additional simulated hardware nonidealities.

Critical strengths

None

Critical weaknesses

The connection to photonic computing is superficial and does not constitute a meaningful technical contribution. Despite the title and framing, the proposed methodology is essentially a conventional higher-order recurrent associative memory model that can be implemented entirely in software. The central energy formulation, recurrent update, learning procedure, and experimental conclusions do not rely on any distinctive property of optical computation. The photonic component is introduced primarily through a generic mapping of linear projections onto programmable MZI meshes. Importantly, the higher-order nonlinear interactions, state updates, normalization, and feedback control are implemented electronically in the proposed hybrid architecture. These are precisely the operations that define the core computational behavior of the method. Consequently, the optical hardware does not introduce a new computational primitive, enable a fundamentally different algorithm, or impose meaningful constraints that shape the proposed learning method. Furthermore, there is no actual photonic hardware demonstration, nor any quantitative evidence of improved latency, throughput, energy efficiency, or computational scalability compared with electronic implementations. The simulated loss, quantization, and drift experiments do not establish a practical advantage of photonic computation. The hardware mapping itself raises substantial scalability concerns. For the flagship latent dimension of (N=1024), the generic optical realization requires approximately 2.1 million MZI cells according to Table 8. Although the authors acknowledge that this is an unoptimized upper-bound mapping, no feasible alternative architecture or convincing system-level analysis is provided. A fundamental question is what would change scientifically if the photonic component were removed entirely. As far as I can tell, all principal methodological and experimental findings would remain unchanged. This makes the photonic framing appear largely incidental rather than an integral contribution of the work.

Questions for authors

See above

1 comment

We thank the reviewer for emphasizing the need to distinguish the learning result from its physical realization and to compare against strong digital implementations. We have revised the paper so these roles are explicit.

Scientific contribution. Our primary finding concerns learning a recurrent computation for a finite execution horizon. For image-latent associative memory, a model trained for four steps achieves 18.44 dB reconstruction PSNR at four steps and 13.13 dB when the same parameters are evaluated at the converged fixed point (P=128; submitted Table 2). State and memory-gradient errors simultaneously converge toward the equilibrium computation. This establishes that improving equilibrium accuracy can change the task output selected by finite-horizon training. The matched-depth experiments demonstrate how the preferred inference horizon depends on the training horizon.

Role of the optical circuit. In the hybrid mapping, interferometric meshes perform linear overlap and backprojection operations, with higher-order products, accumulation, normalization, and feedback control in electronics. The proposal ties the number of physical projection/conversion cycles to the learned execution budget. The mathematics also admits digital implementation, and we have added an exact reduced digital reference for the controlled fourth-order recurrence.

Optical resource reduction. The originally reported 2,095,104 MZIs at N=1024 arise from a square-padded two-core SVD/Clements construction. We implemented full-rank rectangular SVD/Givens synthesis and verified its ideal forward projection, reverse projection, and fourth-order field against dense linear algebra (relative discrepancies on the order of 10^-14).

N=1024Submitted generic two-core countRectangular two-core countConditional reciprocal-core count
P=482,095,10498,208 (21.3× reduction)49,104
P=1282,095,104261,888 (8.0× reduction)130,944

The independent two-core reduction preserves projection rank. The reciprocal-core column evaluates an additional bidirectional-reuse assumption. Physical realization of either structure also depends on routing, optical depth, insertion loss, signed reception, conversion, and phase programming; our P=128 rectangular compiler has a 1,403-cell worst-depth scheduling proxy.

Exact digital baseline. For the specified cue-plus-memory fourth-order recurrence, the ideal state remains in the span of the cue and stored memories. Its updates can be evaluated exactly through a (P+1)-coordinate Gram-space representation. Across the tested loads and seeds, full-state disagreement stays below 6.9×10^-13 through 32 steps. At P=128, T=4, and 128 queries reusing the same memory bank, this formulation uses 2.29× fewer modeled MACs than direct dense recurrence, including Gram setup and reconstruction of the output state. It provides the appropriate strong digital reference for this controlled recurrence.

Component-level energy implications. We extended the source-anchored hardware sensitivity analysis to phase-control holding power, optical loss, conversion, and memory-bank reuse. Under the investigated thermo-optic parameter assumptions, the digital Gram reference has lower modeled energy than the active optical mesh; an illustrative shared-core configuration requires about 72.7 W of modeled phase-tuning power. These are calculated device scenarios, with their coefficients and assumptions recorded in the supplementary evidence. The revised manuscript presents the optical design as an explicit, quantitatively constrained execution architecture whose engineering trade-offs can now be assessed against an efficient digital implementation.

These additions provide a sharper account of what is learned, what the optical hardware executes, and which resource constraints govern practical deployment.

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

The manuscript introduces a concept of a hybrid opto-electronic recurrent neural network that borrows ideas from a Dense Associative Memory, yet, adds its own twist in using the network. The concept is shown to work in training and inference. The manuscript gives a snapshot of a larger body work or project in progress. Maybe for that reason it feels fragmented. The project behind is not revealed which leaves the reader wondering backgrounds and motivations. For example, it is unclear what motivates authors to use optics as part of the computing platform. I am sure they have valid reasons, such as minimization of latency, or to improve throughput, or to conserve energy, but these are not revealed and therefore the use of optical computing feels somewhat artificial. The same functionality could naturally have been achieved by conventional digital computation.

Training of recurrent photonic circuits is interesting. The scientific literature that concentrates more on feedforward networks and inference using physical computing systems.

What I found also interesting and non-intuitive is that the system can be trained to give optimal (albeit non-perfect) answers after a limited number of recurrent steps. That is, not to wait until the system reaches energy-minimum steady-state. I am not sure if this result can be generalized, but assuming that it can, that is an interesting finding from a practical engineering point of view. In physical systems signal propagation delay can limit the performance, and thus limiting the number of iterations can save time. That said, it might be difficult to build an analog computing system where all state vector values can be recorded exactly at the right moment of time, and at least I would imagine that in practice high-dimensional state vector element values are read at different phases of evolution. In the case of energy minimum and steady state, this would not be a problem, because the system does not evolve any longer, but in a dynamic system that would create a new source of error. That does not render this idea impractical, but adds a layer of complexity in real-life implementation.

The finding that higher-order dynamics improves specificity in memory retrieval, yet may harm the sensitivity, makes sense.

The text is dense and hard to follow at times. The English is probably grammatically correct, but occasionally the text mixes ideas and concepts. For example this sentence combines multiple concepts and topics: "Photonic neural networks span interferometric, recurrent, diffractive, and tensor architectures, including systems trained in situ ..." I understand that authors refer to various scientific papers that realize neural networks by different means, but reading the text for the first time is still confusing.

Critical strengths
  1. Training of a neural network including a photonic circuit is uncommon
  2. System can be trained to give optimal response with limited number of iterations
Critical weaknesses

I discussed the weaknesses already in the Paper Summary

Questions for authors

None

1 comment

We thank the reviewer for appreciating the potential of training recurrent associative-memory dynamics for a prescribed number of steps. We have strengthened the motivation and made the feedback interface and synchronization assumptions explicit.

Motivation and main result. We address recovery of a stored image representation from a corrupted cue using a recurrent memory layer with an execution budget. The central learning question is how to optimize the output produced at the executed depth. For the same fourth-order model trained with four steps, the original image-memory experiment gives the following results:

Stored memoriesAt the trained horizon, T=4At the same model's fixed point
P=4895.8% retrieval; 21.59 dB PSNR81.3%; 20.27 dB
P=12889.6%; 18.44 dB32.3%; 13.13 dB

The 5.31 dB difference at P=128 is an inference-horizon comparison using fixed learned parameters. The dense early-depth sweep, matched training-depth controls, and state/gradient-convergence diagnostics in Section 5 and Appendix I explain the effect: finite-horizon optimization shapes a useful transient output, while continued iterations move the same system toward equilibrium. The preferred inference horizon also shifts with the training depth.

Why photonic feedback? The reference hybrid circuit assigns the repeated linear memory projections to programmable interferometric optics; the electronic/mixed-signal path performs conversion, higher-order products, recurrent state updates, normalization, and readout. The number of projection-and-conversion passes creates a concrete physical inference budget. Our revised Section 3 defines clocked sampling between the receiver and the electronic state update and distinguishes slow memory-bank programming from fast per-query modulation.

Follow-up timing and calibration study. We evaluated a clocked fourth-order feedback model with three converter stages, channel-gain variation, and relative sampling skew. The table reports retrieval over controlled bipolar-memory banks, averaged across three memory-bank seeds and 128 held-out queries per seed:

N=1024, P=128, T=4; 44% cue-bit flipsTop-1 retrieval
Ideal arithmetic82.8%
Three 8-bit DAC/ADC stages81.0%
8 bits + 5% gain variation + up to 0.1-step skew77.9%
Above + gain estimated from eight independent pilot probes82.0%

Eight-probe calibration recovered 4.17 percentage points under the joint perturbations. The timing model interpolates current and previous projection outputs; gains are estimated from a separate set of known input probes. Across three seeds the improvement at P=128 is consistent, while P=48 shows mixed behavior. These results quantify the sensitivity of a specific feedback model and show how calibration can mitigate gain mismatch. The revised supplementary section documents probe overhead and the simulated sampling assumptions.

We appreciate the suggestion: explicit timing boundaries now connect the learned finite-step semantics to a reproducible execution model.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.