{"id":"07d1419a-1648-4b5d-84f3-6c22df5ae04e","arxiv_id":"2501.10376","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Delay-conditioned deep joint source-channel coding improves energy-constrained image storage on simulated memristors subject to resistive drift.","lead":"This paper trains a neural network to store images on simulated memristor devices, balancing reconstruction quality against the energy spent setting each device's resistance. It shows that conditioning the encoder and decoder on the storage delay improves image quality, most clearly when the energy budget is tight.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported energy-constrained comparisons may not operate under the stated constraint: Eq. (18) penalizes deviations from the budget in both directions and §V.A admits violations, so the delay-conditioned PSNR gain could reflect extra energy rather than coding.","rationale":"The reader's weakest assumption is external validity: the cGAN and event-based model may not faithfully represent real memristive drift. That is a legitimate concern and the ground-truth model is itself authored by the same group, so it cannot fully resolve it. However, I find a more immediate internal problem: the stated optimization problem is not the one actually solved. Eq. (1) defines an upper bound on average energy, but Eq. (18) penalizes any deviation from the budget, including under-spending. This is not a harmless implementation detail: the energy regularisation shapes the learned encoder distribution, and the paper's own §V.A admits that energy constraints are violated in the reported results. Because the headline claim is explicitly about performance under equal average energy budgets, the comparison is only meaningful if the measured average energy is equalized across conditioning schemes. The reported legend averages in Figure 8 are close to the budget, but the PSNR curves in Figure 6 extend beyond the training delay range and the text acknowledges violations; the paper does not provide per-delay measured energy for all compared schemes. A one-sided constraint retraining is the cleanest way to determine whether the delay-conditioning benefit is a genuine coding gain or a consequence of allowing the conditioned encoder to spend more energy in the regimes where it shows the largest PSNR advantage. This does not invalidate the paper's broader algorithmic contribution, but it does mean the central claim needs an explicit check before being accepted. I therefore keep the verdict conditional, with the required condition being the one-sided energy-constraint experiment rather than only external hardware validation.","tokens_in":17747,"tokens_out":12684,"duration_ms":128307,"concrete_test":"Re-train the unconditioned and both-conditioned autoencoders for the 0.01 J and 0.1 J cases with Eq. (18) replaced by a one-sided soft constraint, e.g. r_energy = max(0, (1/n)Σ_i E(M_i)/e_b - 1)^2, keeping all other hyperparameters identical. Then compare PSNR-versus-delay curves and, crucially, report the measured average energy per scheme per delay. If the conditioned-vs-unconditioned gap shrinks to the size of the energy differences in the original Figure 6, the headline claim is an artefact of the two-sided regulariser; if the gap persists at matched measured average energy, the concern is resolved.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that a delay-conditioned DeepJSCC encoder/decoder pair outperforms a delay-agnostic one under equal average energy budgets. Two parts of the paper make this claim insecure even within the simulation. First, the regulariser used to enforce the constraint is not the constraint in Eq. (1). Eq. (18) defines renergy(M) = (1/n) Σ_i (1 - E(M_i)/e_b)^2, which is minimized when every device spends exactly e_b and penalizes under-spending as much as over-spending. This is an equality target, not an upper bound of the form EM[E(m)] ≤ B. Second, §V.A states explicitly: \"Energy constraints are violated in order to ensure that enough information is retained, following storage and recovery, to reconstruct the image to a reasonable degree.\" If the compared schemes violate the constraint by different amounts, or if the reported PSNR curves are evaluated at delays where only the conditioned encoder is allowed to exceed the budget, the advertised advantage may be an energy artefact rather than a coding gain. The reader's external-validity concern about the cGAN/event-based model is real, but this internal mismatch is more load-bearing because it threatens the headline comparison before any hardware question is reached.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies analogue information storage on memristive devices as a delay-dependent communication problem. The authors derive an energy cost function E(R)=|A ln(B/R)| from a simplified model of programming pulses, formulate the storage problem as an energy-constrained joint source-channel coding (JSCC) problem, and train a DeepJSCC autoencoder with a conditional GDN (cGDN) architecture so that the encoder and/or decoder can be conditioned on the storage delay. The channel is represented by a cGAN trained on a simulated event-based resistive-drift dataset from the authors' prior work. Experiments on CIFAR-10 compare unconditioned, encoder-conditioned, decoder-conditioned, and jointly conditioned settings under energy budgets from 1.0 J to 0.01 J, with reconstruction quality measured in PSNR, and include a ground-truth evaluation on the event-based simulator. The central claim is that the jointly delay-conditioned scheme achieves higher and more balanced PSNR across delays under equal average energy budgets.","tokens_in":18054,"tokens_out":6007,"duration_ms":55174,"significance":"If the central comparison holds, this is a useful step for analogue storage: it demonstrates a practical way to adapt a single JSCC codebook to a time-varying storage channel and an explicit energy-regularised training objective. The cGDN conditioning layer and the separate energy regularisation on a fixed delay grid are interesting architectural ideas, and the ground-truth evaluation in Section V.D is a good sanity check that is often missing in deep-learning-for-channels papers. However, the paper's 'fundamental' energy-retention trade-off is not established beyond a phenomenological model, and the experimental support is entirely internal to a simulation chain in which the cGAN and the ground-truth simulator are both from the authors' prior preprints; the headline energy-constrained comparison is also weakened by the admitted energy-constraint violations and by the equality-form regulariser.","major_comments":[{"comment":"The energy regulariser in Eq. (18) does not implement the constraint in Eq. (1). Eq. (18) penalises (1 - E(M_i)/e_b)^2, which is minimised when every encoded resistance costs exactly e_b; spending less than the budget is penalised as heavily as spending more. This is an equality target, not an upper bound. Section V.A then states that 'Energy constraints are violated in order to ensure that enough information is retained...', so the compared schemes may operate at different actual average energies at the delays used for the PSNR comparison. The reported PSNR advantage of the conditioned scheme could therefore partly reflect higher energy expenditure rather than better coding. Please report the achieved average energy for each setting and delay, and either enforce the constraint during evaluation or compare at equalised average energy.","section":"IV.C, Eq. (18); V.A"},{"comment":"The entire experimental chain is validated only against the authors' own simulation. The cGAN channel model of [29] is trained on the event-based simulator of [30], and Section V.D evaluates the trained autoencoder on that same event-based simulator; there is no independent physical measurement anywhere in the loop. Figure 10 shows visible PSNR drops when switching from the cGAN to the event-based model at several delays, so the generative model is not perfect even within the simulated family. In the absence of device data or at least a broader family of simulators, the results should be presented as valid for the simulated drift model, not for memristive devices in general.","section":"IV.A, IV.H, V.D"},{"comment":"The abstract and contributions describe Eq. (12) as a 'novel, fundamental trade-off' between energy and information lifetime. The derivation rests on several unvalidated simplifying assumptions: constant compliance current throughout the pulse, conductance linearly proportional to pulse duration, symmetric switching energy, and the specific form of Eq. (9). The constants A and B are set to illustrative values in Figure 2, and no measurement is provided. This is a plausible phenomenological model, but 'fundamental' is an overclaim. I recommend re-scoping the claim and stating explicitly that Eq. (12) is an illustrative first-order model.","section":"III.A"},{"comment":"None of the PSNR results in Figures 6, 8, or 10 include error bars, multiple seeds, or significance tests, and the main claim of 'higher and more balanced reconstruction quality' rests on differences between single curves. Given the energy-constraint issue above, it is important to know whether the advantage of the jointly conditioned scheme is stable across training runs and statistically distinguishable from the unconditioned baseline.","section":"V.A, V.B, V.D"}],"minor_comments":[{"comment":"There is a typo in the paragraph following Eq. (2): 'mehtods' should be 'methods'.","section":"II"},{"comment":"The notation for the regularisation weights is inconsistent: the text defines lambda_res and lambda_energy, while Eq. (19) uses lambda_resistance and lambda_energy; in addition, the loss in Eq. (19) is written with a square root and calls the distortion MSE, although l was defined as a Frobenius norm in Eq. (2). Please align these definitions.","section":"IV.C, Eq. (19)"},{"comment":"Section IV.H says the networks are trained for delays d in [0,1000], but Figure 6 reports results up to 1400 s; the recurrent cGAN extension with n=4 makes this possible, but the discrepancy should be stated explicitly in the text.","section":"IV.H, V.A"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new piece here is the cGDN conditioning mechanism: a single encoder/decoder pair that takes the storage delay as a continuous side input, plus the energy-regularized JSCC formulation. That is a fair increment over the prior DeepJSCC storage papers [14], [19], [20], which did not condition on delay or budget. The paper also deserves credit for evaluating on the ground-truth event-based sim (Fig. 10), which shows real effort to check the cGAN surrogate, and for clearly separating encoder-only, decoder-only, and joint conditioning. The architecture work looks competent.\n\nThe soft spots are real but not fatal to the core algorithmic claim. First, the energy constraint is not the upper bound stated in Eq. (1). The regularizer in Eq. (18) is an equality penalty, and Section V.A openly admits that \"energy constraints are violated\" for some delays. That means the headline PSNR gains at low budgets could be partly an energy artifact, not a pure coding gain. The stress-test note is right about this internal mismatch. Still, the paper does report normalized energy curves, and the gains are larger at low budgets where the constraint binds, so the direction is plausible—but the numbers in Figs. 6 and 8 should be re-plotted with actual average energy held equal before you trust the magnitude.\n\nSecond, external validity is thin: the channel and ground truth are both the authors' own preprints [29], [30], no code or data are released, and there are no error bars. That is a serious limitation for a hype-prone field, but it does not invalidate the comparison within the simulation. The cGAN/event-based mismatch in Fig. 10 shows some PSNR drops at certain delays, so the surrogate is not perfect.\n\nThe \"fundamental trade-off\" language for Eq. (12) is overblown—it is a simplified first-order derivation, not a fundamental bound—but the energy cost function itself is a reasonable closed-form parameterization. The citation pattern is fine; self-citations are to the actual channel models used. The math is internally consistent.\n\nWho this is for: people working on analog storage on memristive devices, or on conditioning DeepJSCC for time-varying channels. A serious referee should see it, mainly to force a cleaner constraint-satisfaction analysis and ideally a code/data release. I would not cite it within the next year without independently checking the energy accounting, but it deserves peer review.","headline":"A solid algorithmic contribution—delay-conditioned DeepJSCC with cGDN—but the energy-constraint enforcement is loose, and the channel model is entirely the authors' own simulation.","tokens_in":18582,"tokens_out":622,"would_cite":false,"duration_ms":7909,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A delay-conditioned DeepJSCC encoder/decoder pair beats delay-agnostic storage coding under equal average energy budgets.","keywords":["memristors","resistive drift","deep joint source-channel coding","analogue information storage","conditional GDN","energy-constrained storage","generative adversarial network channel model","delay conditioning"],"falsifier":"Take a real memristor array, program it with the resistance distributions produced by the delay-conditioned encoder at 0.01 J for short and long delays, wait each delay, read back, and compare reconstruction PSNR against the unconditioned encoder under identical conditions; if the conditioned scheme stops outperforming on hardware, the simulated drift model was the load-bearing assumption rather than the coding scheme.","tokens_in":17519,"feed_emoji":"🔋","tokens_out":10146,"duration_ms":82557,"temperature":0.7,"pith_summary":"The paper treats storing an image on an array of memristors as communication through time, where the channel is the delay-dependent resistive drift that degrades stored resistances between write and read. Its central claim is that a single deep joint source-channel coding (DeepJSCC) encoder/decoder pair, conditioned on the storage delay, learns an energy-aware analogue coding scheme that reconstructs natural images at higher and more balanced peak signal-to-noise ratio across delays than delay-agnostic coding under the same average energy budget, with the largest margin at the tightest budget studied, 0.01 J. To make this work, the paper derives a parametric energy cost $E(R)=|A\\ln(B/R)|$ from published device measurements, linking the energy needed to program a memristor to the resistance state reached, and uses it as a regulariser during end-to-end training against a generative model of resistive drift. A sympathetic reader would care because this points to a practical way to use the analogue density of memristors while managing their main weakness, the slow stochastic drift of stored values, under a hard energy constraint.","feed_headline":"Delay-aware codec wins on low-energy memristor storage","feed_subtitle":"One encoder/decoder pair conditioned on storage delay beats one-size-fits-all coding across readout times, especially at 0.01 J.","key_machinery":"The load-bearing machinery is the combination of four components. First, the energy cost function $E(R)=|A\\ln(B/R)|$ derived from published pulse-programming measurements: it turns the physical observation that lower resistance states are more stable but more expensive to program into a differentiable regulariser. Second, conditional GDN (cGDN), a modified form of generalised divisive normalisation in which the per-channel normalisation parameters $\\beta$ are shifted by a learned, fully-connected embedding of the normalised delay; the inverse transform, icGDN, is used in the decoder, so delay information modulates every convolutional layer. Third, a delay- and initial-resistance-conditioned cGAN trained on an event-based drift model serves as a differentiable channel model for end-to-end gradient flow. Fourth, a delay-specialised batch schedule enforces the average energy constraint over a fixed set of delays, letting the network trade energy across delay conditions.","core_discovery":"The central claim is that the delay itself is usable side information for storage, not just a nuisance: an encoder/decoder pair that both know the delay at write and read time learns an energy-aware analogue coding scheme that outperforms delay-agnostic coding under the same average energy budget. The paper supports this with CIFAR-10 experiments over delays from 0 to 1000 seconds, showing the conditioned pair achieves higher and more balanced PSNR than unconditioned, encoder-only, and decoder-only baselines, with the biggest improvement at 0.01 J, where the unconditioned network degrades sharply. The authors also claim a new trade-off between programming energy and information retention, captured by $E(R)=|A\\ln(B/R)|$, with $B=R(0)$ the equilibrium resistance and $A=\\tau_{\\mathrm{final}}K^2(1/R(\\tau_{\\mathrm{final}})-1/R(0))$; this function becomes the energy regulariser in the training loss. They implement conditioning through a modified divisive normalisation layer, conditional GDN (cGDN), which shifts the normalisation parameters by a learned embedding of the delay. The paper's own ground-truth evaluation against the event-based model shows the learned scheme transfers for most delays, with some drops where the surrogate statistics were not perfectly matched.","pith_inferences":["Because the energy regulariser is only a soft constraint, the exact PSNR at 0.01 J should depend on the weighting of the energy term; tightening that weighting could change the delay-energy allocation and is a direct ablation to run.","The derived cost function rests on a linear-conductance approximation and could be tested directly on fabricated devices by measuring pulse energy versus retained resistance, since the paper validates it only through the simulated channel.","cGDN is a generic conditioning mechanism: the same idea could condition a JSCC codec on channel signal-to-noise ratio in wireless fading or on battery state, because it only requires a differentiable scalar condition."],"forward_implications":["A single trained delay-conditioned pair covers the whole delay range, so storage systems need not train or store a separate codec per expected readout time.","Under a tight energy budget, delay conditioning becomes necessary: the unconditioned baseline's reconstruction collapses at 0.01 J, while the conditioned pair maintains usable PSNR.","The network allocates the energy budget across delays, shifting more of it to long, noisy delays and using sparser, higher-resistance encodings for short delays.","The scheme transfers from the differentiable cGAN surrogate to the event-based ground-truth model for most delays, indicating the learned coding strategy is not merely fitted to the surrogate."],"supporting_citations":[{"why":"Supplies the differentiable cGAN channel model that approximates the event-based resistive drift statistics and is used for end-to-end training.","marker":"[29]"},{"why":"Supplies the underlying event-based stochastic drift model used to generate the training dataset and as the ground-truth evaluation channel.","marker":"[30]"},{"why":"Introduces the DeepJSCC autoencoder architecture and training paradigm that this paper adapts to storage, replacing GDN with cGDN.","marker":"[22]"},{"why":"Defines the generalised divisive normalisation transform that cGDN modifies for delay conditioning.","marker":"[32]"},{"why":"Provides the conditional batch normalisation idea that inspired the cGDN conditioning mechanism.","marker":"[33]"},{"why":"Provides the inverse proportionality between compliance current and final resistance used in the derivation of the energy cost function.","marker":"[16]"},{"why":"Provides the proportional relationship between pulse duration and final conductance used in the same derivation.","marker":"[27]"}],"fun_headline_variants":["Delay-aware codec beats agnostic in low-energy memristor storage","Energy budget meets read delay: memristor codec learns trade-off","cGDN codec tunes energy-delay trade-off for memristor storage","Delay-conditioned deep codec lifts low-energy memristor storage quality","Read-delay as side info: energy-aware memristor storage codec"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the computer model of resistive drift used for both training and testing faithfully matches real memristor behaviour; if real devices drift differently, the reported image-quality gains may shrink or disappear.","fun_headline_variants_meta":{"raw":{"variants":["Delay-aware codec beats agnostic in low-energy memristor storage","Energy budget meets read delay: memristor codec learns trade-off","cGDN codec tunes energy-delay trade-off for memristor storage","Delay-conditioned deep codec lifts low-energy memristor storage quality","Read-delay as side info: energy-aware memristor storage codec"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001031,"raw_usage":{"total_tokens":4396,"prompt_tokens":1051,"completion_tokens":3345,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":667,"completion_tokens_details":{"reasoning_tokens":3245}},"tokens_in":667,"tokens_out":3345,"duration_ms":23838,"temperature":1.0,"reasoning_tokens":3245,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:43:56.744360+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a real memristor array, program it with the resistance distributions produced by the delay-conditioned encoder at 0.01 J for short and long delays, wait each delay, read back, and compare reconstruction PSNR against the unconditioned encoder under identical conditions; if the conditioned scheme stops outperforming on hardware, the simulated drift model was the load-bearing assumption rather than the coding scheme.","supporting_citations":[{"cited_title":"Deep Joint Source- Channel Coding for Wireless Image Transmission,","cited_arxiv_id":null,"evidence_quote":"Introduces the DeepJSCC autoencoder architecture and training paradigm that this paper adapts to storage, replacing GDN with cGDN."},{"cited_title":"Modulating early visual processing by language,","cited_arxiv_id":null,"evidence_quote":"Provides the conditional batch normalisation idea that inspired the cGDN conditioning mechanism."},{"cited_title":"Modeling the Universal Set/Reset Characteristics of Bipolar RRAM by Field- and Temperature-Driven Filament Growth,","cited_arxiv_id":null,"evidence_quote":"Provides the inverse proportionality between compliance current and final resistance used in the derivation of the energy cost function."},{"cited_title":"Size-Dependent Retention Time in NiO-Based Resistive-Switching Memories,","cited_arxiv_id":null,"evidence_quote":"Provides the proportional relationship between pulse duration and final conductance used in the same derivation."}],"review_version":1}