{"id":"c49713eb-437d-4b45-b38c-2b0537beaa8f","arxiv_id":"2509.17152","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A stochastic exponential modern Hopfield network trained on MNIST shows a critical noise region, around p from 0.23 to 0.3, where retrieval overlap drops and the overlap dynamics become strongly persistent, with critical noise decreasing as pattern load increases.","lead":"This paper simulates a stochastic modern Hopfield network with an exponential memory function under pixel-flip noise and tracks the average overlap and temporal correlation scaling as the noise level rises. It reports a critical noise range where retrieval fails and the network's overlap dynamics develop strong temporal memory, with a critical noise level that falls as more patterns are stored.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported H≈1.3 critical exponent is not separated from DFA trend artifacts; the long-memory claim needs a surrogate/detrending-order check before the phase-transition claim is accepted.","rationale":"The reader's weakest assumption is exactly the most load-bearing issue: the H≈1.3 value that distinguishes the critical regime is meaningful only if the integrated overlap is monoscaling and the DFA estimate is converged. The paper itself states the self-similarity assumption but never tests it, and the reported crossover in the critical-region DFA curves contradicts it. A bounded increment process with stationary increments cannot produce H>1, so the observed superdiffusive exponent is a red flag for nonstationarity or trend contamination. The concrete protocol of varying DFA order and using surrogates would settle whether the long-range temporal memory claim is real or an artifact. Since the reader already conditioned acceptance on addressing exactly this kind of issue, the CONDITIONAL verdict remains appropriate; no verdict change is needed from this stress-test pass.","tokens_in":12312,"tokens_out":5004,"duration_ms":50906,"concrete_test":"At a critical point (e.g., N=10, p=0.25), run R=20 independent noise realizations (new MNIST draws and noise seeds) for T=200000 steps. For each realization, compute Q_t and its integral X_t; estimate H with DFA order 1 on the same fitting range as in Fig. 4. Then (a) repeat with DFA orders 2 and 3, and (b) apply DFA order 1 to phase-randomized surrogates of Q_t that preserve the marginal distribution but destroy temporal correlations. If the median H drops from ≈1.3 toward 0.5 under higher-order detrending, or if shuffled surrogates also yield H≈1.3, then the reported long-memory critical regime is a nonstationarity/DFA artifact rather than genuine persistent temporal memory.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The key novel claim is the critical-regime long-memory exponent H≈1.3 from Eqs. (13)-(15). Equation (14) is well-posed only if X_t is monoscaling, and for a process with bounded increments (|Q_t|≤1) a fitted H>1 cannot arise from stationary increments; it indicates a nonstationary mean or trend within DFA windows. The linear detrending in Eq. (17) removes only a local linear trend from X_t; if Q_t dwells in one energy well for long intervals, X_t has a quadratic component, and the detrended fluctuation F(Δt) can grow superlinearly, mimicking H>1. The paper gives no stationarity test, no trajectory-length convergence check, no DFA-order variation, no surrogate test, and no error bars (single realization per (N,p)). The observed crossover in critical DFA curves is itself evidence against the monoscaling assumption needed for a single H. Thus the 'long-range correlated, persistent temporal memory' part of the central claim is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a stochastic version of the exponential modern Hopfield network (called SMHN in the body and SEDAM in the abstract) with multiplicative salt-and-pepper noise, trained on MNIST patterns. It defines the time-averaged overlap Q and the detrended fluctuation analysis (DFA) exponent H as order parameters and reports a noise-driven transition in both, with the critical noise level p_c decreasing with the number of stored patterns N. The central claim is that at critical noise levels near p_c≈0.23–0.3 the network enters a long-range correlated dynamical regime with H≈1.3, interpreted as highly persistent temporal memory. The model and simulation protocol are explicit, but the evidence for the H transition and for a genuine phase transition is not yet sufficient.","tokens_in":12638,"tokens_out":5470,"duration_ms":51492,"significance":"If the claims were established, this would be a valuable contribution: it would extend the phase-diagram literature of modern Hopfield networks from equilibrium thermodynamic analysis to out-of-equilibrium, noise-driven temporal dynamics, and it would demonstrate a load-dependent critical noise level in a realistic correlated dataset (MNIST). The manuscript is explicit about the model, the simulation parameters (L=784, 200,000 time steps, 25 noise values, loads up to 10,000), and it is methodologically interesting to use DFA on the overlap time series. The main novelty, however, is the claimed critical-regime exponent H≈1.3, and that result currently rests on an untested monoscaling assumption and is vulnerable to DFA trend artifacts. The comparison with Rademacher patterns and with a stochastic Hopfield network is announced in the abstract but is not actually reported in the body, which weakens the paper's stated scope.","major_comments":[{"comment":"The reported H≈1.3 in the critical regime is not supported because H>1 cannot arise from stationary increments of a bounded signal. Since |Q_t|≤1, the integrated variable satisfies |X(t+Δt)-X(t)|≤Δt, so any stationary-increment process has F(Δt) ≤ C Δt and hence H≤1. A fitted H≈1.3 therefore indicates a nonstationary mean or a local trend that the linear detrending of Eq. (17) does not remove; a long dwell in one energy well followed by a switch creates quadratic segments in X_t, which DFA-1 can mistake for superdiffusion. The manuscript provides no stationarity test, no surrogate data, no DFA-order variation, and no trajectory-length convergence check to rule out this artifact.","section":"Overlap as order parameter, Eqs. (13)–(15) and End Matter, Eq. (17)"},{"comment":"The authors state that the DFA function in the critical region 'always' shows a crossover between short- and long-time scaling, yet Eq. (14) defines H through a single monoscaling exponent. When a crossover is present, the reported 'H≈1.3' is not a well-defined scaling exponent unless the two fitting ranges are justified and the fits are shown to be stable; the paper does not report the fitting ranges, the goodness of fit, or how the short- and long-time regimes are separated. This is load-bearing because the long-memory claim is precisely the claim of a single long-range scaling regime.","section":"Results, Fig. 3 and accompanying text"},{"comment":"The abstract states that the MNIST-based SEDAM is compared with a SEDAM trained on Rademacher patterns and with a stochastic Hopfield network, and that the Rademacher-trained network gives H≈1.5. However, no section, figure, or table in the body reports these simulations, their parameter settings, or the comparison results. As written, the abstract's comparison claim is unsupported by the presented material.","section":"Abstract and body (comparison claims)"},{"comment":"The 'phase transition' in Q and H is inferred from a single realization per (N,p) on a fixed L=784 system with 200,000 time steps; there are no error bars, no averaging over initial conditions or pattern subsets, and no finite-size scaling or thermodynamic-limit analysis. Moreover, p_c is identified from the Q(p) curves and then the same p_c values are used to label the H anomaly in Fig. 4, so the coincidence of the two transitions is partly by construction. A finite-size analysis and an independent identification of p_c are needed before calling the regime 'critical' or describing the transition as first-order.","section":"Results, Figs. 1–4; 'phase transition' language"}],"minor_comments":[{"comment":"The abstract uses 'SEDAM' and 'load K', while the body uses 'SMHN' and 'load N'; please unify the terminology to avoid confusion.","section":"Notation throughout"},{"comment":"The paper should state how p_c is estimated from the Q(p) curves (e.g., threshold criterion, inflection point, or midpoint between high- and low-Q plateaus), because the reported p_c values are the basis for the subsequent H labeling.","section":"Results, Fig. 2 caption"},{"comment":"The notation G^2(t)=σ^2(t) reuses G(t) from Eq. (14) in a way that is potentially confusing; please define G(t) and σ^2(t) separately and consistently.","section":"End Matter, Eq. (15)"},{"comment":"There are several typos and stylistic slips, including 'Browian' for 'Brownian' and 'a one-time statistical average of of the overlap parameter'; these should be corrected.","section":"Discussion section"}],"recommendation":"major_revision","confidential_remarks":"The central novel result, the critical-regime H≈1.3 long-memory claim, is currently at risk of being a DFA artifact, and the paper also omits the announced Rademacher/SHN comparison. If the revision cannot provide the requested stationarity checks, surrogate tests, DFA-order variation, finite-size analysis, and actual comparison results, I would not recommend publication. The paper would also benefit from a data/code availability statement, since the numerical protocol is the main evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nHere is my read on Cafiso and Paradisi, arXiv:2509.17152.\n\nThe paper is worth a look if you follow associative-memory networks. It takes the exponential modern Hopfield model of Demircigil et al., adds a multiplicative salt-and-pepper noise, and studies the temporal dynamics of the overlap with the stored pattern, using time-averaged overlap Q and the DFA exponent H as order parameters. The model and simulation protocol are clearly described (L=784, loads up to N=10000, 25 noise values, 200k steps), so the numerical experiment is reimplementable. The qualitative result that Q drops from near 1 to near 0 over a fairly narrow noise window, with the transition shifting to lower p as N increases, is credible and is a useful data point.\n\nWhat is not established is the richer claim: that at criticality the overlap dynamics become long-range correlated with H≈1.3, i.e., persistent temporal memory, with sub- and super-critical regimes both diffusive at H≈0.5. The numerical support is too thin. There is no finite-size scaling or thermodynamic-limit analysis, no error bars, and from what is in the text a single realization per (N,p). The way p_c is read off the Q(p) curves is not defined, and then the same p_c values are used to label the H anomaly, so there is a mild circularity in the phase identification.\n\nThe H≈1.3 claim has a more specific problem. DFA with first-order detrending will happily return H>1 when the integrated signal contains a trend that is not removed by the local linear fit. Here Q_t is bounded by construction, so a stationary-increment process cannot give H>1; the observed H≈1.3 needs a stationarity check, a detrending-order check, surrogate data, and a check that the DFA crossover noted in the critical region is not itself violating the monoscaling assumption behind Eq. (14). The stress-test note is right: without those controls, the persistent-memory claim could be a DFA artifact.\n\nAlso, the abstract promises comparisons with Rademacher patterns and with a stochastic Hopfield network; the body I have only shows MNIST results. If those comparisons exist, they need to be in the main text; if not, the abstract overstates the scope.\n\nThe citation pattern is fine—Demircigil, Lucibello and Mezard, Amit et al., and related work are cited. No circular derivation is present; H is an independent statistic of the overlap series.\n\nVerdict: a promising numerical study, not a confirmed phase transition. With multi-realization statistics, finite-size scaling, and proper DFA controls, it could become a solid contribution. It deserves a serious referee rather than a desk reject. I would not cite it as evidence for criticality in its current form.","headline":"Plausible Q transition and a cleanly specified stochastic modern Hopfield model, but the H≈1.3 long-memory claim rests on DFA fits that are not separated from trend artifacts.","tokens_in":13043,"tokens_out":3753,"would_cite":false,"duration_ms":31939,"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":"Adding random pixel flips to an exponential associative-memory network produces a sharp noise-driven critical transition in which recall drops and the overlap dynamics acquire persistent long-range memory (H≈1.3).","keywords":["dense associative memory","exponential interaction function","criticality","phase transition","salt-and-pepper noise","detrended fluctuation analysis","anomalous diffusion","MNIST"],"falsifier":"Take the N=1000 run at p≈p_c, split the 200,000-step overlap series into two independent halves, and re-estimate H from each half over a range of detrended-fluctuation-analysis window sizes; if the two halves disagree by more than the reported uncertainty, or if the local log–log slope drifts monotonically with window length instead of plateauing, then the self-similarity assumption fails and H≈1.3 cannot be interpreted as a well-defined exponent.","tokens_in":12090,"feed_emoji":"🧠","tokens_out":9722,"duration_ms":82085,"temperature":0.7,"pith_summary":"This paper claims that a stochastic associative-memory network with an exponential interaction function—a design with very high storage capacity—undergoes a genuine noise-driven phase transition when trained on realistic correlated images (MNIST). Taking the salt-and-pepper flip probability p as the control parameter, the time-averaged recall overlap Q drops sharply at a critical value p_c that falls as the number of stored patterns grows, from about 0.3 at low load to about 0.23 at N=10000. At the same p_c, the overlap time series changes character: below and above the transition its fluctuations are indistinguishable from normal Brownian motion (H≈0.5), while in the critical window they become strongly persistent, with anomalous diffusion exponent H≈1.3 on MNIST and about 1.5 on random Rademacher patterns. If right, this establishes that high-capacity associative memories have a finite-noise critical region in which the network keeps a long, non-Markovian temporal memory, and that this criticality is governed by the interplay of stored-pattern structure and noise rather than by thermal equilibrium.","feed_headline":"Noise p≈0.25 gives memory network persistent memory","feed_subtitle":"Recall overlap drops sharply and the time series turns long-range correlated, H≈1.3.","key_machinery":"The load-bearing object is the stochastic update rule of Eq. (11): the state of each neuron at time t is the sign of a difference between two sums of exponentials of overlap with all stored patterns, and this sign is multiplied by an independent dichotomous random variable that flips the neuron with probability p. The exponential interaction function F(z)=exp(z) inside the energy E=−Σ_i F(x_i^T S) is what gives the model its exponential storage capacity; the multiplicative salt-and-pepper noise turns retrieval into a stochastic process in the overlap Q_t. Two order parameters are then read off Q_t: the time average Q (one-time statistics, recall quality) and the DFA scaling exponent H (two-time statistics, temporal memory), defined through G(t)∼t^H on the integrated overlap X_t. The claimed mechanism is that stored-pattern connectivity and noise compete: at low p the network sticks near one energy well (retrieval but Markovian fluctuations), at high p noise dominates (random walk), and only in between do the residual fluctuations develop long-range correlations.","core_discovery":"In the paper's own terms, the central result is the emergence of a critical transition in both the time-averaged overlap Q and the diffusion scaling H, with the critical noise level p_c decreasing as the load N increases. For each load, H marks a transition between a sub-critical and a super-critical regime, both with short-range correlated dynamics, while in the critical regime, found in the range p_c ≈ 0.23–0.3, the MNIST-based network displays long-range correlated dynamics with highly persistent temporal memory, H ≈ 1.3 (the abstract quotes 1.25). Rademacher-trained networks show the same qualitative behavior with a slightly higher memory index H ≈ 1.5. The paper presents Q and H as two order parameters and emphasizes that no equilibrium condition is assumed: the system is run out of equilibrium with a multiplicative noise that acts like random pixel errors.","pith_inferences":["If the reported H≈1.3 genuinely reflects a power-law scaling, then the overlap time series at p_c should show a 1/f-type power spectrum and algebraically decaying autocorrelation; measuring those directly on the same simulations would be a sharper check than DFA alone.","A testable extension the paper does not perform is an adiabatic sweep of p upward and downward: because the H-jump looks first-order-like, hysteresis in Q or H would indicate a discontinuous transition with practical consequences for noise-tolerant retrieval.","The Q-and-H protocol is a new observable for associative-memory criticality and could be applied to other interaction functions (polynomial n-body, softmax attention) to see whether long-range temporal memory at criticality is unique to the exponential choice or generic to high-capacity memories.","If the critical long-memory regime has a functional role, a network deliberately operated at p_c could serve as a temporal-memory buffer or a source of slow fluctuations; the authors do not claim this, but it is a plausible engineering consequence."],"forward_implications":["Storage capacity and noise tolerance are dynamically coupled: p_c drops steeply from about 0.3 at N=5 to about 0.25 at N=100, then slowly to about 0.23 at N=10000, so larger memories tolerate less pixel-flip noise before losing recall.","In the critical window the overlap time series is non-Markovian and persistent (H≈1.3), meaning retrieval errors are not independent events but carry long temporal correlations.","Sub-critical and super-critical phases both show H≈0.5 Brownian, Markovian fluctuation dynamics, so the transition's special character appears only at criticality rather than as a gradual change in memory.","The same p_c appears in both order parameters Q and H, giving a clean dynamical signature of criticality in an out-of-equilibrium associative memory.","Pattern correlations matter: MNIST (correlated patterns) yields H≈1.3 while Rademacher (uncorrelated patterns) yields H≈1.5, showing the long-memory exponent is sensitive to the statistical structure of stored items."],"supporting_citations":[{"why":"The original two-body associative-memory model with Hebbian weights that the exponential update rule generalizes.","marker":"[1]"},{"why":"The equilibrium statistical-mechanics treatment introducing temperature and the storage threshold, the baseline this non-equilibrium study departs from.","marker":"[6]"},{"why":"Introduces the pattern-sum energy E=−Σ_i F(x_i^T S) underlying the modern dense-associative-memory update.","marker":"[10]"},{"why":"Gives the n-body dense-associative-memory capacity scaling N=αL^{n−1} against which exponential capacity is contrasted.","marker":"[12]"},{"why":"Defines the exponential interaction F(z)=exp(z) and the noiseless update rule that the authors modify with random flips.","marker":"[14]"},{"why":"Provides the retrieval/spin-glass phase diagram and the overlap order-parameter convention used for Q.","marker":"[47]"},{"why":"Recent analysis of storage and critical behavior in the exponential model, the specific setting this paper extends to noise-driven dynamics.","marker":"[49]"},{"why":"The MNIST dataset used for training, whose correlated binary images provide the pattern structure in the main experiments.","marker":"[54]"},{"why":"The detrended fluctuation analysis that defines G(t) and the scaling relation used to estimate H.","marker":"[56]"}],"fun_headline_variants":["Critical noise gives Hopfield net persistent time-memory","Noise-tuned criticality yields long-term memory in Hopfield net","Stochastic Hopfield net shows persistent memory at critical noise","Phase transition in noise gives Hopfield net temporal persistence","H≈1.3 at critical noise: neural net retains memory over time"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim rests on the overlap signal having one consistent scaling law at long times; if the critical-region signal shifts its character over the run, the measured H≈1.3 would be an artifact rather than a real exponent.","fun_headline_variants_meta":{"raw":{"variants":["Critical noise gives Hopfield net persistent time-memory","Noise-tuned criticality yields long-term memory in Hopfield net","Stochastic Hopfield net shows persistent memory at critical noise","Phase transition in noise gives Hopfield net temporal persistence","H≈1.3 at critical noise: neural net retains memory over time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000907,"raw_usage":{"total_tokens":3949,"prompt_tokens":1040,"completion_tokens":2909,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":656,"completion_tokens_details":{"reasoning_tokens":2823}},"tokens_in":656,"tokens_out":2909,"duration_ms":21988,"temperature":1.0,"reasoning_tokens":2823,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:48:21.638995+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the N=1000 run at p≈p_c, split the 200,000-step overlap series into two independent halves, and re-estimate H from each half over a range of detrended-fluctuation-analysis window sizes; if the two halves disagree by more than the reported uncertainty, or if the local log–log slope drifts monotonically with window length instead of plateauing, then the self-similarity assumption fails and H≈1.3 cannot be interpreted as a well-defined exponent.","supporting_citations":[{"cited_title":"catastrophic forgetting","cited_arxiv_id":null,"evidence_quote":"The original two-body associative-memory model with Hebbian weights that the exponential update rule generalizes."},{"cited_title":"On the contrary, we here limit to show the case of synchronous update given in Eq","cited_arxiv_id":null,"evidence_quote":"The equilibrium statistical-mechanics treatment introducing temperature and the storage threshold, the baseline this non-equilibrium study departs from."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the pattern-sum energy E=−Σ_i F(x_i^T S) underlying the modern dense-associative-memory update."},{"cited_title":"Gardner, Nuclear Physics, Section B257, 747 – 765 (1985)","cited_arxiv_id":null,"evidence_quote":"Gives the n-body dense-associative-memory capacity scaling N=αL^{n−1} against which exponential capacity is contrasted."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the exponential interaction F(z)=exp(z) and the noiseless update rule that the authors modify with random flips."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the retrieval/spin-glass phase diagram and the overlap order-parameter convention used for Q."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Recent analysis of storage and critical behavior in the exponential model, the specific setting this paper extends to noise-driven dynamics."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The MNIST dataset used for training, whose correlated binary images provide the pattern structure in the main experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The detrended fluctuation analysis that defines G(t) and the scaling relation used to estimate H."}],"review_version":2}