{"id":"b98e752f-e315-4853-aa39-2b0f911ddde9","arxiv_id":"2501.13997","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An energy-based recurrent state-space model with a continuous attractor memory predicts sensory observations after actions using local Hebbian learning, matching several ML baselines.","lead":"This paper builds a brain-inspired predictive model that combines an energy-based neural network with a continuous attractor memory to guess what the eyes and body will see after an action. It reports that this biologically plausible model predicts images and video frames about as well as several standard machine learning world models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equations (14) and (17) silently replace an expectation over s^l with a point evaluation, so the local Hebbian update (Eq. 15) is not the gradient of the stated prediction bound; the theoretical link between the upper bound L and the implemented learning rule is therefore not established.","rationale":"The reader's weakest assumption is the same point: the step from an expectation to a point evaluation is a silent approximation. I agree that this is the most load-bearing flaw because the paper's contribution explicitly includes establishing a prediction-error upper bound and deriving a local Hebbian rule from it. If the equality is false, the learning rule is not optimizing the stated bound, so the theoretical bridge from Eq. (3) to the implemented algorithm is missing. The concern is addressable: the authors could state that Eq. (15) is a single-sample stochastic gradient, or derive the correct gradient including the variance term, or justify a delta prior. The empirical comparisons may still be valid as demonstrations of the architecture's predictive ability, so a conditional verdict rather than rejection is appropriate. No other objection seems more load-bearing; the lack of code and error bars is a reproducibility issue but secondary to the theoretical gap.","tokens_in":15462,"tokens_out":12806,"duration_ms":109549,"concrete_test":"Run a fixed-parameter forward pass with realistic trained values and compute both sides of Eq. (14) for layer 0: draw M=10^4 samples s1^(i) ~ N(theta^1 f(s^2_hat), (Lambda^1)^-1) and form the Monte Carlo estimate (1/M) sum_i 1/2 (s0 - theta^0 f(s1^(i)))^T Lambda^0 (s0 - theta^0 f(s1^(i))) + C; compare it to the claimed point value 1/2 (s0 - theta^0 f(s1_hat))^T Lambda^0 (s0 - theta^0 f(s1_hat)) + C. If the relative difference exceeds a few percent, the equality in Eq. (14) is false under the stated expectation. In the same run, compare the Monte Carlo gradient (1/M) sum_i Lambda^0 (s0 - theta^0 f(s1^(i))) f(s1^(i))^T with the right-hand side of Eq. (15); a material discrepancy shows the local rule is not the gradient of the defined bound.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing issue is the claimed identity in Eqs. (14) and (17). Each L^l_t is defined as an expectation over s^l_t ~ p(s^l_t|s^(l+1)_t) of a negative log-likelihood, but it is then set equal to a single point-evaluation 1/2 (e^l_t)^T Lambda^l e^l_t + C^l, with e^l_t = s^l_t - theta^l f(s^(l+1)_t). For a Gaussian conditional, the expectation contains an additional positive variance term that is not constant and does not vanish except for a delta prior. The equality therefore requires either a one-sample Monte Carlo replacement or a delta prior; neither is stated. Because Eq. (15) is presented as the gradient of this bound, the local Hebbian update is not in fact the gradient of the stated objective L^l_t. In addition, even under a Monte Carlo reading, the algorithm uses prior samples for layer 0 but posterior samples for higher layers, so the stochastic-gradient interpretation is not uniform. The claimed link between the upper bound L in Eq. (3) and the layer-wise local rules is thus not established; this does not by itself invalidate the empirical results, but it removes the theoretical justification for the learning rule.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an energy-based recurrent state-space model (RSSM) with a hierarchical Gaussian EBM and a continuous attractor neural network (CANN) memory. Prediction, learning, and inference are implemented via neural dynamics and local Hebbian updates, without backpropagation. Experiments cover eye movement on CIFAR-10/Fashion-MNIST, navigation in DeepMind Lab and Google Street View, and static observation on MNIST-rot/TaxiBJ, with comparisons to TransDreamer, tPCN, VPN, FRNN, and ST-ResNet. The central theoretical claim is that the layer-wise local update rules are gradients of a prediction-error upper bound, connecting the objective in Eq. (3) to the implemented algorithm.","tokens_in":1442,"tokens_out":1511,"duration_ms":83823,"significance":"The framework is timely and the breadth of tasks is a strength. If the theory were sound, the paper would offer a useful biologically plausible alternative to BPTT-based world models and a concrete link between predictive coding, EBMs, and CANN memory. The authors provide explicit algorithms (Algorithms 1 and 2), a correct Jensen upper bound in Eq. (3), and a genuine attempt at local Hebbian learning. However, the theoretical justification of the local learning rule contains a load-bearing gap, the CANN is underspecified, and the empirical comparisons lack error bars and code/data. The empirical results are suggestive but do not establish the theoretical claims.","major_comments":[{"comment":"The equalities in these equations are not derivable from the definitions. Each L_t^l is defined as an expectation over a stochastic latent variable (s_t^1 under the prior in Eq. (14), and s_t^l under p(s_t^l | hat s_t^{l+1}) in Eq. (17)) of a negative log-likelihood, but the right-hand sides evaluate the quadratic form at a single point, replacing the expectation by its integrand. For a Gaussian conditional with precision Lambda^l, the expectation of the quadratic form differs from its value at the mean by a positive variance contribution, and the difference is amplified by the nonlinearity f(.). The equality would require either a delta prior or an explicit one-sample Monte Carlo approximation, and neither is stated. Also, in Eq. (17) the right-hand side uses the prior error ehat_t^l = s_t^l - theta^l f(hat s_t^{l+1}), while the left-hand side is the expected negative log-likelihood of the lower layer; the layer indices do not match. Since Eq. (15) is derived as the gradient of this bound, the local Hebbian update is not established as the gradient of the stated objective, and the claimed link between the upper bound L in Eq. (3) and the implemented learning rule is not established.","section":"Section 4, Eqs. (14) and (17)"},{"comment":"The CANN is the memory component that drives the prior p(s_t^L | m_t) and enables imagination, but its dynamics are underspecified: m_t is defined as H(I_t) with H(.) never defined, and the main text states that the transition m_{t+1} is a Dirac delta while Algorithm 2 and Appendix C describe a continuous ODE with adaptation. The relationship between the steady state of Eqs. (36)-(37) and the memory update in Eq. (8) must be stated, and H(.), W, alpha, beta need precise definitions, or the implementation cannot be reproduced as written.","section":"Appendix C, Eq. (35)"},{"comment":"The empirical claim that the model 'matches the performances of machine learning methods' is supported only by point estimates of MSE. No error bars, number of seeds, or statistical tests are reported, and in Table 2 the differences between OURS and tPCN at long sequence lengths are small (e.g., 0.0087 vs 0.0151 for unseen sequences of length 1024). Without repeated-run variability, the comparisons in Tables 1-3 do not establish the claimed equivalence. The absence of code or data release (Appendix E) further limits verification.","section":"Section 5, Tables 1-3 and Figures 3-6"}],"minor_comments":[{"comment":"The column header 'T PC O URS' appears to be a typo for 'tPCN' and 'OURS'; this should be corrected.","section":"Table 2"},{"comment":"The text and figure panels are inconsistent: the text refers to subfigure (d) for the layer-wise loss, but the figure panels show different arrangements of MSE versus epochs, initialized patches, and neuron numbers; the caption should be rewritten to match the actual panel labels.","section":"Section 5, Figure 3(c)-(e)"},{"comment":"The derivation of the mutual information lower bound is not written in a standard form; Eq. (19) is not a valid inequality as stated, and the notation mixes p and q without clear definitions. Since the upper bound in Eq. (3) already follows from Jensen's inequality, this appendix should be corrected or removed.","section":"Appendix A, Eqs. (19)-(26)"},{"comment":"The hyperparameter table has formatting issues (for example, the entry '1253' appears to be a line break inside '125 3'), and no random seed is reported; a cleaned table with seeds and unambiguous column separators would improve reproducibility.","section":"Appendix E, Table 4"},{"comment":"The statement that code 'will be open-sourced after publication' is not a substitute for an availability statement; for a journal submission, a repository link or a clear reason for withholding should be provided.","section":"Appendix E"}],"recommendation":"major_revision","confidential_remarks":"The theoretical gap in Eqs. (14)/(17) is the main obstacle: it is local and fixable by restating the objective as a one-sample approximation or by adding the missing variance term, so I recommend major revision rather than rejection. The underspecified CANN and missing error bars are also important. The manuscript may be better suited to a venue combining computational neuroscience and machine learning; for a pure ML journal, the lack of reproducible code and statistical comparisons is a notable weakness."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You'll want to know about arXiv:2501.13997. It proposes an energy-based recurrent state-space model (E-RSSM) for predictive learning: a hierarchical Gaussian EBM as encoder-decoder, a CANN as temporal memory, and a 'learning before inference' objective that is an upper bound on prediction cross-entropy. The ingredients are all known, but the combination is new: no backprop or BPTT, only local Hebbian updates via error neurons. On the evidence in the paper, the architecture works across four task families (eye movement, navigation, head turning, video prediction), and it matches or beats a few ML baselines. That is a genuine empirical contribution.\n\nThe paper is a mixed bag. The experimental section is the strongest part: four different action types, multiple datasets, and MSE comparisons. But there are no error bars, no multiple seeds, no code released, and the CANN is underspecified — Appendix C gives equations but never defines the nonlinearity H, and the recurrent weight matrix W is said to be 'random low-rank' without details. You could not reproduce the results from the text.\n\nThe theoretical centerpiece has a real gap. Equations (14) and (17) claim that the layer-wise prediction bound L^l_t — an expectation over s^l drawn from a Gaussian conditional — equals a single-point quadratic form 1/2 (e^l)^T Λ^l e^l + C. That equality does not hold for a non-delta prior; the expectation adds a variance term. The authors likely intended a one-sample Monte Carlo approximation, but they write it as exact. Worse, the algorithm uses prior samples for layer 0 and posterior samples for higher layers, so even the stochastic-gradient story is not uniform. This does not kill the empirical results, but it means the local Hebbian rule is not actually derived as the gradient of the stated upper bound L. That's a load-bearing flaw in the theory, though fixable with honest phrasing and a proper derivation.\n\nAlso note Eq. (5) uses contrastive divergence with a negative phase, but for Gaussian conditionals the partition function is constant, so the negative phase vanishes. The rule they end up with is correct Hebbian learning, but the derivation is padded.\n\nWho is this for? Computational neuroscientists and people working on biologically plausible world models. The architecture idea is worth discussing, and the experiments suggest it works. It deserves a serious referee — the core idea is promising and the empirical data is suggestive. But the referee should demand a corrected theoretical derivation, code, and a complete CANN specification.\n\nMy recommendation: send it to peer review, and in your report ask for the Monte Carlo approximation to be made explicit and the missing details to be supplied.\n\nBest,\n\n[Your name]","headline":"Interesting biologically plausible world model with a genuine architecture contribution, but the theoretical derivation of the local learning rule has a gap that needs fixing.","tokens_in":16295,"tokens_out":5903,"would_cite":false,"duration_ms":50966,"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 biologically plausible energy-based model predicts next observations with local Hebbian learning, matching backprop baselines in several tasks.","keywords":["energy-based models","recurrent state-space models","predictive learning","continuous attractor neural networks","Hebbian learning","hierarchical generative models","predictive coding","world models"],"falsifier":"On a trained model, compute the layer-wise loss of Eq. (17) under the exact expectation $\\mathbb{E}_{s^l_t \\sim p(s^l_t|s^{l+1}_t)}[-\\log p(s^{l-1}_t|s^l_t)]$ with many Monte Carlo samples, and compare the resulting gradient to the Hebbian update of Eq. (15); if they differ appreciably, the local rule is not minimizing the stated prediction bound, or, alternatively, train a copy of the model with multi-sample gradient estimates and compare prediction error on the same benchmarks.","tokens_in":15261,"feed_emoji":"🧠","tokens_out":8572,"duration_ms":63427,"temperature":0.7,"pith_summary":"The paper proposes a recurrent state-space model built from energy-based components as a hypothesis for how the brain predicts its next visual observation after acting. The model draws latent state samples from a hierarchical Gaussian energy-based network and updates connection weights with local Hebbian rules, so no backpropagation is used. A continuous attractor neural network acts as the memory that compresses past observations and actions into a low-dimensional state. The authors report that this biologically plausible model predicts visual observations in eye movement, navigation, head turning, and changing-scene tasks, with performance comparable to machine learning world models in several experiments.","feed_headline":"Energy-based world model predicts next frames without backprop","feed_subtitle":"Hierarchical energy-based network with attractor memory matches backprop baselines in predicting after actions.","key_machinery":"The central machinery is a hierarchical Gaussian energy-based model coupled to a continuous attractor neural network (CANN), a recurrent network whose stable states form a continuous manifold for encoding variables such as position or head direction. Each layer $l$ models $p_\\theta(s^l_t|s^{l+1}_t)$ as a Gaussian with mean $\\theta^l f(s^{l+1}_t)$ and precision matrix $\\Lambda^l$, so prediction errors appear in dedicated error neurons $e^l_t = s^l_t - \\theta^l f(s^{l+1}_t)$. The layer-wise prediction bound $\\frac{1}{2}(e^l_t)^\\top \\Lambda^l e^l_t + C$ converts into a Hebbian gradient rule, while Langevin dynamics implement sampling-based inference of the posterior. The CANN supplies the top-layer prior $p(s^L_t|m_t)$ and updates its state $m_t$ from the inferred top-layer activity and the action, enabling closed-loop prediction and imagination over multiple time steps.","core_discovery":"The central claim is that prediction after an action can be implemented as an energy-based recurrent state-space model whose joint distribution factorizes as $p_\\theta(o_t, s_t|m_t) = p_\\theta(o_t|s_t)p(s_t|m_t)$, with learning driven by the upper bound $L = -\\mathbb{E}_{o\\sim p_{\\rm true}}\\mathbb{E}_{\\hat{s}\\sim p(s|m)} \\log p_\\theta(o|\\hat{s})$ on the prediction cross-entropy. In the hierarchical Gaussian implementation, each layer's prediction bound takes the form $\\frac{1}{2}(\\hat{e}^l_t)^\\top \\Lambda^l \\hat{e}^l_t + C$ with error neurons $\\hat{e}^l_t = \\hat{s}^l_t - \\theta^l f(\\hat{s}^{l+1}_t)$, and gradient descent on this bound produces a purely local Hebbian update. The memory is a continuous attractor neural network that keeps a compressed record of past states and actions. The paper presents this architecture as a biologically plausible way for the brain to perform predictive learning, and supports it with experiments on several visual prediction tasks.","pith_inferences":["Editorial: the layer-wise bound in Eq. (17) is treated as an equality with an expectation, so the Hebbian update is effectively a single-sample gradient estimate; a multi-sample Monte Carlo version of the same objective might reveal how much performance depends on this approximation.","Editorial: since the paper does not ablate the CANN, one test the authors did not run is replacing the attractor memory with a non-attractor recurrent network of equal size; if prediction degrades, the attractor structure is doing the memory work claimed.","Editorial: the approach suggests static predictive coding networks can be extended to temporal, action-conditioned prediction, potentially linking the predictive coding and world model literatures."],"forward_implications":["Prediction in this class of world models can be learned fully online with local synaptic updates, eliminating backpropagation and backpropagation through time.","Because the generative model is Markovian, the framework can be extended to model-based reinforcement learning by adding a reward signal, as the authors note.","The attractor memory provides a natural substrate for cognitive maps, connecting the predictive model to hippocampal place-cell and grid-cell phenomena.","If the bound holds, predictive coding becomes a normative implementation of energy-based prediction rather than an ad hoc approximation of backpropagation.","The model gives a concrete, testable account of prediction, learning, and inference as neural dynamics that run in real time."],"supporting_citations":[{"why":"Defines the recurrent state-space model that this work recasts in energy-based form.","marker":"Hafner et al., 2019b"},{"why":"Supplies the hierarchical energy-based model and neural sampling method used for inference and learning.","marker":"Dong & Wu, 2023"},{"why":"Introduces predictive coding and error neurons that make the learning rule local.","marker":"Rao & Ballard, 1999"},{"why":"Shows predictive coding can approximate backpropagation with Hebbian plasticity, grounding the local-learning claim.","marker":"Whittington & Bogacz, 2017"},{"why":"Foundational continuous attractor neural network model underlying the memory module.","marker":"Amari, 1977"},{"why":"Provides the CANN dynamics used to implement memory updates.","marker":"Wu et al., 2008"},{"why":"TransDreamer baseline compared in the eye-movement prediction experiments.","marker":"Chen et al., 2022"},{"why":"Temporal predictive coding baseline compared on the MNIST-rot sequence task.","marker":"Tang et al., 2023"}],"fun_headline_variants":["Attractor energy model predicts next observations without backprop","Hierarchical EBM with attractor memory predicts after actions","Energy-based model with attractor memory learns Hebbian-style","Attractor-memory EBM matches backprop baselines in prediction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the layer-wise prediction loss equals the full expectation over the latent posterior, so that a single latent sample per observation makes the local Hebbian update the exact gradient of the objective.","fun_headline_variants_meta":{"raw":{"variants":["Attractor energy model predicts next observations without backprop","Hierarchical EBM with attractor memory predicts after actions","Energy-based model with attractor memory learns Hebbian-style","Attractor-memory EBM matches backprop baselines in prediction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000776,"raw_usage":{"total_tokens":3437,"prompt_tokens":953,"completion_tokens":2484,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":569,"completion_tokens_details":{"reasoning_tokens":2415}},"tokens_in":569,"tokens_out":2484,"duration_ms":17138,"temperature":1.0,"reasoning_tokens":2415,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:49:40.896564+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a trained model, compute the layer-wise loss of Eq. (17) under the exact expectation $\\mathbb{E}_{s^l_t \\sim p(s^l_t|s^{l+1}_t)}[-\\log p(s^{l-1}_t|s^l_t)]$ with many Monte Carlo samples, and compare the resulting gradient to the Hebbian update of Eq. (15); if they differ appreciably, the local rule is not minimizing the stated prediction bound, or, alternatively, train a copy of the model with multi-sample gradient estimates and compare prediction error on the same benchmarks.","supporting_citations":[],"review_version":1}