{"id":"9bc25295-52a4-4ebf-9e0e-4b545615e3ff","arxiv_id":"2508.06501","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"A canonical microcircuit neural-ODE architecture is trained on MNIST and CIFAR-10, but its headline accuracy numbers are inconsistent across sections.","lead":"This paper turns a standard neuroscience model of a cortical microcircuit into a trainable neural ODE and tests it on image classification. The idea is promising for parameter-efficient AI, but the paper's accuracy claims conflict across sections and lack detailed baselines.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's headline accuracy is internally inconsistent: the abstract's 97.8% single-node MNIST figure appears nowhere in the experiments, where Sections 6.2 and 6.3 report 86.4%/86.5% for one node and 95.2%/99.2% for four; the central empirical claim is thus unverifiable.","rationale":"The paper's stated aim is to demonstrate that biologically derived CMC dynamics, formulated as neural ODEs, deliver high accuracy and parameter efficiency on vision benchmarks. For this claim to hold, the experimental results must be internally consistent and reproducible. The most insecure link in that chain is the reporting of the headline numbers. On one hand, the paper deserves credit for making code publicly available and for describing the model equations (Eqs. 1–4) and the solver choices (Sections 5.1 and 6.1); these are concrete, inspectable components. On the other hand, an internal contradiction in the reported accuracies is a direct, verifiable flaw that does not depend on any outside benchmark or disputed methodology. The reader's weakest-assumption—that the gains could come from the learned front-end rather than the CMC dynamics—is a genuine concern, and the supplementary ablations actually provide partial evidence that recurrent inter-node connections matter (feedforward-only variants underperform). But the more fundamental problem is that the paper never states a single unambiguous MNIST accuracy for a given configuration: the abstract's 97.8% is unsupported by §6.2 (86.4%) and §6.3 (86.5%). This is not a matter of reinterpreting results; it is a direct internal inconsistency in the central evidence. A reviewer cannot verify, even in principle, which number the authors intended to report. Thus the rejection is warranted, and the concrete test—re-running the pinned code—would settle whether the concern is merely a reporting error or a deeper problem with the experiments. If the code reproduces, say, 86.5% for one node and 99.2% for four, the abstract still needs correction; if it reproduces 97.8%, then Sections 6.2/6.3 need correction. Either way, the current manuscript is not internally coherent. I therefore agree with the reader's REJECT verdict, while differing slightly on which assumption is weakest: the reproducibility contradiction is more load-bearing than the missing control, though both are legitimate issues.","tokens_in":13574,"tokens_out":5225,"duration_ms":44685,"concrete_test":"Run the released repository (https://github.com/NeurotrustAI/CMC) at a pinned commit (none is currently identified) with the Section 6.1 settings (Adam, batch size 128, cosine annealing, Euler step 0.05, integration [0,1] with 6 time points) and record 1-, 4-, and 5-node MNIST test accuracies over n=3 independent trials. Then compare against the four reported values: 97.8% (abstract), 99.8% (Section 5.2), 86.4%/95.2% (Section 6.2), and 86.5%/99.2% (Section 6.3). If the reproduced 1-node accuracy is not 97.8%—or if the code cannot be run because it is unversioned, incomplete, or missing dependencies—the abstract's central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that a single CMC node achieves 97.8% on MNIST and that hierarchical configurations improve performance with orders-of-magnitude fewer parameters—is not supported by the paper's own experimental reports. Section 5.2 states that a 5-region model reaches 99.8% on MNIST; Section 6.2 reports 1-node accuracy at 86.4% and 4-node at 95.2%; Section 6.3 reports 1-node at 86.5% and 4-node at 99.2%; Figure 4's MNIST curves appear to plateau near 93–94% for all 1–5 node configurations. These numbers cannot all describe the same experimental setup, yet no reconciliation, errata, or confidence intervals are provided. The 97.8% figure in the abstract does not correspond to any number reported in the experiments, and the claimed 'orders of magnitude fewer parameters' comparison lacks a tabulated baseline. The supplementary ablations add a further tension, noting that the full model was outperformed by each ablation for a single V1 node, which contradicts the narrative that each component is essential. Consequently, the central empirical result is unverifiable from the manuscript as submitted. This is a correctness risk, not merely a clarity issue: without a single consistent accuracy value for the 1-node configuration, the headline claims about CMC dynamics' efficiency and scaling behavior cannot be assessed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a neural ODE implementation of canonical microcircuits (CMCs) for image classification. The architecture combines a learnable retinal preprocessing front-end, state-initialization projection, one or more CMC nodes with eight-dimensional dynamics derived from Bastos et al., learnable inter-node connectivity, and a learned readout. The authors report MNIST and CIFAR-10 results for one to five nodes, claim orders-of-magnitude parameter efficiency relative to conventional deep networks, and present phase-space trajectories as evidence of interpretable class-specific dynamics.","tokens_in":13906,"tokens_out":4779,"duration_ms":39923,"significance":"If the empirical claims held, this would demonstrate that a biologically derived recurrent ODE module with roughly 150K parameters can match much larger convolutional or transformer models on standard vision benchmarks while offering interpretable dynamics. The paper's strengths include the use of established CMC dynamics, a clear biological motivation, and an exploratory ablation study. However, the headline accuracy figures are internally inconsistent across sections, no matched control isolates the effect of the CMC ODE, and the parameter-efficiency comparison lacks a quantitative baseline, so the significance cannot be assessed from the manuscript as submitted.","major_comments":[{"comment":"The abstract's headline claim that a single CMC node achieves 97.8% accuracy on MNIST is not reported anywhere in the experimental sections: Section 6.2 reports 86.4% for one node and 95.2% for four nodes, Section 6.3 reports 86.5% and 99.2% for the same two configurations, Section 5.2 reports 99.8% for a five-region model, and the learning curves in Figure 4 appear to plateau near 93-94% for all node counts. These numbers cannot all describe the same experimental setup, and the manuscript provides no reconciliation or confidence intervals, so the central empirical claim is unverifiable.","section":"Abstract; Sections 5.2, 6.2, 6.3; Figure 4"},{"comment":"Every trained model contains a learnable retinal preprocessing front-end, a learned state-initialization projection, a learned readout, and learnable inter-node connectivity matrices, yet the paper never compares against a matched control in which the CMC ODE is replaced by an ordinary feedforward or linear map with the same surrounding modules. Without such a control, the accuracy and parameter-efficiency gains cannot be attributed to the microcircuit dynamics, which is the central claim of the paper.","section":"Sections 4.1, 5.2, 5.3"},{"comment":"The claimed 'orders of magnitude fewer parameters' comparison is not substantiated: no table lists baseline architectures with their parameter counts and accuracies, and the reported parameter count for the model is inconsistent (Section 5.3 says a 5-region model has approximately 150,000 parameters, while Section 6.3 attributes 150K parameters to the 4-node model). This prevents any quantitative assessment of the parameter-efficiency claim.","section":"Section 5.3; Section 6.3; Figure 1"},{"comment":"The supplementary ablation section states that 'the original model was outperformed by each of the ablation tests for a single V1 node,' yet Section 7 summarizes the ablations as showing that 'both architecture and neural population dynamics contribute substantially to overall performance.' These statements are contradictory, and the single-node result in particular undermines the claim that each cell type is essential to the CMC's performance.","section":"Supplementary Material, Section A; Section 7"},{"comment":"The numerical solver specification is internally inconsistent: Section 5.1 says the integration time span is [0,1] with 6 time points and a step size of 0.15, while Section 6.1 says Euler integration with a step size of 0.05 was used. The relation between '6 time points' and the time span is not defined, so the experiments cannot be reproduced exactly from the description.","section":"Sections 5.1 and 6.1"}],"minor_comments":[{"comment":"The text says 'Figure 2 shows test accuracy across datasets' but Figure 2 depicts laminar structure; the accuracy curves appear in Figure 4.","section":"Section 6.2"},{"comment":"The text refers to 'Figure 4' when describing tracked voltage evolution during training, but Figure 4 contains learning curves; the phase portraits are in Figure 5 and Figure 6, and the voltage dynamics are in Figure 3. The cross-references need correction.","section":"Section 6.4"},{"comment":"The paper refers to 'Section 3.1' for the ODE dynamics, but the equations appear under the malformed subsection heading '3.0.1 Neural Dynamics'; the section numbering should be repaired.","section":"Section 3.0.1 and Section 3.1"},{"comment":"The lower portion of Figure 3 appears to reproduce the abstract of Chen et al. (2018) and a 'Model Performance' bar chart without visible attribution or discussion; if this is an unintended insertion, it should be removed, and if intentional, proper permission and citation are required.","section":"Figure 3"},{"comment":"There are several typographical errors, including 'acheived' in Section 6.3 and 'feed foward' in the supplementary material; a careful proofread is needed.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"The internal inconsistency in the headline accuracy figures is severe enough that I cannot recommend a normal revision path; the authors would need to re-run and re-report experiments with matched controls and reconcile the numbers in the abstract, Sections 5-6, and the supplementary material. The checklist's claim that error bars are reported (item 7) is not supported by the figures shown in the manuscript, and the apparent embedded reproduction of Chen et al.'s abstract within Figure 3 raises a presentation and copyright concern that should be addressed in any future submission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know: the paper has a genuinely new combination—taking Bastos et al.'s canonical microcircuit equations and training them as a neural ODE classifier on MNIST/CIFAR-10—and the paper is open about where the dynamics come from. That part is fine. But the headline empirical claim is not supported by the paper's own experimental section. The abstract says a single CMC node gets 97.8% on MNIST; Section 6.2 says 86.4%, Section 6.3 says 86.5%, and the figure shows accuracy plateauing around 93-94% across 1-5 nodes. A 5-region model is claimed to hit 99.8% in Section 5.2, but Figure 4 shows 94.3% as the best. These numbers cannot all describe the same setup. The reader's stress-test is right: the central result is unverifiable.\n\nThe paper does have virtues. The idea of treating a cortical microcircuit model as a trainable module rather than a simulation tool is sensible, and the explicit use of Bastos et al.'s equations rather than inventing a new dynamic makes it easy to see what is being trained. The phase-space analysis showing class-specific attractors is a nice observation. The code is on GitHub, though I didn't run it, and I'd want a versioned release. The ablation supplement is honest enough to report that each ablation outperformed the full model on a single V1 node, which is a useful piece of information even though it undercuts the narrative.\n\nWhere the paper is soft, in proportion: the missing matched control is a real problem. Every trained model includes a learned retinal front-end, a learned state-initialization projection, and a learned readout. Without replacing the CMC ODE by an ordinary feedforward map with the same front-end and readout, the efficiency claim can't be attributed to the microcircuit. The n=3 trials are mentioned but no intervals are shown. The baseline parameter comparison (1.3e3x fewer than ViT-Small) isn't tabulated. These are fixable, but they're load-bearing.\n\nWho should read this: researchers working on neural ODEs or neuromorphic computing who want to explore biologically structured dynamics as trainable modules. The paper doesn't deserve to be accepted as is, but the underlying idea is worth pursuing. A serious referee could help the authors resolve the inconsistencies and add the control. I'd send it to review rather than desk reject, because the architecture is novel and the code is available, and the empirical contradictions are exactly the kind of thing that referees are for.","headline":"A genuinely new CMC-as-nODE combination undermined by internally inconsistent accuracy numbers; the central claim can't be trusted as submitted.","tokens_in":14460,"tokens_out":4320,"would_cite":false,"duration_ms":38279,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proposes that canonical microcircuit motifs from the neocortex, implemented as learnable 8-dimensional neural ODEs, can classify images competitively and interpretably using far fewer parameters than conventional deep networks.","keywords":["canonical microcircuits","neural ODE","recurrent dynamics","neuromorphic computing","parameter efficiency","visual hierarchy","MNIST","phase space analysis"],"falsifier":"Train the same architecture with the canonical microcircuit ODE replaced by a plain feedforward or linear layer of comparable parameter count, keeping the retinal preprocessing, state-initialization projection, readout, and solver schedule fixed; if MNIST accuracy stays near 97.8 percent, the microcircuit dynamics are not the source of the claimed capability.","tokens_in":13292,"feed_emoji":"🧠","tokens_out":8323,"duration_ms":80461,"temperature":0.7,"pith_summary":"The paper tries to establish that the stereotyped four-population circuits of the neocortex, translated into learnable neural ordinary differential equations, are viable building blocks for computer vision. It reports that even a single canonical microcircuit node reaches 97.8 percent accuracy on MNIST, and that stacking one to five such nodes improves performance on more demanding image benchmarks while using only about 150,000 parameters, orders of magnitude fewer than conventional convolutional networks. If true, this would mean that biologically derived recurrent dynamics can replace much larger learned feature extractors and also supply interpretable, class-specific trajectories in state space. The practical payoff claimed is a path toward parameter-efficient, interpretable neuromorphic models grounded in real cortical circuit motifs.","feed_headline":"A cortically inspired ODE hits 97.8% on MNIST","feed_subtitle":"A recurrent circuit of four neural populations classifies digits with a tiny parameter budget.","key_machinery":"The load-bearing object is the canonical microcircuit (CMC) neural ODE: a four-population recurrent dynamical system in which spiny stellate cells, inhibitory interneurons, superficial pyramidal cells, and deep pyramidal cells each obey a second-order differential equation derived from a standard predictive-coding microcircuit model. Written as first-order equations, each node is 8-dimensional; its connectivity strengths ($\\gamma$) and time constants ($\\kappa$) are learnable, and the whole system is integrated with a fixed solver over a short time window as the hidden-state evolution of a neural ODE. This machinery converts a neuroscience circuit model into a trainable module, and repeated modules with inter-node connectivity form the hierarchy used for vision tasks.","core_discovery":"The central claim is that canonical microcircuits—stereotyped laminar patterns of spiny stellate, inhibitory, and pyramidal neurons—can be operationalized as an 8-dimensional neural ODE system and trained end-to-end on raw image classification. Each cortical area is modeled by this ODE, with learnable synaptic strengths and time constants, and multiple areas are connected by learnable forward and backward matrices that mirror the V1-to-V5 visual hierarchy. The paper reports competitive MNIST and CIFAR-10 accuracy with roughly 150,000 parameters for a five-region model, and shows that different digit classes converge to distinct phase-space attractors during inference. It concludes that biologically plausible circuit dynamics, not just brain-inspired layer wiring, can deliver both efficiency and interpretability.","pith_inferences":["The paper leaves open whether the microcircuit dynamics or the learnable preprocessing and readout surrounding them explain the gains; a matched control that replaces the CMC ODE with a plain feedforward layer would isolate the contribution.","If the recurrent ODE is doing the heavy lifting, any sufficiently expressive recurrent ODE with comparable weight sharing might reproduce the parameter efficiency, meaning the biological detail may matter less than recurrence itself.","The observed class-specific attractors suggest a testable extension the author does not pursue: using distance to each class attractor as a confidence score for out-of-distribution detection or few-shot classification.","Because the solver uses a fixed short time window, the model is effectively a depth-unrolled RNN; comparing against a discrete RNN with the same integration schedule would quantify what continuous-time dynamics add."],"forward_implications":["A single CMC node can serve as a parameter-efficient feature extractor for MNIST, with a five-node hierarchy reaching about 99 percent test accuracy at roughly 150,000 parameters.","Scaling from one to five nodes consistently improves accuracy on CIFAR-10, suggesting that deeper hierarchies of the same recurrent motif extract progressively more abstract features.","Removing inter-regional recurrence yields worse generalization and rougher loss surfaces than the recurrent version, so recurrent feedback across nodes is doing real work.","Class-specific attractors in phase space give a built-in interpretability channel: the dynamics themselves, not a post-hoc explanation, distinguish input classes."],"supporting_citations":[{"why":"Supplies the four-population canonical microcircuit equations and laminar connectivity that the model adopts as its ODE dynamics.","marker":"[6]"},{"why":"Provides the neural ODE framework and adjoint training method that make the CMC dynamics learnable end-to-end.","marker":"[22]"},{"why":"Grounds the excitatory-inhibitory microcircuit connectivity in established cortical circuit anatomy.","marker":"[17]"},{"why":"Shows how the connectivity parameters generalize to a spatial field formulation, which the paper uses to justify flexible inter-node coupling.","marker":"[7]"},{"why":"Provides the biological vision hierarchy comparison that motivates the V1-to-V5 model structure and the comparison to CNNs.","marker":"[33]"}],"fun_headline_variants":["8D microcircuit ODE hits 97.8% MNIST","Microcircuit ODEs: 97.8% MNIST, 150k params","Phase-space attractors in microcircuit ODEs identify digits","Hierarchical microcircuit ODEs: efficient, interpretable","Canonical microcircuit ODEs rival deep nets with fewer weights"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper attributes its accuracy and parameter efficiency to the microcircuit, but every model also contains a learnable retina-like preprocessing layer, a learned projection that sets the initial state, a learned readout, and hand-selected solver settings, and no control replaces the microcircuit ODE with an ordinary layer to show those parts are not responsible for the gains.","fun_headline_variants_meta":{"raw":{"variants":["8D microcircuit ODE hits 97.8% MNIST","Microcircuit ODEs: 97.8% MNIST, 150k params","Phase-space attractors in microcircuit ODEs identify digits","Hierarchical microcircuit ODEs: efficient, interpretable","Canonical microcircuit ODEs rival deep nets with fewer weights"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.002493,"raw_usage":{"total_tokens":9544,"prompt_tokens":902,"completion_tokens":8642,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":518,"completion_tokens_details":{"reasoning_tokens":8541}},"tokens_in":518,"tokens_out":8642,"duration_ms":54878,"temperature":1.0,"reasoning_tokens":8541,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:59:28.618239+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same architecture with the canonical microcircuit ODE replaced by a plain feedforward or linear layer of comparable parameter count, keeping the retinal preprocessing, state-initialization projection, readout, and solver schedule fixed; if MNIST accuracy stays near 97.8 percent, the microcircuit dynamics are not the source of the claimed capability.","supporting_citations":[{"cited_title":"Neural Ordinary Differential Equations,","cited_arxiv_id":null,"evidence_quote":"Provides the neural ODE framework and adjoint training method that make the CMC dynamics learnable end-to-end."},{"cited_title":"A functional microcircuit for cat visual cortex.,","cited_arxiv_id":null,"evidence_quote":"Grounds the excitatory-inhibitory microcircuit connectivity in established cortical circuit anatomy."}],"review_version":2}