{"id":"8cc32e05-83f6-449f-b0ba-fc7d031d897a","arxiv_id":"2504.14665","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A hybrid local-global feedback predictive coding network with learned dynamic modulation and a composite predictive consistency loss reports modest accuracy gains over backpropagation and PCN baselines on four image benchmarks.","lead":"This paper proposes a predictive-coding image classifier that combines local and global feedback signals and gates them with a learned modulation factor, plus a four-part loss function. It reports small accuracy gains over a standard predictive-coding baseline and backprop networks on CIFAR-10/100 and MNIST/FashionMNIST, but provides no code and leaves key implementation details unspecified.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The superiority claim rests on an unspecified PCN baseline and test-set-tuned hyperparameters, so Table I cannot support it; a reproducible comparison is needed.","rationale":"The paper's central claim is explicitly comparative: DMPCN beats BP and PCN on four benchmarks. For that claim to hold, the PCN baseline must be a fair, sufficiently strong implementation, and the model-selection procedure must not use test labels. Both conditions fail in the text: PCN is not described, and §V-A says grid search and early stopping were used to ensure the highest test accuracy surpassing BP, with no validation split. This is the same load-bearing weakness the Reader identified, and I agree it is decisive. A secondary issue is the convergence metric: Fig. 2 uses Eq. 19, the proposed spatial loss, so 'faster convergence' is partly a statement that DMPCN's own loss decreases as DMPCN optimizes it. Numerical inconsistencies (0.28 vs. 0.33 in Table II versus §V-C; 98.78% vs. 99.72% MNIST accuracy between Table I and Tables II/III) further reduce confidence in the reported values. None of this proves the architecture is useless; it means the paper as written does not provide evidence adequate to accept the comparative claim. A re-run with specified baselines and validation-based selection would settle the question. Therefore the Reader's REJECT verdict stands, and no change is needed.","tokens_in":8916,"tokens_out":4698,"duration_ms":42368,"concrete_test":"Ask the authors to release code and reproduce Table I with a fixed, named baseline: the global PCN of Wen et al. (2018) and/or local PCN of Han et al. (2018), using the same AlexNet/VGG9 skeleton, same recurrent-step count, same optimizer schedule, and hyperparameters chosen on a held-out validation split (e.g., 10% of the training set) rather than on test accuracy. Report mean ± std over at least 5 seeds for each cell. If PCN's entries shift by more than its reported std, or if DMPCN's margin is no longer significant, the abstract's superiority claim should be withdrawn.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The main evidence for the central claim is Table I, but the comparison in §V-A cannot support it. The PCN baseline is never concretely specified: no architecture variant, recurrent-step count, update schedule, or hyperparameters are reported, so it is impossible to tell whether DMPCN is compared with a strong published PCN or a weak implementation. The selection protocol is also circular with respect to the claim: models are 'rigorously tuned ... through grid search to ensure the highest test accuracy surpassing BP,' and early stopping selects the best model, with no validation split described. That makes the DMPCN-over-BP advantage partly tautological and risks test-set overfitting in all columns. The faster-convergence evidence (Fig. 2) is similarly weak because it plots Eq. 19, which is DMPCN's own proposed spatial-loss term, so convergence of that curve is at least partly a tautology. Internal inconsistencies (Table II gives MNIST error 0.28 while §V-C says 0.33; Table I's LeNet/MNIST DMPCN 98.78% conflicts with the 99.72% implied by Tables II/III) further reduce confidence in the reported numbers. As written, the paper does not establish the headline comparative advantage.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DMPCN, a predictive coding network variant that combines local and global recurrent feedback paths with a sigmoid-based dynamic modulation of the feedback error, and introduces a four-term predictive consistency loss (hybrid cross-entropy with modulation, spatial consistency, spatial MSE, and reconstruction). The authors claim that DMPCN achieves higher predictive accuracy and faster convergence than both backpropagation-trained networks and a standard predictive coding network on MNIST, FashionMNIST, CIFAR-10, and CIFAR-100, and they report additional calibration and ablation experiments.","tokens_in":9263,"tokens_out":5044,"duration_ms":47010,"significance":"The conceptual direction of combining local and global recurrent updates with input-dependent modulation is reasonable, and a loss tailored to predictive coding dynamics is a plausible avenue for improving PCN training. The paper also includes ablations over loss components and calibration analyses, which go beyond a single accuracy table. However, the manuscript provides no code, no data, and no machine-checked derivations, and the empirical evidence is not reproducible from the text. The central comparative claim is therefore not established; the significance would be moderate if the experiments were rigorously controlled and the inconsistencies resolved.","major_comments":[{"comment":"The PCN baseline is never concretely specified. The text states that each model was 'rigorously tuned for hyperparameter optimization through grid search to ensure the highest test accuracy surpassing BP,' but it does not report the PCN architecture variant, the number of recurrent steps, the update schedule, the hyperparameter ranges, the number of seeds, or a validation split. Because the selection protocol is explicitly driven by test accuracy, the DMPCN-over-BP margin is partly tautological and the comparison may reflect a weak or under-tuned baseline. Without a reproducible baseline specification, Table I cannot support the headline superiority claim.","section":"§V-A, Table I"},{"comment":"The faster-convergence evidence plots a quantity defined in Eq. 19, L_SP, which is a component of DMPCN's own total loss in Eq. 21. Since DMPCN is explicitly trained to minimize this term while the PCN baseline is not, comparing the two models on this metric measures objective alignment rather than learning efficiency. The claimed faster convergence is therefore at least partly a consequence of the chosen metric, not independent evidence of superior optimization dynamics.","section":"Fig. 2, Eq. 19, Eq. 21"},{"comment":"The reported numbers are internally inconsistent. Table II lists the DMPCN+PCL MNIST error as 0.28, while §V-C text states 0.33 for the same configuration. Table I reports LeNet on MNIST DMPCN accuracy as 98.78%, implying a 1.22% error, whereas Tables II and III report accuracies of 99.72% for the full DMPCN on MNIST. These contradictions mean that at least some of the reported numeric results are incorrect, and they undermine confidence in every empirical comparison in the paper.","section":"Tables I, II, III and §V-C"},{"comment":"There is a mathematical inconsistency in the derivation of the representation update. With L_l defined in Eq. 3 as (1/(2σ^2))||e_l||^2, the gradient with respect to r_l is (1/σ^2)(r_l - p_l), not the factor-2 expression in Eq. 6. Equation 7 then compounds this factor in the update. Unless an alternative normalization is intended and stated, the recurrent dynamics of the proposed model are not correctly derived from the stated loss.","section":"Eq. 6 and Eq. 7"}],"minor_comments":[{"comment":"The datasets paragraph says the model is evaluated on three datasets (CIFAR-10, CIFAR-100, MNIST), but the abstract and Table I include FashionMNIST as a fourth dataset; the discrepancy should be corrected.","section":"§IV-E"},{"comment":"The paragraph introducing Eq. 19 calls it the 'spatial consistency loss,' which is the same name used for the spatial consistency term in Eq. 18; use distinct names to avoid confusion.","section":"§IV-D-3"},{"comment":"Equation 9 contains stray formatting ('[(]') and an undefined function φ; it is not used in the proposed method and should be removed or rewritten.","section":"§III-B, Eq. 9"},{"comment":"The y-axis is described only as 'prediction error' in the text, but the plotted quantity is L_SP from Eq. 19; the figure caption should state this explicitly so readers do not mistake it for a task-level error.","section":"Fig. 2"},{"comment":"These tables report point estimates without standard errors, number of seeds, or the underlying network architecture for each dataset (e.g., LeNet for MNIST, AlexNet or VGG9 for CIFAR); this information is necessary for reproducibility.","section":"Tables II and III"}],"recommendation":"reject","confidential_remarks":"The paper's central empirical claim rests on an underspecified and test-set-tuned comparison, and the reported numbers are internally inconsistent. The missing code and lack of baseline detail mean that even the strongest results cannot be independently verified. If the authors can provide a rigorous, validation-based comparison with a fully specified PCN baseline and correct the numerical inconsistencies, a resubmission could be considered, but the current manuscript does not meet the bar for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing you should know: this paper combines local recurrent processing from Han et al. and global recurrent processing from Wen et al. into one network, adds a learned sigmoid gate on the error feedback, and stacks a four-term loss. That exact combination is new, and the calibration analysis under corruption is a nice extra. If you work in predictive coding, this is a reasonable variant to know about.\n\nThe actual architecture seems coherent. Eqs. 10-16 define a straightforward hybrid: local conv-transpose error correction, global top-down feedback, and a modulation factor computed from the prediction error. The PCL loss is a bag of known terms—cross-entropy with a modulation-scaled weight, a quadrant-variance spatial consistency term, an MSE between representations, and a reconstruction loss. Nothing here is revolutionary, but it is a legitimate extension of the cited work.\n\nThe soft spot is the evaluation. I checked the numbers and the stress-test note is right. The PCN baseline is never fully specified: no architecture variant, no recurrent-step count, no update schedule. The tuning protocol says they grid-searched \"to ensure the highest test accuracy surpassing BP,\" which reads like test-set tuning, and there is no validation split described. That makes the DMPCN-over-BP advantage partly a selection artifact. Fig. 2 plots Eq. 19, which is the spatial loss DMPCN is explicitly trained to minimize, so the faster-convergence claim is at least partly tautological. There are also internal inconsistencies: Table II gives MNIST error 0.28 while Section V-C says 0.33; Table I gives LeNet/MNIST DMPCN accuracy 98.78% while Tables II/III imply 99.72%. No code or data are released. The architecture idea may still be sound, but the empirical evidence as written does not establish superiority over a strong baseline.\n\nFor whom is this paper? Read it if you track PCN variants and want to see one way to mix local and global feedback. It does not deserve acceptance without a much more careful, reproducible comparison—exact baseline, validation-based tuning, error bars over seeds, and a convergence metric that is not part of the proposed loss. On that basis, I would send it to a serious referee, because the combination is plausible and worth a rigorous test, but I would expect heavy revision.\n\nRecommendation: engage with the idea, but do not take the tables at face value yet.","headline":"The architecture is a plausible composition of known local and global PCN ideas, but the experimental comparison as reported cannot support the headline gains.","tokens_in":9737,"tokens_out":2166,"would_cite":false,"duration_ms":20908,"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":"Predictive coding with hybrid local–global feedback, dynamic input-dependent modulation, and a custom loss reports higher accuracy and faster convergence than backpropagation and plain predictive coding on four image benchmarks.","keywords":["predictive coding network","hybrid feedback mechanism","dynamic modulation","recurrent processing","object recognition","predictive consistency loss","calibration","image classification"],"falsifier":"Re-run the benchmark with a published, fixed PCN architecture, matching the number of recurrent updates, training epochs, and data preprocessing across BP, PCN, and DMPCN, and select the early-stopping epoch using only a validation split; if plain PCN then matches or exceeds DMPCN on VGG9/CIFAR-100 or AlexNet/CIFAR-10, the hybrid feedback and modulation are not the source of the reported advantage.","tokens_in":8738,"feed_emoji":"🧠","tokens_out":8034,"duration_ms":64402,"temperature":0.7,"pith_summary":"Predictive coding networks, modeled on the brain's error-prediction loop, usually use either local or global recurrent feedback, and their loss functions are generic. This paper argues that both choices are suboptimal and that a hybrid feedback path—combining fine-grained local reconstruction with global error propagation—guided by a per-layer modulation factor can fix them. The proposed DMPCN also uses a custom predictive consistency loss that ties spatial and reconstruction errors to the classification objective. The stake is that a biologically inspired learning rule would match or beat backpropagation-trained baselines on four standard image benchmarks while converging faster and using fewer layers.","feed_headline":"Predictive coding with hybrid feedback beats backprop on four benchmarks","feed_subtitle":"Dynamic input-dependent modulation and a tailored loss speed convergence and lift accuracy.","key_machinery":"The central object is the dynamic modulation factor $m_l = \\sigma(\\mathrm{Conv2D}(e_l; W^{(l)}_{\\mathrm{mod}}))$, computed from the prediction error $e_l$ at layer $l$ and applied as an element-wise gate to the global feedback signal before it updates the representation $x'_l = x_l + m_l \\odot e^{\\mathrm{feedback}}_l$. It carries the paper's argument because it lets the network scale feedback per spatial location and per input, while the hybrid feedback path itself—local transpose-convolution reconstruction plus global error propagation—supplies both fine detail and context. The predictive consistency loss $L_{\\mathrm{total}} = L_{\\mathrm{hybrid}} + \\mu L_{\\mathrm{SCT}} + L_{\\mathrm{SP}} + \\gamma L_{\\mathrm{recon}}$ is the second load-bearing component, tying the recognition objective to prediction-error minimization.","core_discovery":"On its own terms, the paper's central discovery is that a predictive coding network can combine two previously separate feedback strategies—local recurrent updates that reconstruct each layer's input with transpose convolutions, and global recurrent updates that propagate prediction errors down the hierarchy—and that a per-layer modulation factor learned from the error signal can govern how much of the global feedback reaches each representation. The same network also uses a four-part predictive consistency loss: a hybrid cross-entropy term scaled by modulation weights, a spatial consistency term over quadrant activations, a spatial mean-squared-error between actual and predicted layer representations, and a reconstruction loss. The paper reports that this combined architecture converges faster and reaches higher test accuracy than backpropagation and a plain PCN across MNIST, FashionMNIST, CIFAR-10, and CIFAR-100, with the largest reported gain on CIFAR-100 with VGG9 (74.84% versus 72.58% for PCN and 62.17% for BP).","pith_inferences":["A direct test of the loss's role would be to train a standard backprop CNN with the same predictive consistency loss; if the gain survives without hybrid feedback, the recurrence is not the main driver.","The modulation factor is a sigmoid gate over error maps, so it is close to learned attention on feedback; comparing against an attention-gated recurrent CNN would show whether the benefit is specific to predictive coding.","Because weights are still trained with SGD, the approach is only partially biologically plausible; a stricter test would constrain weight updates to local prediction-error signals and see whether the advantage over backprop persists.","The benchmarks are small-to-medium; extending to ImageNet-scale classification or dense prediction such as segmentation would test whether hybrid local-global recurrence matters where context and detail are both needed."],"forward_implications":["DMPCN should capture both fine-grained local details and broad contextual information in a single recurrent pass, because its local update reconstructs each layer's input while its global update propagates errors from higher layers.","The per-layer modulation factor is computed from the error signal, so the network can adapt feedback strength to each input; corrupted or variable inputs should be handled more gracefully than with a fixed feedback schedule.","The predictive consistency loss contributes measurable gains: on CIFAR-10 it reports a 5.67% error rate, versus 6.20% with KL divergence and 6.35% with cross-entropy, and similar gains on MNIST.","DMPCN reports lower calibration error under data corruption than BP and PCN, which matters for applications where uncertainty estimates must stay reliable under distribution shift.","Across the four benchmarks DMPCN reports the highest test accuracy except on LeNet/MNIST, where BP leads by 0.05 percentage points; on VGG9/CIFAR-100 it leads PCN by 2.26 points and BP by 12.67 points."],"supporting_citations":[{"why":"Defines predictive coding theory and the hierarchical error-prediction framework the paper builds on.","marker":"[16]"},{"why":"Supplies the local recurrent processing idea that DMPCN's local convolutional/transpose-convolutional update extends.","marker":"[6]"},{"why":"Introduces the global recurrent PCN that DMPCN combines with local feedback and uses as a comparison point.","marker":"[23]"},{"why":"Shows predictive coding can approximate backpropagation, motivating the BP-vs-PCN comparison.","marker":"[24]"},{"why":"Provides the CIFAR-10 and CIFAR-100 datasets used for the main classification comparisons.","marker":"[9]"},{"why":"Provides the MNIST dataset and the LeNet reference architecture.","marker":"[10]"},{"why":"Provides the FashionMNIST dataset used in the LeNet comparison.","marker":"[25]"},{"why":"Supplies the KL-divergence loss baseline that the predictive consistency loss is compared against.","marker":"[7]"},{"why":"Supplies the cross-entropy baseline for the loss-function comparison.","marker":"[11]"},{"why":"Supplies the calibration-error analysis approach used to evaluate robustness under corruption.","marker":"[4]"}],"fun_headline_variants":["Hybrid feedback predictive coding beats backprop on four datasets","Dynamic-modulated predictive coding outruns backprop on 4 benchmarks","Predictive coding with learned modulation tops backprop on four sets","DMPCN: hybrid feedback + dynamic modulation surpasses backprop on 4","Modulated predictive coding converges faster, hits 74.84% on CIFAR-100"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the plain PCN baseline was implemented and tuned as carefully as DMPCN—same backbone, same number of recurrent update iterations, and model selection on a validation split—so the reported accuracy and convergence gaps reflect the new mechanisms rather than an undertuned comparison.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid feedback predictive coding beats backprop on four datasets","Dynamic-modulated predictive coding outruns backprop on 4 benchmarks","Predictive coding with learned modulation tops backprop on four sets","DMPCN: hybrid feedback + dynamic modulation surpasses backprop on 4","Modulated predictive coding converges faster, hits 74.84% on CIFAR-100"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000227,"raw_usage":{"total_tokens":1469,"prompt_tokens":943,"completion_tokens":526,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":559,"completion_tokens_details":{"reasoning_tokens":428}},"tokens_in":559,"tokens_out":526,"duration_ms":4591,"temperature":1.0,"reasoning_tokens":428,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:43:52.038029+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the benchmark with a published, fixed PCN architecture, matching the number of recurrent updates, training epochs, and data preprocessing across BP, PCN, and DMPCN, and select the early-stopping epoch using only a validation split; if plain PCN then matches or exceeds DMPCN on VGG9/CIFAR-100 or AlexNet/CIFAR-10, the hybrid feedback and modulation are not the source of the reported advantage.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the local recurrent processing idea that DMPCN's local convolutional/transpose-convolutional update extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the global recurrent PCN that DMPCN combines with local feedback and uses as a comparison point."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows predictive coding can approximate backpropagation, motivating the BP-vs-PCN comparison."},{"cited_title":"LeCun, L","cited_arxiv_id":null,"evidence_quote":"Provides the MNIST dataset and the LeNet reference architecture."},{"cited_title":"LeCun, L","cited_arxiv_id":null,"evidence_quote":"Supplies the cross-entropy baseline for the loss-function comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the calibration-error analysis approach used to evaluate robustness under corruption."}],"review_version":1}