{"id":"5c3e9d4e-ad70-4c0d-a730-d8b6a5308b0b","arxiv_id":"2501.04700","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A biologically inspired 'planarian' architecture that swaps StemBlock weights between two parallel ResNets slightly reduces image classification error versus an ensemble, but only the CIFAR-10 gain reaches statistical significance.","lead":"A team at Harrisburg University trained pairs of ResNet-style networks that periodically swap early-layer weights, mimicking a planarian nervous system, and tested them on CIFAR-10 and CIFAR-100. The best swapped pair beat a plain ensemble by 0.17 percentage points on CIFAR-10, a difference that was statistically significant, and by 0.12 points on CIFAR-100, a difference that was not.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline claim rests on a single p=0.04 U-test computed after selecting the best patience value (PNN15) without multiplicity control or a demonstrated validation-only selection protocol; the CIFAR-100 result is not significant, so 'outperformed' is not established.","rationale":"The reader's weakest assumption points to the same vulnerable spot: the highlighted gains are selected after fine-tuning and are not a held-out evaluation of a preselected architecture. I agree with that concern and with the CONDITIONAL verdict. I would refine the framing: the patience gate is not a separate confound from cross-communication, because the gate only controls when swaps occur; a no-op patience gate would change nothing. The real load-bearing weakness is statistical selection. The paper reports exact error tables and uses nonparametric tests, which is a reasonable practice, and the raw means are consistent with a small benefit. However, the only statistically significant result in the paper is a best-of-three patience selection on CIFAR-10 with p=0.04 and five seeds, with no multiplicity correction and no explicit statement that the test set was untouched during patience selection. On CIFAR-100 the equivalent comparison is clearly not significant. This does not make the method impossible, but it means the central 'outperformed' claim is not supported by the reported statistics. The conditions the reader set (release code, hold out patience selection, add ablations) are appropriate; I would particularly emphasize the need for a corrected post-hoc test or a true held-out evaluation. Therefore the reader's CONDITIONAL verdict remains the right call.","tokens_in":16442,"tokens_out":11140,"duration_ms":114040,"concrete_test":"Re-run the CIFAR-10 dual-network experiment under a pre-registered protocol: split the official 10,000-image test set into a 2,000-image tuning split and an 8,000-image evaluation split. Tune the patience gate (5, 10, 15) only on the tuning split, then freeze PNN15 and evaluate once on the untouched evaluation split; compute the Mann-Whitney U statistic against the Ensemble baseline and apply a multiple-comparison correction across the three patience values. If the corrected p is not <0.05, the reported 'outperformed' claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that cross-communication alone improves accuracy. The decisive evidence is the CIFAR-10 dual-network comparison: after trying patience values 5, 10, and 15, the paper highlights PNN15 and reports a Mann-Whitney U test against the Ensemble baseline with p=0.04. This p-value is not a valid test of a pre-specified hypothesis. Section IV says the PNN models were 'fine-tuned' and then 'presented'; the paper never states that the patience value was chosen on validation before any test evaluation. If the three patience values were compared on the same 10,000-image test set used for the highlighted error rates, the p=0.04 is a post-selection artifact. If, more favorably, patience was chosen on the validation loader, the paper still reports no protocol or code showing this, and the best-of-three selection is not accounted for in the statistical test. On CIFAR-100 the same comparison is not significant (H test p=0.54; U test p=0.64 for PNN20 vs Ensemble), so the only quantitative support for 'outperformed' is one borderline, unadjusted p-value on five seeds. Without a held-out evaluation or a corrected post-hoc procedure, the strongest claim is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Planarian Neural Networks (PNNs), a dual-network architecture inspired by the planarian nervous system, in which two ResNet variants (a deep network and a wide network) are trained in parallel and periodically exchange StemBlock weights through an artificial brain, with the swap frequency controlled by a patience gate. The authors evaluate PNNs against single-network baselines and a soft-voting ensemble baseline on CIFAR-10 (five seeds) and CIFAR-100 (seven seeds), reporting mean error rates and nonparametric significance tests. They claim that cross-network communication improves test accuracy, with PNN15 on CIFAR-10 and PNN20 on CIFAR-100 outperforming all other models, and they attribute the improvement to the weight-exchange mechanism rather than to increased neuron count.","tokens_in":16720,"tokens_out":2072,"duration_ms":19855,"significance":"If the central claim were established, the paper would offer a modest but interesting biologically motivated training mechanism: periodically swapping early-layer weights between two networks during training, gated by validation-stagnation, could yield a small accuracy gain over a matched ensemble without increasing parameter count. The manuscript has genuine strengths: it provides per-seed result tables for all models, uses nonparametric tests appropriate for small samples, matches the parameter counts between ensemble and PNN, and reports both significant and non-significant results rather than only favorable ones. However, the headline claim is not currently supported because the main favorable result is a single borderline p-value obtained after selecting the best patience value on the test set, and the CIFAR-100 result is not statistically significant. The paper also lacks an ablation that isolates the weight-swap mechanism from the altered training schedule (subepochs and patience gating).","major_comments":[{"comment":"The U test comparing PNN15 with Ensemble (p = 0.04) is not a valid test of a pre-specified hypothesis. Section IV.A states that 'after the PNN models were fine-tuned, we presented the PNN5, PNN10, and PNN15 models,' and the same test set is then used both to select the best patience value (PNN15) and to compute the highlighted error rate. Because three patience values were compared on the same test data, the reported p-value is a post-selection artifact unless the authors demonstrate that the patience value was chosen on the validation split before any test evaluation; the manuscript provides no such protocol or code. The paper should report a held-out evaluation of a pre-selected architecture, or apply a multiplicity correction, or present the comparison of all three patience values as exploratory.","section":"IV.A and V.A (CIFAR-10 dual-network U test)"},{"comment":"The conclusion in Section VI that PNN20 'outperformed the other models' on CIFAR-100 is not supported by the statistical tests reported in Section V.B: the H test across the four dual networks gives p = 0.54 and the U test for PNN20 versus Ensemble gives p = 0.64, both far above alpha = 0.05. The authors themselves note that the improvements were not statistically significant, yet the Conclusions section still asserts outperformance. This contradiction between the stated results and the concluding claim must be resolved, for example by reframing the CIFAR-100 result as a non-significant trend or by providing additional evidence.","section":"V.B (CIFAR-100 dual-network tests)"},{"comment":"The paper's core attribution claim, stated in Section III, is that 'the improvements in test accuracy were induced by the inclusion of cross-network communication rather than an increase in the number of neurons.' The comparison between PNN and Ensemble holds neuron count constant, but the two procedures differ in multiple ways: the PNN trains each network in alternating subepochs, uses a patience gate that changes the effective learning-rate schedule, and swaps StemBlock weights. Without an ablation that trains two networks in the same subepoch schedule with the same patience gating but without the weight swap (or, conversely, with weight swap but no soft voting), it is impossible to attribute the observed difference specifically to cross-network communication. Such an ablation is load-bearing for the paper's central claim and should be added.","section":"III.A and IV (ablation isolating the swap mechanism)"},{"comment":"The manuscript does not specify a validation-based model-selection procedure. Section II.A mentions that 10% of the training images are used as validation, but Section IV.A's phrase 'after the PNN models were fine-tuned' suggests that the patience values (5, 10, 15 on CIFAR-10; 10, 15, 20 on CIFAR-100) were tuned using the test error. If the patience value was chosen on validation, the paper must state this explicitly, describe the selection rule, and account for the selection in the statistical analysis. If it was chosen on test error, then the reported significance levels overstate the evidence and a held-out evaluation is required.","section":"IV (fine-tuning and test-set reuse)"}],"minor_comments":[{"comment":"The index term 'plana1rian neural networks' contains a typo; it should read 'planarian neural networks.'","section":"Index Terms"},{"comment":"In Section IV.B, the text refers to 'PNN (ResNet164 + WideResNet108)' and later 'WideResNet108,' but the models described and listed in Tables V and VI are WideResNet110; the naming is inconsistent and should be corrected.","section":"IV.B (CIFAR-100 model name)"},{"comment":"Reference [19] is cited as 'University of Utah, 2023, Mann-Whitney U test' but the URL points to Wikipedia; the reference should be corrected to the actual source.","section":"References"},{"comment":"The phrase 'CIFAR0-10' in Section II.B appears to be a typo for 'CIFAR-10.'","section":"Section II.B"},{"comment":"The citation 'Eguyen et al.' in Section III appears to be a misspelling of 'Nguyen et al.' (reference [11]); please correct the in-text name.","section":"Section III"},{"comment":"The term 'patience gate' is sometimes written as 'patient gate' (e.g., Section IV.B: 'fine-tuned using patient gate values'); unify the terminology to 'patience gate' throughout.","section":"Section II.B and IV.A"},{"comment":"Algorithm 2 resets patience_level to 0 after a swap in the pseudocode, but the surrounding text states that 'each weight swap reset the patience level to 1.' This discrepancy between the pseudocode and the prose should be resolved.","section":"Algorithm 2 (Table II)"},{"comment":"The paper does not provide a statement on code availability or reproducibility details beyond the seed-selection description; sharing the training and evaluation code, including the exact weight-swap and validation logic, would strengthen the reproducibility of the results.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper's empirical contribution is close to publishable in a workshop or a venue with modest standards for significance, but the central claim is currently overclaimed relative to the statistical evidence. The authors should be asked to either (a) provide a validation-based, pre-registered selection protocol with appropriate multiplicity control, or (b) reframe the results as an exploratory finding and add the missing ablation. I would not recommend rejection outright because the per-seed tables are transparent and the idea is testable, but the manuscript in its current form does not establish the stated conclusion."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper is a legitimate small empirical study of a concrete architecture—two parallel ResNets that exchange StemBlock weights on a patience-gated schedule and then soft-vote. The gains over a matched ensemble are small (0.17 pp on CIFAR-10, 0.12 pp on CIFAR-100) and only the CIFAR-10 difference reaches p<0.05. That said, the authors do several things right: they give seed-level tables, use nonparametric tests, match total neuron counts between ensemble and PNN, and they honestly report the CIFAR-100 comparison as not significant in the discussion, even if the conclusions overstate it.\n\nThe main problem is the statistics behind the headline. The paper tunes the patience gate over values 5, 10, 15 (and 20 on CIFAR-100), then reports the best one (PNN15) against the ensemble. The p=0.04 U-test is computed on that chosen configuration, with no correction for the selection and no statement that patience was fixed on validation before touching the test set. So the 0.04 is not a valid test of a pre-specified hypothesis. On CIFAR-100 the equivalent comparison is not significant (p=0.64), so 'outperformed the other models' is only supported by one borderline, unadjusted p-value on five seeds. The stress-test note's reading is fair: this is a post-selection artifact unless the authors can show a validation-only selection protocol.\n\nThere is also no ablation that separates the weight-swapping mechanism from the soft voting. Since the PNN is an ensemble plus swapping, the natural control is an ensemble with the same training schedule but no swaps. Without that, the 'cross-communication rather than an increase in the number of neurons' claim is not actually demonstrated. And the paper does not discuss the obvious related literature—ensemble distillation, snapshot ensembles, weight averaging—so the novelty claim is more modest than the framing suggests. No code is released, which makes the protocol hard to check.\n\nNone of this is fatal to the underlying idea. The mechanism is concrete, the write-up is clear, and the effect, if real, is the kind of small accuracy improvement that could matter in some settings. But as it stands, the central claim is not established. A referee should ask for code, a validation-only patience selection protocol, and an ablation that isolates the swap. With those, the paper could be a solid workshop-level contribution. For now, I would not cite it, but I would send it out rather than desk reject—a serious referee could get it into shape.","headline":"A clearly described ensemble variant with a small, possibly real gain, but the headline claim leans on a post-selection p-value and no code.","tokens_in":17255,"tokens_out":2228,"would_cite":false,"duration_ms":19912,"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":"Periodically swapping early-layer weights between two parallel networks lowers error below an ordinary ensemble on CIFAR-10 and CIFAR-100.","keywords":["planarian neural networks","cross-network communication","StemBlock weight exchange","patience gate","ensemble learning","wide residual networks","CIFAR image classification","biologically inspired computing"],"falsifier":"Retrain the paired networks with the patience gate fixed before any test-set evaluation, or chosen on a separate validation split; if PNN15 or PNN20 no longer beat their ensembles, the reported advantage was an artifact of test-set model selection. Alternatively, train the same two networks with identical schedules but no weight swapping and soft-vote them; if the no-swap control matches the PNN's test error, the exchange contributes nothing.","tokens_in":16239,"feed_emoji":"🧠","tokens_out":7934,"duration_ms":71657,"temperature":0.7,"pith_summary":"This paper claims that coupling two neural networks the way a planarian couples its brain with its two nerve cords improves image classification accuracy. The two networks train in parallel, and every few epochs the weights of their early StemBlocks are exchanged through an artificial brain, with the exchange frequency set by a patience gate. The coupled pair, called a PNN, is compared with an ordinary soft-voting ensemble of the same two networks, and the paper reports lower average test error on both datasets: 5.81 versus 5.98 percent on CIFAR-10, and 20.67 versus 20.79 percent on CIFAR-100. The CIFAR-10 difference is reported as statistically significant; the CIFAR-100 difference is not. The aim is to show that cross-network communication during training, not extra neurons or a stronger final combination, is what drives the gain.","feed_headline":"Twin networks that swap early-layer weights beat a static ensemble","feed_subtitle":"CIFAR-10 error drops from 5.98% to 5.81%; CIFAR-100 shows the same trend with a smaller gap.","key_machinery":"The mechanism is a dual-network training loop with a StemBlock portal and a patience gate. A StemBlock is the shallow initial convolutional block that the PNN uses as the weight-exchange site; each global epoch, the two nerve-cord networks train for their own subepochs, and when the patience gate fires, the StemBlock weights of both networks are copied into an artificial brain and interchanged. The patience gate fires only after validation accuracy fails to improve for a set number of consecutive epochs, so swapping is adaptive rather than fixed-frequency. At test time the two networks' prediction probabilities are merged by soft voting. The comparison to a static ensemble isolates the swap: identical network members, identical total neuron count, with only the gated StemBlock exchange differing.","core_discovery":"On the paper's own terms, the central discovery is that periodic StemBlock weight exchange between two parallel networks improves test accuracy over both the individual networks and a conventional ensemble. PNN15, the best PNN on CIFAR-10, reaches 5.81 percent mean test error against 5.98 percent for the soft-voting ensemble, and the difference has a Mann-Whitney U p-value of 0.04. On CIFAR-100, PNN20 reaches 20.67 percent against 20.79 percent for the ensemble, a smaller difference with p = 0.64. Because the PNN and the ensemble contain the same total number of neurons, the paper attributes the improvement to the cross-communication mechanism rather than to capacity. The paper also reports that individual networks embedded in a PNN match or slightly beat their baselines, with the best single member improving from 7.44 to 7.27 percent mean error on CIFAR-10.","pith_inferences":["A natural testable extension is to swap weights at different depths or to replace discrete swaps by an exponential moving average; if the gain depends specifically on early-layer synchronization, depth and averaging schedule should matter.","The monotone trend with patience values suggests mapping the full patience curve beyond 15 on CIFAR-10 and 20 on CIFAR-100; an optimum would show where the reported trend turns around.","If the benefit is real, a no-swap control with identical training schedules but ordinary soft voting should reproduce only the ensemble baseline; if that control equals the PNN's accuracy, the exchange itself is not the active ingredient."],"forward_implications":["A paired ResNet and WideResNet trained with gated StemBlock exchange can beat a soft-voting ensemble of the same pair without adding parameters.","Increasing the patience gate value from 5 to 15 on CIFAR-10 and from 10 to 20 on CIFAR-100 lowered the average test error in the tested range, pointing to less frequent exchange as beneficial.","The best coupled PNN member on CIFAR-10, PNN15's ResNet20, improves mean test error from 7.44 to 7.27 percent over its baseline, so the mechanism helps not just the combined system but the individual learners.","On CIFAR-100 the dual-network gain is directionally consistent with CIFAR-10 but not statistically significant with seven runs, so a larger seed count is the natural next check."],"supporting_citations":[{"why":"Supplies the ResNet20 architecture and the five-run evaluation convention used for CIFAR-10 baselines.","marker":"[4]"},{"why":"Supplies ResNet164 and the mean/median evaluation conventions used for CIFAR-100 baselines.","marker":"[5]"},{"why":"Provides the CIFAR-10 and CIFAR-100 datasets on which all models are trained and tested.","marker":"[8]"},{"why":"Provides the biological principle that fusion of nerve centers drives nervous-system evolution, which motivates the brain-nerve-cord design.","marker":"[9]"},{"why":"Defines the StemBlock, the early convolutional block used as the weight-exchange portal between nerve cords.","marker":"[16]"},{"why":"Supplies the wide residual network design used for WideResNet14 and WideResNet110 baselines.","marker":"[17]"}],"fun_headline_variants":["Planarian-style twin nets beat static ensembles via weight exchange","Weight-swapping parallel networks outperform traditional ensembles","CIFAR-10 error drops to 5.81% with planarian weight swap","Twin nets exchange early weights, beat static ensembling","Planarian-inspired parallel nets swap weights, top CIFAR accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim stands on the premise that the small error advantage of PNN15 on CIFAR-10 and PNN20 on CIFAR-100 comes from the periodic StemBlock weight exchange itself, and not from the fact that the patience gate values were selected using the same test data that produced the highlighted error rates.","fun_headline_variants_meta":{"raw":{"variants":["Planarian-style twin nets beat static ensembles via weight exchange","Weight-swapping parallel networks outperform traditional ensembles","CIFAR-10 error drops to 5.81% with planarian weight swap","Twin nets exchange early weights, beat static ensembling","Planarian-inspired parallel nets swap weights, top CIFAR accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000678,"raw_usage":{"total_tokens":3075,"prompt_tokens":933,"completion_tokens":2142,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":549,"completion_tokens_details":{"reasoning_tokens":2068}},"tokens_in":549,"tokens_out":2142,"duration_ms":15704,"temperature":1.0,"reasoning_tokens":2068,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:25:53.293984+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the paired networks with the patience gate fixed before any test-set evaluation, or chosen on a separate validation split; if PNN15 or PNN20 no longer beat their ensembles, the reported advantage was an artifact of test-set model selection. Alternatively, train the same two networks with identical schedules but no weight swapping and soft-vote them; if the no-swap control matches the PNN's test error, the exchange contributes nothing.","supporting_citations":[{"cited_title":"‘Identity mappings in deep residual networks,’ Vis. ECCV,","cited_arxiv_id":null,"evidence_quote":"Supplies ResNet164 and the mean/median evaluation conventions used for CIFAR-100 baselines."},{"cited_title":"Krizhevsky, Learning Multiple Layers of Features from Tiny Images , 2009, pp","cited_arxiv_id":null,"evidence_quote":"Provides the CIFAR-10 and CIFAR-100 datasets on which all models are trained and tested."},{"cited_title":"From nerve net to nerve ring, nerve cord and brain — Evolution of the nervous system,","cited_arxiv_id":null,"evidence_quote":"Provides the biological principle that fusion of nerve centers drives nervous-system evolution, which motivates the brain-nerve-cord design."},{"cited_title":"Pelee: A real -time object detection system on mobile devices,","cited_arxiv_id":null,"evidence_quote":"Defines the StemBlock, the early convolutional block used as the weight-exchange portal between nerve cords."}],"review_version":1}