{"id":"589484de-df92-40d1-9e48-39763cc8ede2","arxiv_id":"2412.01193","paper_version":3,"verdict":"REJECT","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"DEN is a shared-trunk multi-branch ensemble that reports roughly 4-6x faster inference than standard ensembles on simple benchmarks, with similar accuracy.","lead":"This paper introduces DEN, an ensemble network that shares one early layer and then splits into several independent branches. The authors report faster inference than standard ensembles on MNIST, NotMNIST, and a toy regression problem.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The efficiency gain is plausible, but the 'maintaining ensemble diversity' half of the central claim is unsupported: there is no matched-capacity comparison, no inter-branch diversity metric, and no ablation of the shared layer, so the uncertainty estimate may be a single model's confidence.","rationale":"The reader's weakest_assumption is the same one I would flag: the shared representation layer must be rich enough to keep branches diverse. I agree with the reader's REJECT-level skepticism. In good faith, the paper does provide code (Appendix A) and reproducible timing on a single core (Appendix B), which supports the efficiency claim as an engineering observation. However, the central scientific claim is not just speed; it is speed without sacrificing ensemble diversity. That second half is unsupported for the reasons above. The inference-time comparison is not controlled for total parameters, and no diversity metric is reported. The paper's own Section 5.1 concedes sensitivity to the shared representation choice, yet no ablation is run. The only test that would distinguish the bottleneck hypothesis from a benign architecture is a matched-capacity comparison with inter-branch disagreement metrics. Since the current evidence cannot rule out branch collapse, the paper should not be accepted as establishing the central claim. The verdict therefore remains REJECT, with no adjustment needed from the reader's assessment.","tokens_in":6261,"tokens_out":3879,"duration_ms":37155,"concrete_test":"Re-run the MNIST classification with three matched configurations: (i) DEN as published; (ii) a deep ensemble with per-member width adjusted so total parameter count equals DEN's total; (iii) a DEN variant with the shared layer removed so each branch sees the raw input. For each, report per-branch test accuracy, average pairwise prediction disagreement (or variance of branch probabilities) on the in-distribution test set, and inference time. If DEN's pairwise disagreement is substantially below the matched deep ensemble's, the uncertainty estimate is not maintained and the central claim fails; if the timing advantage disappears when total parameters are matched, the speedup is not from shared representation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is two-part: shared layers reduce redundancy, and independent branches preserve diversity. The first part is partly supported by the timing tables, but the second has no direct evidence. Table 2's only diversity-related statistics are average variance and entropy on NotMNIST, where every method is at chance (~10% accuracy); DEN's variance (0.0375) is actually lower than the deep ensemble's (0.0406), and no error bars are provided. Table 3 reports no uncertainty metric at all for regression. More importantly, the speedup comparisons are not matched in total parameter count or per-branch capacity. DEN's 0.056 s vs 0.263 s could simply reflect a smaller network, since no parameter counts or FLOPs are reported, so the 6x is not attributable to the shared representation. The paper's own Section 5.1 concedes that DEN's dependence on the choice of shared representations may require substantial hyperparameter tuning. If the single shared layer from Section 2.1 compresses information needed by the branches, branch outputs become correlated and the ensemble collapses to a single model's confidence. The paper never measures inter-branch agreement (e.g., pairwise KL divergence, prediction disagreement, or rank correlation of branch errors), so the load-bearing assumption of maintained diversity is untested.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Divergent Ensemble Networks (DEN), an architecture with a single shared input/representation layer followed by multiple independently trainable branches, intended to reduce parameter redundancy while preserving the diversity needed for ensemble uncertainty estimation. The authors report experiments on MNIST classification, NotMNIST out-of-distribution classification, and a toy sine regression, claiming roughly 4-6x faster inference than deep ensembles, MC Dropout, and bootstrap while maintaining comparable accuracy and uncertainty. The manuscript includes a code availability statement, a hardware description, and a limitations section acknowledging sensitivity to the shared representation choice.","tokens_in":6573,"tokens_out":3101,"duration_ms":28388,"significance":"If substantiated, the core idea is attractive: a shared-to-branching ensemble could make deep-ensemble-style uncertainty estimation practical in real-time settings. The paper's strengths are its simple and clear architectural proposal, the direct comparison against standard baselines, a reproducible toy regression setup with known noise, public code, and an explicit statement of the single-core hardware setup. These are useful starting points. However, the significance is currently limited by the evidential gap between the two-part central claim ('reduces parameter redundancy while maintaining ensemble diversity') and what the experiments actually measure. The efficiency part rests on uncontrolled timing comparisons, and the diversity part is not directly measured at all. The paper would be substantially strengthened by matched-capacity comparisons, inter-branch diversity metrics, error bars, and a proper regression uncertainty evaluation.","major_comments":[{"comment":"The central claim that DEN 'maintains ensemble diversity' is not supported by the evidence presented. No metric directly measures inter-branch diversity (e.g., pairwise KL divergence, prediction disagreement, or rank correlation of branch errors). In Table 2, all methods are at chance accuracy on NotMNIST (~10%), and DEN's reported average variance (0.0375) is actually lower than the deep ensemble's (0.0406). With no error bars, this does not even establish that DEN's variance is comparable, let alone that diversity is maintained. The paper needs a matched-capacity comparison and a direct diversity metric to substantiate the central claim.","section":"Section 3.3, Tables 1-2"},{"comment":"The efficiency comparison is not properly controlled. The paper does not report parameter counts, FLOPs, layer widths, or number of branches for DEN or any baseline; the timing measurements on a single CPU core therefore cannot isolate the effect of the shared representation from the effect of simply using a smaller network. The claimed 6x speedup in Table 1 and 5x speedup in Table 3 could be an artifact of model size differences. The authors should report model capacities and compare against ensembles with matched total parameter count or matched per-branch capacity.","section":"Section 3.3 and Appendix B"},{"comment":"The regression experiment does not evaluate uncertainty at all, despite Section 3.2 describing entropy-based uncertainty evaluation and out-of-domain testing. Table 3 reports only MSE, MAE, R², and inference time; there is no predictive variance, entropy, confidence interval, or any out-of-domain uncertainty metric. Consequently, the claim that DEN 'achieves similar or superior uncertainty estimation' for regression is unsupported.","section":"Section 3.2 and Table 3"},{"comment":"All empirical results appear to be single-run point estimates without error bars, standard deviations, or a description of hyperparameters (e.g., optimizer, learning rate, batch size, epochs, dropout rate, activation functions, number of branches, and branch widths). Given the small reported differences (e.g., DEN MNIST accuracy 98.78% vs. Ensemble 98.56%), the reader cannot assess whether any accuracy or uncertainty difference is statistically meaningful, and the experiments are difficult to reproduce from the text alone.","section":"Sections 3.3 and Appendix B"}],"minor_comments":[{"comment":"The text of Section 2.2 and Section 2.3 is essentially identical; the 'Shared Input and Common Representation Layer' and 'Divergent Branches for Independent Predictions' subsections appear to be duplicated and should be merged or rewritten.","section":"Sections 2.2 and 2.3"},{"comment":"There are several typos and formatting issues, including 'F unction' in the Section 2.4 heading, 'T est' in Section 3.2, 'V ariance' above the variance equation, and 'F uture W ork' in Section 5.2.","section":"Throughout"},{"comment":"The equations for entropy and variance are not numbered, and the entropy equation is introduced with a dangling phrase 'where ˆyi is the predicted probability or confidence score for each sample' that belongs to the variance formula; these need to be cleaned up.","section":"Section 3.2"},{"comment":"The sentence 'All of the methods show low entropy as expected' is confusing, because the stated goal is to see high entropy on unseen classes; the authors should clarify what is being compared and why low entropy is expected in that context.","section":"Section 3.3"},{"comment":"The statement 'We tested the model on 9000 examples from each class' is ambiguous: it could mean 9000 examples per class or 9000 total; please specify the exact evaluation protocol.","section":"Appendix B"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads as a preliminary workshop-style report rather than a complete journal article. The architectural idea is simple and potentially useful, but the experiments currently do not substantiate the two-part central claim. The main missing pieces—matched-capacity comparisons, direct diversity metrics, error bars, and a regression uncertainty evaluation—are all within the scope of a revision and could be addressed with additional experiments. I therefore recommend major revision rather than rejection, though the bar for the revised version will be high."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read on 2412.01193. The architecture is a shared feature extractor with independent branches, which is a well-known pattern in the ensemble literature; the paper doesn't cite any of it and doesn't add a new mechanism, loss, or theoretical angle. What it does do is report a concrete inference-time advantage on MNIST and a toy regression, roughly 4-6x over deep ensembles, and it ships the code. That efficiency result is plausible enough to be worth a footnote for someone working on cheap ensembling for small problems.\n\nThe soft spots are substantial. The central claim is two-part: shared layers reduce redundancy, and independent branches preserve diversity. The first part gets some support from the timing tables, but the second is simply not tested. There are no inter-branch diversity metrics, no pairwise agreement, no error bars, and no ablation of the shared layer. The NotMNIST results are at chance for every method, so the variance numbers there don't tell you anything. The regression table reports MSE, MAE, R² and time but no uncertainty metric at all. And the speedup comparison doesn't match total parameter count or per-branch capacity; without FLOPs or parameter counts, the 6x could just be a smaller network. The paper's own Section 5.1 concedes that the shared representation may require substantial hyperparameter tuning, which undercuts the efficiency claim further.\n\nThe duplicated text in Sections 2.2 and 2.3 and the vague experimental details (no learning rate, optimizer, or epochs) don't help. I agree with the stress-test note: the diversity half of the claim is load-bearing and unsupported.\n\nWho is this for? An engineer who wants a quick, code-available trick for cheaper ensembling on tiny datasets might find it useful. As a research paper, it doesn't clear the bar: the novelty is low, the evidence is thin, and the key claim about maintained diversity is untested. I would desk reject this version and suggest the authors resubmit with matched baselines, error bars, a real diversity analysis, and a literature review that situates the work against existing shared-trunk methods.","headline":"A familiar shared-trunk ensemble with a plausible speedup, but the diversity claim is untested and the evaluation is too thin to support the paper's novelty.","tokens_in":7038,"tokens_out":2287,"would_cite":false,"duration_ms":21554,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that sharing one representation layer across ensemble branches preserves uncertainty while cutting inference time by roughly sixfold.","keywords":["uncertainty estimation","deep ensembles","shared representations","divergent branches","out-of-distribution detection","inference efficiency","ensemble diversity"],"falsifier":"Take a trained DEN and shrink the shared representation layer from its original width down to about 16 units, then measure the variance of branch predictions on an out-of-distribution set such as NotMNIST. If branch variance collapses toward zero while accuracy changes little, the shared layer is an information bottleneck; if branch variance and OOD entropy stay high even at small widths, the independent branches carry the diversity themselves.","tokens_in":6068,"feed_emoji":"⚡","tokens_out":9573,"duration_ms":79313,"temperature":0.7,"pith_summary":"Deep ensembles get their uncertainty from many independently trained networks, but running many full networks costs time and memory. This paper proposes a Divergent Ensemble Network (DEN), in which all ensemble members share one input-processing layer and then split into independently trained branches. The claim is that common features can be computed once, so the shared-to-branching design removes redundant computation without removing the diversity that uncertainty estimation needs. In the reported comparisons on MNIST, NotMNIST, and a toy regression, DEN reaches roughly five- to six-fold faster inference than deep ensembles, bootstrap, and MC Dropout while keeping accuracy, error, and out-of-distribution uncertainty at similar levels. The intended payoff is an ensemble method that is practical for real-time or resource-limited systems yet still provides calibrated uncertainty.","feed_headline":"One shared layer makes deep ensembles 6x faster","feed_subtitle":"DEN splits after a shared representation, keeping ensemble accuracy while cutting inference time about 5-6x.","key_machinery":"The shared-to-branching architecture: one input layer and one shared representation layer compute features used by all branches, and multiple branches with separate weights, biases, and dropout layers then produce independent predictions. The shared layer is the speed mechanism, because the common feature computation runs once per input instead of once per ensemble member; the independently trained branches are the diversity mechanism, because they average over distinct hypotheses. Branch count and layer widths are hyperparameters, and each branch is trained with a task-specific loss, cross-entropy for classification and mean squared error for regression.","core_discovery":"The central discovery is architectural: a single shared representation layer followed by divergent branches produces the variance of a deep ensemble at a fraction of the inference cost. The paper demonstrates this by measuring accuracy, error metrics, and the entropy or variance of predictions on in-distribution and out-of-distribution data. On MNIST, DEN reaches 98.78 percent ensemble accuracy with an average inference time of 0.056 seconds against 0.263 seconds for the deep ensemble; on the toy regression, its R-squared is 0.9059 against 0.9088 and inference time drops from 0.299 to 0.066 seconds. On NotMNIST, DEN matches the out-of-distribution entropy and variance of bootstrap and ensemble methods rather than the overconfident pattern the paper attributes to MC Dropout. The conclusion is that the redundancy in fully independent ensemble training is avoidable without sacrificing predictive diversity.","pith_inferences":["Beyond the paper, the shared layer is the unexamined bottleneck: ablating its width or depth against inter-branch variance would show whether the observed diversity comes from the branching or simply from the shared representation being large enough.","Beyond the paper, comparisons with matched total parameter count would isolate the efficiency claim from any capacity advantage, since DEN and a conventional ensemble are not matched for total parameters in the reported tables.","Beyond the paper, the same shared-to-branching design should transfer to convolutional or transformer trunks, where the shared backbone dominates computation and the expected speedup is larger than in the small MLPs tested here."],"forward_implications":["If DEN holds up, ensemble uncertainty estimation can run about five to six times faster on a single CPU core, making it usable in real-time settings where running several full networks is too slow.","On out-of-distribution inputs, DEN behaves like a full ensemble rather than like MC Dropout, which the paper reports as overconfident on some unseen classes, so it is a safer drop-in where wrong confident predictions are costly.","Because the shared computation runs once, adding branches to increase diversity is cheaper than adding full ensemble members, letting practitioners scale the uncertainty estimate more affordably.","The speed gain does not require an accuracy trade-off on the tested benchmarks: accuracy, MSE, MAE, and R-squared stay comparable between DEN and the comparison methods."],"supporting_citations":[{"why":"Defines the deep-ensemble baseline that DEN must match on accuracy and uncertainty.","marker":"[8]"},{"why":"Supplies the MC Dropout baseline that the paper compares against and contrasts with its own OOD behaviour.","marker":"[3]"},{"why":"Provides bagging, one of the ensemble baselines whose inference cost DEN reduces.","marker":"[1]"},{"why":"Provides bootstrapping, a resampling-based baseline used for uncertainty comparison.","marker":"[7]"},{"why":"Supports the motivation that MC Dropout can compromise prediction independence.","marker":"[14]"},{"why":"Frames the redundancy-versus-diversity trade-off in ensemble deep learning that DEN addresses.","marker":"[4]"}],"fun_headline_variants":["Shared layer cuts deep ensemble inference 6x","One shared layer, divergent branches: cheaper deep ensembles","Ensemble variance at 1/6 the inference cost","Divergent branches off a shared layer match ensemble accuracy","Cut inference 6x without losing ensemble diversity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that one shared representation layer can supply every branch with all the features it needs; if that layer becomes an information bottleneck, the branches will make correlated predictions and the ensemble uncertainty will start to resemble a single model's confidence.","fun_headline_variants_meta":{"raw":{"variants":["Shared layer cuts deep ensemble inference 6x","One shared layer, divergent branches: cheaper deep ensembles","Ensemble variance at 1/6 the inference cost","Divergent branches off a shared layer match ensemble accuracy","Cut inference 6x without losing ensemble diversity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000247,"raw_usage":{"total_tokens":1473,"prompt_tokens":808,"completion_tokens":665,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":424,"completion_tokens_details":{"reasoning_tokens":590}},"tokens_in":424,"tokens_out":665,"duration_ms":6282,"temperature":1.0,"reasoning_tokens":590,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:36:01.942367+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a trained DEN and shrink the shared representation layer from its original width down to about 16 units, then measure the variance of branch predictions on an out-of-distribution set such as NotMNIST. If branch variance collapses toward zero while accuracy changes little, the shared layer is an information bottleneck; if branch variance and OOD entropy stay high even at small widths, the independent branches carry the diversity themselves.","supporting_citations":[{"cited_title":"Handling estimation uncertainty with bootstrapping: Empirical evaluation in the context of hybrid prediction methods","cited_arxiv_id":null,"evidence_quote":"Provides bootstrapping, a resampling-based baseline used for uncertainty comparison."},{"cited_title":"Ensemble deep learning: A review","cited_arxiv_id":null,"evidence_quote":"Frames the redundancy-versus-diversity trade-off in ensemble deep learning that DEN addresses."}],"review_version":1}