{"id":"04a1c0c8-f579-407c-a630-9fdcd88254e6","arxiv_id":"1908.09066","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Deep negative correlation learning trains a shared-feature ensemble of regressors with no extra weights, improving accuracy on crowd counting, personality analysis, age estimation, and super-resolution.","lead":"This paper trains a single deep network as a team of small predictors that are penalized for agreeing too much, then averages their outputs. The authors report accuracy gains on four image regression tasks at the same parameter cost as one network, plus a theoretical claim about reduced learning complexity.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 1 is salvageable with bounded weights; the load-bearing gap is empirical: no code, no error bars, and several super-resolution gains are within 0.1 dB.","rationale":"The reader identified Proposition 1's proof as the weakest assumption. My analysis shows the proof can be repaired: because the group-conv groups partition the channels, the sum of the K regressor outputs is representable as a full convolution with a block weight vector, so the class of sums equals the conventional class. With explicit bounded weights, the average shrinks the effective weight bound by 1/K, giving the claimed Rademacher reduction. Thus the theoretical objection is a rigor gap, not a fundamental flaw. The more load-bearing issue is the empirical claim, which depends on unverified numerical results: no code, no error bars, and several improvements are tiny. The paper's CONDITIONAL verdict is still appropriate, but for reproducibility rather than the Rademacher argument. The concrete test of rerunning with seeds and significance testing would settle whether the central empirical claim holds.","tokens_in":29179,"tokens_out":30482,"duration_ms":306375,"concrete_test":"Re-implement DNCL and the baselines (L2, SmoothL1, Tukey) on the four tasks with at least 5 random seeds, using the described backbones and hyperparameters, and report mean +/- std for each metric. Apply a paired significance test (e.g., Wilcoxon signed-rank or bootstrap) comparing DNCL against the best baseline per dataset. Additionally, train a true deep ensemble of K independently initialized networks with matched total parameter budget and compare against DNCL. If the super-resolution PSNR differences fall within one standard deviation, the universality claim fails; if the true ensemble matches or beats DNCL, the 'no extra parameters' advantage is not unique.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The reader's Proposition 1 concern is not fatal. In the proposed architecture the group-convolution groups partition the feature channels, so the sum of K group-conv outputs is exactly a full 1x1 convolution with a weight vector whose blocks are the group kernels. Thus the class of sums coincides with the conventional single-head class, and with bounded weights the averaging operator scales the function class by 1/K, yielding the claimed Rademacher factor. The proof in Eqn. (22) is terse and omits the boundedness assumptions, but the step is justifiable. The real load-bearing concern is empirical: the paper releases no code, reports no variance or error bars, and the super-resolution tables show DNCL ahead of the L2 baseline by at most 0.1 dB on several entries (e.g., Set5 x4, Set14 x2/x3). Without multiple seeds or significance tests, the headline claim of universal superiority over L2/SmoothL1/Tukey is not verifiable. Moreover, Table 8's 'Conventional Ensemble' is a shared-backbone multi-head network with lambda=0, not an ensemble of independently trained networks, so the claim of beating conventional deep ensembles is not directly tested.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Deep Negative Correlation Learning (DNCL), a deep regression ensemble in which K regressors share the lower convolutional feature extractor and are separated only by a group-convolution layer at the top. Each regressor is trained with the amended loss L_k = 1/2(G_k - Y)^2 - lambda (G_k - \\tilde G)^2 (Eq. 11), where \\tilde G is the ensemble average. The authors claim that, without extra network parameters, the method controls the bias-variance-covariance trade-off, that the averaged ensemble class has empirical Rademacher complexity 1/K of a conventional network (Proposition 1), and that DNCL outperforms L2, SmoothL1, and Tukey losses as well as several task-specific state-of-the-art methods on crowd counting, personality analysis, age estimation, and image super-resolution.","tokens_in":29377,"tokens_out":8746,"duration_ms":91057,"significance":"If verified, DNCL is an attractive and simple mechanism for obtaining ensemble-like gains without multiplying parameters: it is end-to-end trainable, complementary to backbone architecture choices, and tested on four rather diverse regression tasks. The paper is also honest in attributing the loss to the negative correlation learning literature and in disclosing the relation to the authors' CVPR 2018 paper. The theoretical claim, once repaired, would be a concise formal statement about the complexity of the averaged ensemble class. However, the empirical core is not yet fully established: the super-resolution gains are mostly within 0.1 dB, no error bars or multiple-seed statistics are reported, no code is released, and the 'conventional ensemble' baseline used in Table 8 is not an independently trained deep ensemble. The significance for TPAMI therefore depends on the authors addressing these verification gaps.","major_comments":[{"comment":"The proof of Proposition 1 is incomplete as written and should be rewritten. The equality inside the supremum where the sum of group-convolution outputs is replaced by a single conventional convolution must be justified explicitly: each W_k^Q is nonzero only on its channel group, so the sum over k is exactly a 1x1 convolution with a concatenated weight vector; then the factor 1/K follows by homogeneity of the empirical Rademacher complexity under scaling of the function class. The proof also uses Lemma 1, whose N=1 derivation drops the absolute value that appears in Definition 1, and Lemmas 2-3 omit a clear statement of the boundedness assumptions. In addition, the abstract claims that each sub-problem has reduced Rademacher complexity, whereas Proposition 1 as stated concerns the averaged ensemble class \\tilde G; the relation between the two claims should be reconciled.","section":"Sec. 3.2.3, Eq. (22)"},{"comment":"The baseline called 'Conventional Ensemble' in Table 8 is the same shared-backbone architecture with lambda = 0, not an ensemble of independently trained networks. This is clear from the text in Section 3.2.2, where the authors state that setting lambda = 0 achieves conventional ensemble learning. The comparison therefore isolates the effect of the correlation penalty but does not test the claim of outperforming conventional deep ensembles, which typically cost K times more compute. The authors should either include an independently trained K-network ensemble baseline, matching total parameter/compute budget where possible, or qualify the claim to say that DNCL outperforms a shared-backbone multi-head ensemble trained independently.","section":"Sec. 4.5, Table 8"},{"comment":"The central empirical claim of universal superiority over L2, SmoothL1, Tukey, and task-specific baselines is not supported by significance information. No standard deviations, multiple-seed runs, or paired tests are reported; the only repeated-evaluation case, MORPH age estimation, averages five random partitions but still reports no variance. In Table 7 the DNCL gains over DRRN are at most 0.1 dB in every super-resolution row, and several entries are equal (e.g., Set5 x4: 31.7 vs 31.7; Set14 x2: 33.2 vs 33.2). Differences of this magnitude are within typical run-to-run variability in PSNR. The authors should report error bars over at least 3-5 seeds for the main comparisons, or perform paired significance tests, and ideally release code or per-run results to make the single-shot numbers reproducible.","section":"Sec. 4, Tables 1-7"}],"minor_comments":[{"comment":"The word 'Radamacher' in Definition 1 is a typo and should read 'Rademacher'.","section":"Sec. 3.2.3, Definition 1"},{"comment":"In the first line of Eq. (11), the summation index j should run over j != k, not j != i, for consistency with the regressor index k.","section":"Sec. 3.2.1, Eq. (11)"},{"comment":"The text refers to the 'MPRPH' dataset; this should be MORPH, as in Table 6.","section":"Sec. 4.3"},{"comment":"The citation 'citewang2015deep' appears unresolved in the text and should be replaced with a proper reference.","section":"Sec. 2.2, Crowd Counting"},{"comment":"The authors state that lambda is set in [10^-3, 10^-2], but no ablation table or plot for lambda is reported; a sensitivity analysis for the central hyperparameter would strengthen the paper.","section":"Sec. 4.5, Effect of lambda and K"},{"comment":"The SSIM values in Table 7 appear to be scaled by 100, as noted in the text, but the table header simply says SSIM; the header should explicitly state the scaling factor.","section":"Table 7"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for TPAMI and the proposed method is simple and potentially useful. However, the empirical verification is the main weakness: the super-resolution margins are very small, no variance estimates are given, and the 'conventional ensemble' baseline is a shared-backbone variant rather than an independently trained ensemble. The Rademacher proof appears salvageable once the group-convolution structure is spelled out. I would ask the authors to provide per-run results and code, and to repair the theoretical proof."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper is a legitimate, reasonably well-scoped extension of the group's CVPR 2018 crowd-counting DNCL work to personality, age, and super-resolution, plus an attempted Rademacher complexity analysis. The loss itself is old (NCL from Liu/Yao), and the group-convolution shared-feature ensemble trick is not brand new, but the multi-task validation is useful and the results on crowd counting and age estimation are genuinely strong. The paper is also honest that it builds on their earlier paper; the self-citation is disclosed and appropriate.\n\nThe best parts: the method is simple and pluggable, the comparison to L2/SmoothL1/Tukey on the same backbone is the right experimental design, and the reported gains are substantial in places—UCF CC 50 MAE 288 vs 394 for L2, MORPH 2.85 vs 2.91 for DRF. The ablations on K and lambda, and the comparison against a shared-backbone lambda=0 ensemble, add real information.\n\nThe main soft spot is Proposition 1. The proof in Eqn. (22) replaces the supremum over the group-convolution ensemble with a supremum over the conventional class, asserting equality. That is not generally valid. Summing K group-convolutions on disjoint channel blocks gives a function class that is a subset of a single dense convolution class (block-diagonal weights), so the exact equality \\hat R_N(\\tilde G)=1/K \\hat R_N(G) is not established; at best the argument gives an inequality. This is a real gap in the paper's theoretical hook, not a typo.\n\nThe empirical side is thinner than the claims. No code, no error bars, no multiple seeds. On super-resolution several DNCL-vs-L2 differences are within 0.1 dB, which is below the usual run-to-run noise on those benchmarks; the 'clear margin' language oversells those entries. Table 8's 'Conventional Ensemble' is a shared-backbone lambda=0 multi-head network, not an ensemble of independently trained networks, so the claim about beating conventional deep ensembles goes beyond what is actually tested.\n\nFor a reader: this is a practical paper that belongs in the conversation on deep ensemble regression. It is not a breakthrough, but the empirical method is plausible and several of the gains look real. My recommendation for peer review: send it to referees, with the expectation of major revision—fix or re-frame the Rademacher claim, release code and variance numbers, and tone down the margin claims.","headline":"Useful incremental extension of DNCL with real empirical gains on crowd counting and age estimation; Proposition 1 has a genuine proof gap and the evaluation needs code and seeds before the 'clear margin' claims hold.","tokens_in":29947,"tokens_out":4893,"would_cite":true,"duration_ms":53058,"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 single convolutional network, trained with a negative-correlation penalty on K heads, can outperform both single-network and conventional-ensemble baselines on regression tasks while adding no parameters.","keywords":["deep regression","negative correlation learning","ensemble learning","deep ensemble","Rademacher complexity","crowd counting","age estimation","image super-resolution"],"falsifier":"Train the same backbone twice on one vision regression task, once with $\\lambda=0$ (conventional ensemble) and once with $\\lambda>0$, and also compute the empirical Rademacher complexity of the $K$-head group-convolution class on random labels; if the $\\lambda>0$ ensemble does not beat the $\\lambda=0$ ensemble on held-out error, or if the measured complexity ratio is not near $1/K$, the central claim fails.","tokens_in":28936,"feed_emoji":"📈","tokens_out":8211,"duration_ms":82779,"temperature":0.7,"pith_summary":"The paper claims that a deep regression network can be made to behave like an ensemble of $K$ regressors at the parameter cost of one network, by adding a negative-correlation penalty to the usual squared-error loss. The per-member loss is $L_k = \\frac12(G_k - Y)^2 - \\lambda(G_k - \\bar G)^2$, where $\\bar G$ is the average of all $K$ outputs; the second term actively decorrelates the members so their errors cancel on unseen data. Across crowd counting, apparent personality analysis, age estimation, and image super-resolution, the authors report that this one change beats a single network, several robust losses, and a conventional ensemble of the same architecture. If correct, the result gives practitioners a cheap \"ensemble for free\" plug-in that works with any convolutional backbone.","feed_headline":"Negative-correlation loss makes one CNN beat deep ensembles","feed_subtitle":"Four vision tasks—crowd counting, age, personality, super-resolution—show the gain over single and standard ensemble baselines.","key_machinery":"The load-bearing object is the amended loss $L_k = \\frac12(G_k-Y)^2 - \\lambda(G_k-\\bar G)^2$ together with the bias-variance-covariance decomposition of ensemble error, $\\mathbb{E}[(\\bar G-Y)^2] = \\frac{1}{K}\\sum_k \\mathbb{E}[(G_k-Y)^2] - \\frac{1}{K}\\sum_k \\mathbb{E}[(G_k-\\bar G)^2]$. The decomposition shows that penalizing each member's deviation from the ensemble average subtracts covariance from the total error, which is exactly the diversity mechanism. The implementation uses group convolution at the top layer to give each of the $K$ regressors a disjoint subset of the shared feature map, so the ensemble has no more parameters than a single network. The theoretical engine is Rademacher complexity: the paper proves Proposition 1, that the ensemble class has empirical Rademacher complexity $1/K$ times that of a conventional class, and uses it to argue that each subproblem fits random noise less easily.","core_discovery":"The central discovery is that negative correlation learning, previously used for shallow regression ensembles, transfers directly to deep convolutional regressors and yields a stronger ensemble without extra parameters. The authors define $K$ base regressors mounted on top of shared convolutional features, with the top feature map split into $K$ subsets by group convolution so each regressor sees a private slice. Each regressor minimizes its own squared error minus $\\lambda$ times its squared deviation from the ensemble average; minimizing the deviation term drives the members to be negatively correlated on the training data, which through the bias-variance-covariance decomposition reduces the ensemble's mean squared error. The paper further claims that the empirical Rademacher complexity of the resulting ensemble class is $1/K$ that of a standard network, so each subproblem is easier to optimize, and supports the claim with experiments on four vision regression tasks.","pith_inferences":["An extension the paper does not pursue: apply the same loss to other dense continuous regression tasks, such as monocular depth estimation or optical-flow prediction, where the top feature map naturally splits into $K$ groups.","A corollary of the paper's complexity claim is that the relative gain should grow as the training set shrinks, because Rademacher-complexity control matters most when data are scarce; a controlled study that varies training-set size could test this.","The paper does not measure the empirical Rademacher complexity of its actual trained models; computing it on random labels for the $K$-head group-convolution class would directly confirm or refute the claimed $1/K$ factor."],"forward_implications":["Swapping the loss in an existing regression CNN is a drop-in change: the method adds no parameters, so training and inference cost stay close to that of a single network.","The improvement should persist across regression tasks with continuous targets, since the loss is task-agnostic and the paper demonstrates it on density maps, trait scores, ages, and pixel-level residuals.","The optimal ensemble size $K$ is bounded by the feature dimension, because each member needs a private slice of the top feature map; the paper reports degradation at very large $K$ on crowd counting.","Combining NCL with other loss functions in the first term is possible, but the authors report it is usually weaker than using the squared-error first term, which guides how the loss should be deployed.","Because the method only changes the loss and the top layer, it is complementary to task-specific backbones and to other regularization strategies."],"supporting_citations":[{"why":"Defines negative correlation learning for ensembles of shallow networks; the amended cost function is the direct antecedent of the paper's loss.","marker":"[18]"},{"why":"Supplies the bias-variance-covariance decomposition of ensemble error used to motivate and derive the objective.","marker":"[19]"},{"why":"Provides the strength-correlation argument that accurate and decorrelated members improve ensemble generalization, the theoretical motivation for the diversity penalty.","marker":"[12]"},{"why":"Defines Rademacher complexity and the Lipschitz contraction lemma used in the paper's complexity bound.","marker":"[85]"},{"why":"Gives the generalization bound (Theorem 1) that connects empirical Rademacher complexity to expected error.","marker":"[86]"},{"why":"Introduces group convolutions, the mechanism that gives each of the $K$ regressors a private feature subset without extra parameters.","marker":"[29]"},{"why":"Provides the deep regression forest backbone and evaluation protocol for the age estimation experiments.","marker":"[27]"},{"why":"Reports the earlier crowd-counting application of the method that this paper extends to new tasks and additional theory.","marker":"[1]"}],"fun_headline_variants":["Negative correlation makes deep regression ensembles cheaper and better","Deep regression: negative correlation learning without extra parameters","One network, negative correlation, beats deep ensemble for regression","Negative correlation learning improves deep regression ensembles","Divide and conquer: negative correlation for deep regression"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The theoretical guarantee rests on identifying the averaged ensemble of regressors with the single-network class inside the Rademacher complexity calculation; if that identification is not valid, the claim that each subproblem is $1/K$ as complex is unsupported, even if the trained method still works empirically.","fun_headline_variants_meta":{"raw":{"variants":["Negative correlation makes deep regression ensembles cheaper and better","Deep regression: negative correlation learning without extra parameters","One network, negative correlation, beats deep ensemble for regression","Negative correlation learning improves deep regression ensembles","Divide and conquer: negative correlation for deep regression"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000186,"raw_usage":{"total_tokens":1324,"prompt_tokens":940,"completion_tokens":384,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":313}},"tokens_in":556,"tokens_out":384,"duration_ms":4578,"temperature":1.0,"reasoning_tokens":313,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:23:23.716734+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same backbone twice on one vision regression task, once with $\\lambda=0$ (conventional ensemble) and once with $\\lambda>0$, and also compute the empirical Rademacher complexity of the $K$-head group-convolution class on random labels; if the $\\lambda>0$ ensemble does not beat the $\\lambda=0$ ensemble on held-out error, or if the measured complexity ratio is not near $1/K$, the central claim fails.","supporting_citations":[{"cited_title":"Rademacher and gaussian complex- ities: Risk bounds and structural results,","cited_arxiv_id":null,"evidence_quote":"Defines Rademacher complexity and the Lipschitz contraction lemma used in the paper's complexity bound."},{"cited_title":"Empirical margin distributions and bounding the generalization error of combined classifiers,","cited_arxiv_id":null,"evidence_quote":"Gives the generalization bound (Theorem 1) that connects empirical Rademacher complexity to expected error."}],"review_version":1}