{"id":"beb11bad-c4bd-4636-a0a1-5758e12ad8c5","arxiv_id":"2502.08474","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"LBYL restores pruned CNNs by spreading each pruned filter's contribution across multiple preserved filters via a closed-form delivery matrix, outperforming one-to-one neuron merging without data or fine-tuning.","lead":"This paper introduces LBYL, a method to restore pruned neural networks without retraining or access to the original data. It distributes each removed filter's information across many remaining filters using a closed-form solution, and reports accuracy gains over the previous one-to-one compensation approach on several benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The step from Eq. (7) to Eq. (8) treats convolution with a fixed filter as scalar proportionality, so Theorem 2's closed-form solution is only proven to minimize the surrogate Lre, not the true next-layer reconstruction error.","rationale":"The reader's weakest_assumption correctly identifies the gap between the pre-convolution residual and the actual next-layer reconstruction error, located at the transition from Eq. (7) to Eq. (8) in Section IV-A. My analysis confirms this is the most load-bearing concern: Theorem 2's closed-form solution minimizes Lre, which is derived from Eq. (8) after dropping the data-dependent activation A^{(l-1)} and switching L1 norms to L2. The paper presents no argument that the minimizer of Lre approximates the minimizer of Eq. (4); treating the convolution with fixed filters as scalar proportionality is mathematically unjustified. This does not necessarily invalidate the empirical method—the reported gains over Neuron Merging are consistent across CIFAR, ImageNet, COCO, and FashionMNIST, and the method's practical utility is plausible. However, the central theoretical claim, namely that the closed-form solution is derived from a principled reconstruction-error formulation, is not established. The reader's CONDITIONAL verdict is appropriate: the concern is addressable by either strengthening the theory (e.g., bounding the surrogate gap via operator norms) or reframing the contribution as a heuristic with empirical support. I therefore recommend UNCHANGED rather than rejecting, because the empirical evidence is substantial and the theoretical flaw is not a demonstrated failure of the method itself.","tokens_in":26738,"tokens_out":5476,"duration_ms":49676,"concrete_test":"Using a pretrained ResNet-50 on CIFAR-100, prune one convolutional layer at 30% and compute, over the validation set: (i) the value of Lre in Eq. (11) at the s given by Theorem 2 with the paper's reported λ1, λ2; (ii) the true next-layer error of Eq. (4) at that s; (iii) the minimum achievable value of Eq. (4) obtained by solving a least-squares problem for s directly on the validation activations A^{(l)}. If the s from Theorem 2 yields a true error more than 10% above the empirical minimizer of Eq. (4), the surrogate objective is misaligned with the claimed reconstruction error. Repeating on 2–3 layers would directly settle whether the proportionality step between Eq. (7) and Eq. (8) holds.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section IV-A, the paper derives the exact next-layer reconstruction error in Eq. (7) as sum_i ||(A_j - sum_k s_k A_k) ⊛ W_{i,j}||_1. It then asserts that, since W_{i,j} is constant, minimizing Eq. (8), ||A_j - sum_k s_k A_k||_1, has the same effect. This is not valid: convolution with a fixed filter is a linear operator with a nontrivial kernel and non-uniform frequency response. Two residuals with identical L1 norm can produce outputs whose L1 norms differ by an arbitrarily large factor, and the argmin of ||A ⊛ W||_1 over s need not coincide with the argmin of ||A||_1. The paper then replaces the L1 objective with the L2 loss Lre = ||E||_2^2 + λ1||B||_2^2 + λ2||s||_2^2 in Eq. (11), dropping the data-dependent activation A^{(l-1)} entirely. Even if the algebra in Theorem 2 is correct, it establishes only that s minimizes Lre, a double surrogate for Eq. (4). The abstract's claim that the method is 'based on a theoretical analysis on how to formulate the reconstruction error' is therefore not supported by the derivation: the link between Lre and the true loss is an unproven proportionality assumption. The empirical gains are credible and consistent, but they do not validate the theoretical derivation, and the paper's own Eq. (10) shows the exact auxiliary loss contains A^{(l-1)} ⊛ E, which cannot be minimized data-free without additional assumptions. This gap is the most load-bearing weakness because the closed-form solution is the paper's central theoretical contribution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LBYL, a training-free and data-free method to restore pruned convolutional neural networks by compensating each pruned filter with a linear combination of preserved filters, defined via a delivery matrix. The authors derive a data-free surrogate loss Lre that combines a residual error term, a batch-normalization error term, and an L2 regularizer, and provide a closed-form solution for its minimizer in Theorem 2. They evaluate LBYL against Neuron Merging and a no-recovery baseline across VGG-16, ResNet-34/50/101, MobileNet-V2, SSD on COCO, and LeNet-300-100, reporting consistent accuracy gains, especially at higher pruning ratios.","tokens_in":27149,"tokens_out":3111,"duration_ms":31798,"significance":"If the empirical results are taken at face value, this is a practically useful contribution to the small but growing area of data-free, fine-tuning-free pruning recovery. The paper reports consistent improvements over the one-to-one baseline across many architectures, datasets, pruning criteria, and downstream transfer settings, and it releases code for reproducibility. However, the paper's advertised theoretical contribution is not substantiated: the derivation connecting the true reconstruction error to the closed-form surrogate loss contains an unjustified proportionality step and a broken inequality in Lemma 2. The method may well be a strong heuristic; the manuscript would need either a rigorous link between Lre and the original objective or a substantial reframing of its claims to make the theory load-bearing.","major_comments":[{"comment":"The reduction from Eq. (7) to Eq. (8) is not valid as stated. The paper claims that because W_{i,j}^{(l+1)} is constant, minimizing ||A_j - sum_k s_k A_k||_1 has the same effect as minimizing sum_i ||(A_j - sum_k s_k A_k) \\circledast W_{i,j}^{(l+1)}||_1. Convolution with a fixed filter is a linear operator with a nontrivial kernel and non-uniform frequency response; the argmin of ||A \\circledast W||_1 need not coincide with the argmin of ||A||_1, and the L1 norms of two residuals can be ordered differently before and after convolution. Because this step is the only link between the surrogate Lre and the original objective in Eq. (4), Theorem 2 is not proven to minimize the true reconstruction error. The authors should either provide a rigorous bound (e.g., using the operator norm of convolution) that connects the surrogate to Eq. (4), or revise the theoretical claims accordingly.","section":"Section IV-A, Eq. (11) and Lemma 2"},{"comment":"The transition from Eq. (10) to the data-free loss Lre in Eq. (11) is not justified by an inequality: replacing the L1 norms with squared L2 norms, dropping the data-dependent term A^{(l-1)} \\circledast E, and absorbing the activation error R into a regularization term changes the objective in a way that is not shown to be bounded by or equivalent to the original error. Furthermore, the proof of Lemma 2 contains a mathematically invalid equality: ||sum_k s_k min(0,N(Z_k)) - min(0,N(Z_j))||_1 is not equal to ||sum_k s_k min(0,N(Z_k))||_1 - min(0,N(Z_j)) for arrays whose entries have mixed signs. Consequently, the asserted upper bound in Eq. (12) is unproven, which removes the stated justification for the AE-regularization role of lambda2. This is a load-bearing gap because Theorem 2 solves Lre, not Eq. (4), and the paper's abstract claims the method 'is based on a theoretical analysis on how to formulate the reconstruction error.'","section":"Theorem 2"},{"comment":"Even assuming the algebra of Theorem 2 is correct, it establishes only that the closed form minimizes the specifically constructed surrogate Lre. Since the preceding derivation does not connect Lre to the next-layer reconstruction error of Eq. (4), the central theoretical contribution is not yet demonstrated. The empirical sections are credible and consistent, and they may be enough to support a heuristic method, but the manuscript should either add a formal approximation guarantee or clearly state that the closed form is a heuristic design choice validated empirically rather than an exact solution to the reconstruction problem.","section":"Algorithm 1"}],"minor_comments":[{"comment":"The paragraph about overlapping work and a pending plagiarism dispute does not belong in a scientific paper's main text; such matters are normally handled through bibliographic due diligence and editorial communication. As written, it leaves the paper without two references that appear directly relevant to the derivation, and it is not appropriate to ask readers to adjudicate an unresolved dispute.","section":"Section II, 'Clarification regarding overlapping work'"},{"comment":"The final sentence of the abstract, 'The very first version of this work... was submitted to NeurIPS 2021 and ICML 2022,' is unusual for an abstract and should be removed or moved to a footnote.","section":"Abstract"},{"comment":"The hyperparameters lambda1 and lambda2 are tuned per dataset, pruning criterion, and pruning ratio via grid search. The manuscript does not report sensitivity to these hyperparameters or to the number of grid points, leaving open the question of how much of the reported gain relies on this per-setting selection.","section":"Appendix C, Tables A3-A6"},{"comment":"The column header 'Parmas' should be 'Params'.","section":"Table X"},{"comment":"The caption of Figure 4 appears to have duplicated and misaligned subfigure labels: 'Averge Norm of Scales' is repeated, and the labels do not match the panels described in the text.","section":"Figure 4"}],"recommendation":"major_revision","confidential_remarks":"The manuscript contains a passage asserting that two published papers share core derivations with an earlier non-public version of this work and stating the authors have formally reported concerns. This is a serious integrity-related claim that the editor will need to handle separately from the technical review. On the technical side, the empirical study is extensive and the improvements over Neuron Merging are consistent; the main obstacle is the unproven theoretical bridge from the surrogate loss to the true reconstruction error. I would encourage the authors to either supply a rigorous mathematical link (even with multiplicative constants) or rewrite the contributions so that the closed-form solution is presented as a principled heuristic, which the data supports. The invalid equality in the proof of Lemma 2 must be corrected or the lemma restated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is worth reading for its empirical claims: LBYL consistently beats Neuron Merging across CIFAR, ImageNet, COCO, and FashionMNIST, often by large margins at high pruning ratios. The core idea—represent each pruned filter as a linear combination of all remaining filters, with a closed-form solution that accounts for batch-norm affine parameters—is a natural extension of one-to-one compensation, and the experiments are extensive enough to make the gains credible. The method itself is simple and reproducible, and the code appears to be available.\n\nThe soft spot is the theory. The derivation from Eq. (7) to Eq. (8) treats convolution with a fixed filter as scalar proportionality, which is not generally valid. Convolution is linear but has a nontrivial kernel and non-uniform frequency response; two residuals with the same L1 norm can lead to output errors differing by an arbitrary factor. So minimizing ||A_j - sum s_k A_k||_1 is not equivalent to minimizing the actual next-layer reconstruction error. The subsequent loss Lre in Eq. (11) drops the data-dependent activation entirely, making it a double surrogate. The closed form in Theorem 2 exactly minimizes Lre, but Lre is not the true error. The proof of Lemma 2 also contains an invalid L1 equality, even though the bound itself might be recoverable. These are not minor typos; they are the difference between claiming 'theoretically rigorous' and offering a well-motivated heuristic.\n\nThere is also a separate issue: the inserted 'Clarification regarding overlapping work' says two published papers share core derivations with an earlier version of this submission, and that the authors are deliberately not citing them while a dispute is unresolved. That is an explicit gap in the citation pattern. It does not invalidate the experiments, but it means external novelty cannot be assessed from the paper alone. An editor should be aware.\n\nThe hyperparameters lambda1 and lambda2 are tuned per dataset and per pruning ratio by grid search. That is not a fatal flaw because the baseline NM is also tuned, but it weakens the 'data-free' claim—the method needs a labeled validation set to pick hyperparameters.\n\nOverall: the empirical contribution is real, and the method deserves a serious referee. The theory needs to be reframed as a heuristic with analysis of when the surrogate correlates with the true error, and the Lemma 2 proof fixed. I would not cite it in my own work until the overlap dispute is resolved, but I would send it to review. Bring it to reading group if you want a concrete case of a surrogate loss running ahead of its theory.","headline":"Empirically strong, theoretically overclaimed restoration method; the surrogate-loss gap between Lre and the true reconstruction error is real.","tokens_in":27623,"tokens_out":4574,"would_cite":false,"duration_ms":40229,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","15A69"],"pacs":[],"model":"deepseek-v4-flash","headline":"A pruned network can be restored without any data or retraining by spreading each removed filter's role across the surviving filters.","keywords":["network pruning","filter pruning","data-free recovery","training-free restoration","closed-form solution","reconstruction error","batch normalization","model compression"],"falsifier":"Apply LBYL to a pretrained CNN and, in one deep layer, prune a filter whose nearest surviving filters have near-zero cosine similarity; then measure both the filter-domain residual $\\|E\\|_2$ and the true layer-wise reconstruction error (for instance the WARE metric) on a small held-out sample. If any layer shows a small residual with a large feature-map error, the proportionality step between Eq. (7) and Eq. (8) is invalid there and the closed form is minimizing a surrogate objective.","tokens_in":2105,"feed_emoji":"✂️","tokens_out":2311,"duration_ms":113629,"temperature":0.7,"pith_summary":"Network pruning normally buys a smaller model at the price of a long fine-tuning phase that needs the original training data. This paper claims that the damage from filter pruning can instead be repaired with no data and no training at all, by re-routing each pruned filter's contribution through a weighted combination of the filters that survive. The method, called LBYL (Leave Before You Leave), replaces the one-to-one compensation trick of finding a single similar neuron to stand in for the pruned one with a many-to-one delivery matrix. The paper proves that the resulting data-free loss has a closed-form solution, and reports that at 30% pruning LBYL restores an average of 17.92 percentage points of accuracy over Neuron Merging on ResNet-101, with gains that persist across pruning criteria and even random masks. If the claim holds, pruning pipelines could regain most of their lost accuracy before any fine-tuning budget is spent, and the recovered network is also a better starting point for fine-tuning when data does become available.","feed_headline":"Pruned networks regain 18 accuracy points with zero retraining","feed_subtitle":"Spreading a pruned filter's role across all surviving filters lifts accuracy far past one-to-one replacement.","key_machinery":"The central object is the delivery matrix $S^*$, a relaxation of the pruning matrix in which the row belonging to a pruned filter holds coefficients $s_k$ (one per surviving filter) instead of zeros, so that the next layer's weights are rebuilt as $W^{(\\ell+1)} \\times_2 S^{*\\top}$. The argument rides on two proved results: Theorem 1 decomposes the next-layer reconstruction error into $\\|\\frac{\\gamma_j}{\\sigma_j}(A^{(\\ell-1)} \\circledast E) + B + R\\|_1$, and Theorem 2 gives the closed-form minimizer of the data-free loss, which is the only per-filter computation in Algorithm 1. The analysis enters through the $n$-mode product formalism, and the batch-normalization statistics $\\gamma, \\sigma, \\mu, \\beta$ are folded into $X$ and $p$, so the entire recovery procedure reads only the pretrained weights and never sees a data sample.","core_discovery":"The paper's central claim is that a pruned convolutional layer can be restored, with no data and no training, by giving every pruned filter a delivery row of scalar coefficients $s_k$ that re-route its output through all surviving filters at once, and that the optimal coefficients are obtained in closed form from the pretrained weights and batch-normalization statistics. The reconstruction error between the original network and the approximation splits into three components—a filter-domain residual $E$, a batch-normalization error $B$, and an activation error $R$ whose $\\ell^1$ norm is bounded by a constant times the delivery coefficients—and Theorem 2 proves that minimizing $L_{\\mathrm{re}} = \\|E\\|_2^2 + \\lambda_1\\|B\\|_2^2 + \\lambda_2\\|s\\|_2^2$ is a convex problem with a unique solution $s = [X^\\top X + \\lambda_1 \\frac{\\gamma_j^2}{\\sigma_j^2} p p^\\top + \\lambda_2 I]^{-1}[X^\\top y + \\lambda_1 \\frac{\\gamma_j}{\\sigma_j}(\\frac{\\mu_j \\gamma_j}{\\sigma_j} - \\beta_j) p]$. Because a weighted combination of many survivors approximates the pruned filter far better than any single similar filter, the restored network keeps much more of its original accuracy: the authors report an average of 17.92 percentage points over the one-to-one Neuron Merging baseline on ResNet-101 at 30% pruning, with consistent gains across pruning criteria, ratios, architectures, and even random pruning masks.","pith_inferences":["The per-filter closed-form solve is effectively a ridge regression in the space of surviving filters, so its cost grows with layer width; a batched or iterative linear-solve variant would be the natural way to scale the method to very wide layers, and nothing in the derivation forbids it.","The proportionality step between the filter-domain residual and the true feature-map error is the most isolated testable link: a diagnostic that compares layer-wise WARE against the predicted RE+BE values on networks where survivor filters are nearly orthogonal to the pruned one would show exactly where the surrogate can mislead.","The delivery-matrix idea transfers beyond weight pruning: quantizing a layer or truncating a tensor decomposition also removes a component that could be re-expressed as a combination of the surviving components, though the three-term error split would need re-derivation for nonlinearities other than ReLU, which the paper itself names as future work.","Because every surviving filter participates, the method inherits the quality of the pruning mask; a joint criterion that prefers filters whose information is cheapest to redistribute (smallest achievable residual) should push restored accuracy higher than fixing the mask first."],"forward_implications":["At high pruning ratios, where one-to-one compensation collapses because no sufficiently similar survivor exists, LBYL keeps accuracy largely intact: on ResNet-101/ImageNet at 30% pruning it reports on average 17.92 points higher accuracy than Neuron Merging, and it also improves recovery under random pruning masks.","The same closed-form restoration carries over to fully-connected neuron pruning, where LBYL outperforms the training-free recovery of Coreset on LeNet-300-100/FashionMNIST by clear margins at high ratios.","Because the residual error and batch-normalization error are computable without data, the paper shows they can act as a data-free proxy for the layer-wise weighted average reconstruction error, enabling pruning-ratio selection without any data.","When fine-tuning data later becomes available, networks restored by LBYL converge in a few epochs and end higher than one-to-one restored networks after the same 20-epoch budget, indicating that the data-free step is a better initialization.","On a detection task, LBYL restores the SSD/ResNet-50 backbone to higher AP and AP50 than Neuron Merging on COCO2017, with the gap widening as the pruning ratio grows."],"supporting_citations":[{"why":"The one-to-one compensation baseline (Neuron Merging) that LBYL must beat; supplies the comparison method, its default hyperparameters, and the pretrained VGG-16 model used for CIFAR-10.","marker":"[21]"},{"why":"The other one-to-one data-free pruning-recovery approach in the same problem setting, whose similarity assumption LBYL is designed to relax.","marker":"[22]"},{"why":"Defines the n-mode tensor product that formalizes the pruning matrix and the delivery matrix in Eqs. (1)-(3).","marker":"[39]"},{"why":"Supplies the L2-GM pruning criterion used to generate pruned networks in all experiments.","marker":"[8]"},{"why":"Supplies the L2-norm pruning criterion used to generate pruned networks in all experiments.","marker":"[14]"},{"why":"Supplies the L1-norm pruning criterion used to generate pruned networks in all experiments.","marker":"[42]"},{"why":"Provides the ResNet architectures (ResNet-50, ResNet-101, ResNet-34) whose recovery results carry the main empirical claims.","marker":"[23]"},{"why":"The ImageNet benchmark behind the headline ResNet-101 and MobileNet-V2 recovery numbers.","marker":"[24]"}],"fun_headline_variants":["Pruned nets restored without training, data, or guesswork","Closed-form fix recovers pruned network accuracy","No retraining, no data: pruned filters rebuilt via closed form","Spreading pruned roles boosts accuracy by 18 points"],"cache_read_input_tokens":29696,"weakest_assumption_plain":"The method's load-bearing premise is that shrinking the filter-domain residual—the difference between the pruned filter and the weighted sum of survivors—automatically shrinks the true next-layer feature-map error, because the next layer's convolution weights are treated as a harmless constant multiplier.","fun_headline_variants_meta":{"raw":{"variants":["Pruned nets restored without training, data, or guesswork","Closed-form fix recovers pruned network accuracy","No retraining, no data: pruned filters rebuilt via closed form","Spreading pruned roles boosts accuracy by 18 points"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000992,"raw_usage":{"total_tokens":4309,"prompt_tokens":1155,"completion_tokens":3154,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":771,"completion_tokens_details":{"reasoning_tokens":3085}},"tokens_in":771,"tokens_out":3154,"duration_ms":21638,"temperature":1.0,"reasoning_tokens":3085,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T00:43:08.321953+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Apply LBYL to a pretrained CNN and, in one deep layer, prune a filter whose nearest surviving filters have near-zero cosine similarity; then measure both the filter-domain residual $\\|E\\|_2$ and the true layer-wise reconstruction error (for instance the WARE metric) on a small held-out sample. If any layer shows a small residual with a large feature-map error, the proportionality step between Eq. (7) and Eq. (8) is invalid there and the closed form is minimizing a surrogate objective.","supporting_citations":[{"cited_title":"Neuron merging: Compensating for pruned neurons,","cited_arxiv_id":null,"evidence_quote":"The one-to-one compensation baseline (Neuron Merging) that LBYL must beat; supplies the comparison method, its default hyperparameters, and the pretrained VGG-16 model used for CIFAR-10."},{"cited_title":"Data-free parameter pruning for deep neural networks,","cited_arxiv_id":null,"evidence_quote":"The other one-to-one data-free pruning-recovery approach in the same problem setting, whose similarity assumption LBYL is designed to relax."},{"cited_title":"Filter pruning via geometric median for deep convolutional neural networks acceleration,","cited_arxiv_id":null,"evidence_quote":"Supplies the L2-GM pruning criterion used to generate pruned networks in all experiments."},{"cited_title":"Soft filter pruning for accelerating deep convolutional neural networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the L2-norm pruning criterion used to generate pruned networks in all experiments."},{"cited_title":"Pruning filters for efficient convnets,","cited_arxiv_id":null,"evidence_quote":"Supplies the L1-norm pruning criterion used to generate pruned networks in all experiments."},{"cited_title":"Deep residual learning for image recognition,","cited_arxiv_id":null,"evidence_quote":"Provides the ResNet architectures (ResNet-50, ResNet-101, ResNet-34) whose recovery results carry the main empirical claims."},{"cited_title":"Imagenet: A large-scale hierarchical image database,","cited_arxiv_id":null,"evidence_quote":"The ImageNet benchmark behind the headline ResNet-101 and MobileNet-V2 recovery numbers."}],"review_version":1}