{"id":"78083bce-f197-40e9-b5f0-513e04d15ad3","arxiv_id":"2411.16901","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A gravity-inspired regularization method for structured pruning is proposed, but its no-retraining results show large accuracy drops and its fine-tuned comparisons use lower baselines than competing methods.","lead":"This paper adds a 'gravity' penalty to neural network training that drives some convolution filters toward zero so they can be pruned. It claims this supports pruning at many ratios with no retraining, but its own results show large accuracy drops and its comparisons rely on lower baselines.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The no-retraining claim is contradicted by the paper's own Table I, where 10% pruning drops ResNet-56/CIFAR-10 accuracy from 93.39% to 74.52% with no fine-tuning.","rationale":"I focused on the strongest part of the central claim: compression at variable pruning ratios without retraining. The paper tests this directly in Table I, and the measured drops (e.g., 93.39% -> 74.52% at 10% pruning on ResNet-56/CIFAR-10) are far outside any reasonable notion of accuracy preservation. The later comparisons in Tables III and IV include fine-tuning, so they cannot rescue the no-retraining claim. I also checked the derivation and found an internal contradiction: Eq. 4 replaces the index distance d with its reciprocal, so the force term in Eq. 10 grows as the filter moves away from the attracting filter, yet Section VI.C states that larger d decreases the penalty. This does not change the verdict, but it further weakens confidence in the method's explanation. I partially agree with the reader: their weakest_assumption about using index distance as a pruning coordinate is plausible, but the more decisive failure is the direct contradiction between the headline claim and the paper's own Table I. Therefore the reader's REJECT verdict should stand.","tokens_in":8778,"tokens_out":5824,"duration_ms":52486,"concrete_test":"Reproduce the exact Table I protocol: take a pretrained ResNet-56 on CIFAR-10, train it with the gravity regularizer at alpha_g=1e5 (G=6.7e-11), prune with Algorithm 2 at ratios 10%, 20%, 30%, 40%, 50%, and evaluate immediately after pruning with no fine-tuning. If the 10%-pruned accuracy is near 74.5%, the no-retraining claim is empirically refuted under the authors' own setup. If it is within 1-2 points of the 93.39% baseline, then Table I is erroneous and the claim deserves re-evaluation. The same check should be run with at least three random seeds to rule out seed sensitivity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central contribution is the claim that a gravity-trained DCNN can be pruned at arbitrary ratios without retraining while preserving accuracy. Table I is the direct test of that claim and it fails: pruning ResNet-56/CIFAR-10 at only 10% (no fine-tuning) drops accuracy by 18.87 points, and VGG-19/CIFAR-100 drops by 10.37 points. These are not 'preserving accuracy' results, and Discussion VI.A concedes that 40% pruning reduces accuracy by roughly 37%. The favorable numbers in Tables III and IV are reported after standard fine-tuning, which is exactly the stage the method claims to eliminate. Figure 2 is the only evidence for a no-retraining curve, but it is presented without code, seeds, or error bars, so it cannot be checked. There is also an internal inconsistency in the mechanism: Eq. 4 defines d = 1/|p1-pn|, making the penalty grow with index distance, while Section VI.C says larger d decreases the penalty. The decisive issue remains that the headline claim is contradicted by the paper's own experimental section.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes a structured pruning method for DCNNs based on a 'gravity' regularizer added during training. Filters are assigned masses equal to their L1 norms, an attracting filter is chosen as the largest-mass filter, and a penalty term proportional to the product of masses and inversely proportional to a squared distance (defined as the reciprocal of the index difference) is added to the loss. The authors claim that after training, filters can be pruned at arbitrary ratios without retraining. They report experiments on CIFAR-10/100 with ResNet-56 and VGG-19, including pruning-only results (Table I), fine-tuned comparisons with existing methods (Tables III–IV), and ablation studies over the gravity rate. The paper concludes that the method achieves competitive results while eliminating the fine-tuning stage.","tokens_in":9018,"tokens_out":6834,"duration_ms":58093,"significance":"If the no-retraining claim were true, this would be an attractive, architecture-agnostic approach because it would decouple the pruning ratio from the training procedure. However, the paper provides no code, no error bars, and no seed information; its only evidence for the headline claim is a set of curves (Figure 2) that cannot be checked. The fine-tuned comparisons in Tables III and IV are weakened by the fact that gravity training itself lowers the baseline accuracy, so the reported 'accuracy drop' values are relative to a weaker model rather than to the standard baselines used by the comparison methods. The manuscript therefore does not currently provide a sound basis for its claimed advantage.","major_comments":[{"comment":"The central claim that the method preserves accuracy without retraining is contradicted by the paper's own Table I: pruning ResNet-56/CIFAR-10 at 10% drops accuracy from 93.39% to 74.52% (an 18.87-point drop), and VGG-19/CIFAR-100 drops from 73.76% to 63.39% (a 10.37-point drop). Section VI.A concedes that 40% pruning reduces baseline accuracy by about 37%. The only support for the no-retraining claim is Figure 2, which is presented without error bars, seeds, or code; the headline claim must be either removed or replaced by evidence from a reproducible experimental protocol.","section":"V.A, Table I"},{"comment":"The comparisons against existing methods are made after fine-tuning, which is exactly the stage the method claims to eliminate, and they compare accuracy drops relative to weakened gravity-trained baselines. For example, the VGG-19 Gravity (p) baseline is 70.99%, 2.77 points below the standard VGG-19 baseline of 73.76% used in Table I, and the ResNet-56 Gravity (r) baseline is 88.02% versus roughly 93.5%. Several 'accuracy drops' are therefore negative (e.g., Gravity (r) at 2.17x speedup in Table III), which does not indicate improved pruning but rather recovery from a degraded baseline after fine-tuning. These comparisons do not support the claim of 'minimal accuracy drop compared to all existing methods.'","section":"V.C, Tables III and IV"},{"comment":"The gravity rate α_g is selected as the value yielding the best pruning curves in Figure 2 (Section VI.D), and the same curves are used to demonstrate the method's headline no-retraining performance. No separate validation set or explicit model-selection procedure is described, so the reported pruning results are optimistically biased and the sensitivity of the method to α_g is not honestly assessed.","section":"VI.D and Figure 2"},{"comment":"The distance formulation is internally inconsistent. Equation (4) defines d = 1/|p1-pn|, making the penalty in Eq. (5) proportional to |p1-pn|²; yet Section VI.C states that 'When the distance d is larger, the penalty term decreases' and 'when d is smaller, the penalty term increases.' Under Eq. (4), a larger d corresponds to a smaller index separation, so the explanation is at odds with the usual meaning of distance and the derivation of the pruning behavior is not self-consistent. In addition, using the filter index as a spatial coordinate is not justified because filters in a convolutional layer are permutation-invariant; an index-based distance defines an arbitrary ordering with no known relation to functional importance.","section":"III.B and VI.C"}],"minor_comments":[{"comment":"The gradient derivation treats m1 and d as constants when differentiating the penalty term, but the attracting filter (and hence m1 and the index ordering) can change during training; the authors should state explicitly that this is an approximation or discuss its effect on the learning dynamics.","section":"III.B, Eq. (9)"},{"comment":"The reference to the CIFAR dataset cites reference [2] (a paper on electrostatic-force pruning) rather than a dataset reference; the dataset citation is missing.","section":"IV.A"},{"comment":"Section IV.B says the fine-tuning uses 'the same hyperparameters as other pruning methods [7]' but does not specify what those hyperparameters are; a concrete list is needed for reproducibility.","section":"IV.B"},{"comment":"Figure 1 is not referenced in the text, and Figure 2 is never described in terms of the number of runs or variance; both should be integrated into the narrative with proper captions.","section":"V.A, Figure 1 and 2"},{"comment":"The notation d is used both for the index difference in Eq. (3) and its reciprocal in Eq. (4); distinct symbols or an explicit naming convention would remove avoidable ambiguity.","section":"III.B"}],"recommendation":"reject","confidential_remarks":"The paper overlaps heavily with the author's own 'Electrostatic force regularization' paper (reference [2]) in motivation, experimental setup, and presentation; the relationship is not discussed. The experimental reporting in Tables III and IV is misleading because the baselines are weakened by gravity training, and the only evidence for the no-retraining claim is an unverifiable figure. In my view, the contribution does not meet the bar for publication in its current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The key thing to know: the paper's central promise — prune at any ratio without retraining while preserving accuracy — is not just unproven; it is refuted by the paper's own Table I. Pruning ResNet-56/CIFAR-10 at only 10% drops accuracy from 93.39% to 74.52%, and VGG-19/CIFAR-100 drops from 73.76% to 63.39% at the same ratio. Those are not \"preserving accuracy\" numbers, and the Discussion even concedes that 40% pruning costs roughly 37% accuracy. So the core contribution, as stated, collapses on the evidence provided.\n\nWhat the paper does well: it is clearly written, gives precise pseudocode for training and pruning, and is upfront about training overhead (80% slower for ResNet-56). The fine-tuned results in Tables IV and V are competitive with existing methods, and the authors include a hyperparameter sweep. The mathematical description is internally consistent once you notice that d in Eq. 4 is the reciprocal of the index distance, so the \"larger d\" language in Section VI.C is not an inconsistency — the stress-test note claiming otherwise misreads that.\n\nThe soft spots are substantial. The no-retraining claim is the reason to care, and it fails. The fine-tuned comparisons use weakened baselines: the gravity-trained VGG-19 baseline is 70.99% versus roughly 73.8% for other methods, so the accuracy drop in Table IV is measured from a lower starting point, flattering the result. The gravity-rate hyperparameter is selected by inspecting Figure 2, which is the very curve used to report the headline pruned accuracies — that is fitting to the evaluation. No code, seeds, or error bars are supplied, so Figure 2 cannot be checked. Novelty is thin: this is essentially the author's earlier electrostatic-force regularization (Ref. [2]) with Coulomb's law swapped for Newton's, which is the same inverse-square structure on index distance. The index distance itself is asserted as a meaningful coordinate without any supporting analysis.\n\nWho is this for? Someone cataloging simple training-time regularizers for filter pruning might find the idea worth a footnote, but they should not rely on the no-retraining claim. The paper is a coherent write-up of a plausible-sounding method that its own experiments show does not work as advertised. A serious editor would desk-reject it: the primary contribution is self-contradicted, and the residual fine-tuned results are not novel enough to justify referee time.","headline":"The paper's headline no-retraining pruning claim is directly contradicted by its own Table I; the method is a minor variant of the author's earlier electrostatic force work, though the write-up is clear and the fine-tuned results are competitive.","tokens_in":9549,"tokens_out":3479,"would_cite":false,"duration_ms":34956,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A gravity-inspired regularizer during training makes convolutional filters prunable at any ratio without retraining.","keywords":["structured pruning","deep convolutional neural networks","gravity regularization","L1 norm","filter pruning","CIFAR","ResNet-56","VGG-19"],"falsifier":"Train the same architecture twice with the same data and hyperparameters, but in the second run randomly permute the order of filters within each convolutional layer before gravity training begins; if pruning the two models at the same ratio yields essentially the same accuracy, the index-distance term is not responsible for the method's effect.","tokens_in":1557,"feed_emoji":"🧲","tokens_out":2741,"duration_ms":86350,"temperature":0.7,"pith_summary":"The paper claims that adding a gravity-inspired penalty to the training loss of a convolutional neural network prepares the network for structured pruning at any desired ratio, with no retraining and no change to the architecture. The penalty pulls filters far from the layer's largest $\\ell^1$ filter toward zero weights while letting nearby filters keep their weights, so that after training a simple $\\ell^1$-norm ranking identifies which filters to delete. If the claim holds, a single trained model could be deployed at several compression levels, and pruning stops being a separate post-training search-and-fine-tune stage. The paper reports experiments on ResNet-56 with CIFAR-10 and VGG-19 with CIFAR-100, comparing against existing structured pruning methods.","feed_headline":"Gravity loss prunes CNNs at any ratio without retraining","feed_subtitle":"A training-time gravitational force pushes far filters to zero, so pruning is just ranking filters by magnitude.","key_machinery":"The key object is the gravity regularizer $F_{n,l} = G m_{1,l} m_{n,l} / d_{n,l}^2$, appended to the standard loss as a penalty with strength $\\alpha_g$. Filter mass is the $\\ell^1$ norm of the filter weights, and the distance between filters is the reciprocal of their absolute index difference, so the regularizer is strongest for filters that the final pruning step will rank as least important. The attracting filter is the filter with the largest mass in the layer, which exerts no force on itself. This device converts structured pruning into a training-time weight redistribution: it uses the same $\\ell^1$ norm both as 'mass' during training and as the pruning criterion afterward, so the ranking and the removal step are automatically aligned.","core_discovery":"The central discovery is that a Newton-style gravitational force, applied as a regularizer during training, can structure the weight distribution of convolutional filters so that filters located far from an attracting filter--the filter with the largest $\\ell^1$ mass in the layer--are driven to zero, while filters close to it remain dense. The force is defined by $F = G m_1 m_n / d^2$, with mass $m$ equal to the $\\ell^1$ norm of a filter's weights and with distance $d$ taken as the reciprocal of the absolute index difference between filters. Minimizing the regularized objective therefore multiplies a $\\operatorname{sign}(w)$ penalty by $G m_1 / d^2$, making the penalty strongest for filters whose index lies far from the attracting filter. After training, the method prunes filters with the smallest $\\ell^1$ norms at a uniform rate across layers; because the weight distribution already separates important from unimportant filters, the same trained model can be pruned at different ratios without retraining. The paper presents this as a way to configure a model for pruning during training rather than after it.","pith_inferences":["If the method works as described, the index-distance geometry may be replaceable: any ordering that assigns larger penalties to filters the network can afford to lose would produce the same effect, so the physics analogy could be swapped for a simpler per-filter penalty schedule.","The 'no retraining' claim is demonstrated only for small CIFAR models; extrapolating to ImageNet-scale networks or to detection and segmentation heads would require testing, since redundancy patterns differ.","A direct test of the distance mechanism would be to randomly permute filter indices before gravity training; if pruning accuracy is unchanged, the attraction is not coming from index distance.","The method's value over standard $\\ell^1$ regularization may come from its position-dependent penalty rates: filters in different index neighborhoods are penalized differently, which can be seen as a hand-designed sparsity curriculum."],"forward_implications":["A gravity-trained model can be compressed to different sizes from a single training run, since the pruning ratio is chosen after training and no retraining is needed when the ratio changes.","The method applies to both randomly initialized and pretrained weights, and requires no modification to the original architecture.","The pruning step is a local $\\ell^1$-norm ranking within each layer, so implementation is simple and the reported speedup and compression ratios grow with the pruning ratio.","The training overhead is nontrivial, about 80 percent longer than baseline for ResNet-56 on CIFAR-10 in the reported settings, so the benefit is a tradeoff against training cost.","After optional fine-tuning, the pruned models are competitive with several existing structured pruning methods at comparable speedups on CIFAR-10 and CIFAR-100."],"supporting_citations":[{"why":"The predecessor electrostatic-force pruning method that the gravity formulation directly extends and compares against.","marker":"[2]"},{"why":"The torque-based structured pruning baseline used as a state-of-the-art comparison in the pruning tables.","marker":"[3]"},{"why":"The growing regularization method used as a comparison baseline and whose hyperparameters are reused for the fine-tuning step.","marker":"[7]"},{"why":"The soft filter pruning baseline compared at similar speedups.","marker":"[10]"},{"why":"The EigenDamage and Kronecker-factor structured pruning baselines compared on CIFAR-100 with VGG-19.","marker":"[17]"}],"fun_headline_variants":["Gravity pulls far filters to zero for instant pruning","Train once, prune anytime: gravity-based CNN pruning","Physics-inspired pruning: rank filters by gravitational pull","Gravity regularizer zeroes distant filters, ready for pruning","CNN pruning made simple: gravity decides which filters die"],"cache_read_input_tokens":11648,"weakest_assumption_plain":"The load-bearing premise is that a filter's position within its layer is a meaningful coordinate for pruning importance, so that filters far from the largest $\\ell^1$ filter can be driven to zero without hurting accuracy.","fun_headline_variants_meta":{"raw":{"variants":["Gravity pulls far filters to zero for instant pruning","Train once, prune anytime: gravity-based CNN pruning","Physics-inspired pruning: rank filters by gravitational pull","Gravity regularizer zeroes distant filters, ready for pruning","CNN pruning made simple: gravity decides which filters die"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000168,"raw_usage":{"total_tokens":1275,"prompt_tokens":974,"completion_tokens":301,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":224}},"tokens_in":590,"tokens_out":301,"duration_ms":3582,"temperature":1.0,"reasoning_tokens":224,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:45:17.483533+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same architecture twice with the same data and hyperparameters, but in the second run randomly permute the order of filters within each convolutional layer before gravity training begins; if pruning the two models at the same ratio yields essentially the same accuracy, the index-distance term is not responsible for the method's effect.","supporting_citations":[{"cited_title":"Torque based structured pruning for deep neural network,","cited_arxiv_id":null,"evidence_quote":"The torque-based structured pruning baseline used as a state-of-the-art comparison in the pruning tables."},{"cited_title":"Neural pruning via growing regularization,","cited_arxiv_id":null,"evidence_quote":"The growing regularization method used as a comparison baseline and whose hyperparameters are reused for the fine-tuning step."},{"cited_title":"Eigen- damage: Structured pruning in the kronecker-factored eigenbasis,","cited_arxiv_id":null,"evidence_quote":"The EigenDamage and Kronecker-factor structured pruning baselines compared on CIFAR-100 with VGG-19."}],"review_version":1}