{"id":"fead6d47-8dff-4b14-ae30-cfd85383a714","arxiv_id":"2411.11079","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A Coulomb's-law-inspired regularizer is proposed for structured pruning, but the claimed repulsion and attraction mechanism is not supported by the paper's gradient derivation.","lead":"This paper proposes a training-time penalty based on an electrostatic force analogy to make convolutional filters prunable after training. The authors claim simpler pruning without fine-tuning, but the paper's own equations do not support the claimed attraction and repulsion mechanism.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The derivation of Eq. 10 from Eq. 6 is invalid: the gradient of the L1-norm-based charge is sign(w), not w, and the distance term contributes too. Since Eq. 11 is the paper's only mechanism for attraction/repulsion, the central pruning claim is unsupported.","rationale":"The reader's weakest-assumption analysis correctly identifies the invalid transition from Eq. 6 to Eq. 10 as the load-bearing flaw. The central claim of the paper depends on Eq. 11 implementing a signed, distance-dependent force that attracts opposite-charge filters to non-zero weights and repels same-charge filters to zero. The actual partial derivative of Fe with respect to a weight is proportional to sign(w_i), not w_i, and also contains derivatives of the distance r. Consequently, Eq. 11 does not follow from Eq. 6. The method as written is either an L1 regularizer with per-filter coefficients (if the gradient is corrected) or an L2 regularizer (if Eq. 10 is implemented literally). In both cases every filter is shrunk toward zero; the alleged attraction is at best a smaller shrinkage rate for filters with large r. That undermines the novelty and the mechanism, not merely the presentation. The empirical tables cannot compensate because no code, error bars, or ablations are provided to show the claimed bimodal filter-norm behavior, and Figure 3 reports only aggregate normalized L1 norms. The reader's REJECT verdict is therefore appropriate, and my stress-test does not change it.","tokens_in":12221,"tokens_out":3462,"duration_ms":74436,"concrete_test":"Run a minimal two-filter experiment with the correctly derived gradient of Fe from Eq. 6 and compare against Eq. 10's literal gradient. If no filter's L1 norm increases while an opposite-charge filter is present, the attraction claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. 6 defines Fe = ke |q1| |q| / r^2 with |q| = ||W||_1 and r = |q1 - q|. The paper then asserts in Eq. 10 that the gradient contribution is alpha_e ke |q1| / r^2 times (w). This does not follow. For the L1 norm, d|q|/dw_i = sign(w_i), and r depends on q, whose derivative is also nonzero. The correct gradient includes terms proportional to sign(w_i) and additional terms from dr/dw_i; it is not proportional to the weight vector w. Thus Eq. 11, which is the sole theoretical justification for the claimed electrostatic mechanism, is not derived from the stated objective. Moreover, even if one ignores the algebra, the regularizer is a penalty on |q| (or, if Eq. 10 is implemented literally, a penalty on ||W||_2^2). In either case every filter's weights are pushed toward zero; the 'attractive' case is at most a weaker shrinkage when r is large, never a force toward non-zero weights. Filter sign enters only through the distance r, not through a signed force direction. The central claim that opposite-charge filters are preserved by attraction and same-charge filters are driven to zero is therefore not supported by the equations as written. The empirical comparisons in Tables 4-6 cannot rescue this because the mechanism is unverified and no code or ablations isolate the alleged force behavior.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a structured pruning method for deep convolutional neural networks in which an electrostatic-force-inspired regularizer is added to the training loss. The charge of a filter is defined as the sign of the sum of its weights multiplied by the L1 norm of the weights, the source filter is the filter with the largest L1 norm, and the force between filters is defined by Fe = ke |q1||q|/r^2, with r the absolute difference of the two scalar charges. The authors claim that minimizing this regularizer causes filters with like charges to shrink toward zero while filters with opposite charges grow toward non-zero values, allowing filters to be pruned after training without architectural changes or retraining. Experiments are reported on MNIST, CIFAR-10, CIFAR-100, and ImageNet with ResNet and VGG architectures, comparing pruned accuracy and speedup against several prior pruning methods.","tokens_in":12511,"tokens_out":3452,"duration_ms":36506,"significance":"If the proposed mechanism worked as described, the method would be practically attractive: it promises simple structured pruning at arbitrary ratios without fine-tuning and with no architectural modification. The paper also provides an interesting analogy between Coulomb's law and filter importance. However, the central mathematical derivation is incorrect: the gradient of the proposed regularizer is not what the paper claims, and the update rule in Eq. 11 does not realize the advertised attraction/repulsion behavior. Because the pruning mechanism is unsupported and no code or targeted ablations are provided to verify the claimed force dynamics, the empirical results cannot establish the central claim. The paper does include comparisons to several published pruning methods and reports training overhead, which are useful, but these do not compensate for the unsupported mechanism.","major_comments":[{"comment":"The derivation from Eq. 6 to Eq. 10 is not a valid gradient computation. From the definition Fe = ke |q1||q|/r^2, with |q| = ||W||_1 and r = |q1 - q|, the partial derivative of Fe with respect to a weight w_i is not proportional to w_i. The derivative of the L1 norm is sign(w_i) (where defined), and the distance r depends on q, so its derivative contributes additional terms through dq/dw_i. Therefore Eq. 10, and consequently the update rule in Eq. 11, do not follow from the stated objective.","section":"Section 3.2, Eqs. (6)-(11)"},{"comment":"Even if Eq. 10 were taken as the implemented update rather than as a derivative, the force term alpha_e ke |q1|/r^2 (w) is a shrinkage term proportional to w: it pushes all weights toward zero, with the magnitude of the push depending on the distance r. It contains no signed direction that would pull opposite-sign filters toward non-zero weights. The 'attractive' case is at most a weaker shrinkage when r is large, not a force toward non-zero values. Thus the paper's central claim that same-sign filters are repelled to zero while opposite-sign filters are attracted to non-zero values is not supported by the equations as written.","section":"Section 3.2, Eq. 11"},{"comment":"The hyperparameter alpha_e is selected per model and dataset after observing the pruning curves. Section 6.4 states that the 'appropriate' values were determined from Figure 2, and Table 2 uses different alpha_e values for ResNet-56, VGG-19, and ResNet-34. Because alpha_e controls the strength of the alleged force and is tuned to produce good pruning results, the reported comparisons partly demonstrate hyperparameter fitting rather than a validated mechanism. Without an ablation that isolates the claimed attraction/repulsion behavior or code to reproduce the force dynamics, the experiments cannot confirm the mechanism.","section":"Section 6.4 and Table 2"},{"comment":"The claim that the method yields 'superior top-1 accuracy after pruning compared to all existing methods' is overstated because the comparison groups methods with different speedups and different baseline accuracies. For example, at 2.17x speedup the electrostatic force (p) model reaches 93.88% pruned accuracy, but Torque at 2.15x reaches 93.76% and ABC Pruner at 2.18x reaches 93.23%; at 2.62x the electrostatic force (r) model drops to 93.04%, which is worse than several methods at similar speedups. A direct comparison under matched speedup and baseline is needed to support the comparative claim.","section":"Section 6.3.1, Table 4"}],"minor_comments":[{"comment":"The sentence beginning 'In the rest of this paper. In Section 2...' contains a period after 'paper' and should be rewritten as a single grammatical sentence.","section":"Section 1"},{"comment":"The value '74,49%' uses a comma as a decimal separator, which is inconsistent with the decimal-point notation used throughout the rest of the manuscript.","section":"Section 6.1"},{"comment":"The line 'if n = source_f ilter_index or (nl) = 0then' contains a formatting artifact and should read 'if n = source_filter_index or sign(q_n,l) = 0 then'.","section":"Algorithm 1"},{"comment":"The Figure 2 caption mentions 'four distinct electrostatic force rates', while Section 6.4 reports experiments with three values of alpha_e (10^-11, 10^-12, and 10^-14). The manuscript should reconcile this discrepancy.","section":"Figure 2 caption and Section 6.4"},{"comment":"The author name 'Y oucef Ferdi' contains a stray space and should be corrected to 'Youcef Ferdi'.","section":"Author list"}],"recommendation":"reject","confidential_remarks":"The manuscript's central mechanism is derived from an invalid gradient computation, and the update rule in Eq. 11 does not implement the advertised attraction/repulsion. The reported experiments are not sufficient to validate the mechanism because the key hyperparameter is tuned per experiment and no code or targeted ablations are provided. I recommend rejection, as fixing the core derivation would require redefining the method and re-running the empirical study, which is beyond a standard revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, the paper's central claim is not backed by its own equations: Eq. 10 does not follow from Eq. 6, and the update rule in Eq. 11 is just per-filter weighted L2 shrinkage. Second, the experiments are more extensive than the analysis: ResNet/VGG on MNIST/CIFAR and ResNet-34 on ImageNet, with a L1-norm comparison, all without fine-tuning. If the regularization works, it works for a more mundane reason than electrostatics.\n\nThe genuinely new thing is the Coulomb analogy for structured pruning, and the practical setup: a training-time penalty that lets you prune at arbitrary ratios without retraining. That is useful for edge deployment. The paper also compares against L1-norm and shows better pruning accuracy at comparable speedups, though at higher training cost.\n\nThe soft spots are serious, not cosmetic. The gradient of F_e = k_e |q1||q_n|/r^2 with |q_n| = ||W_n||_1 is proportional to sign(w_i), not w_i, and r depends on q_n, so its derivative contributes too. Eq. 10 replaces that with (w), which is invalid. The implemented Eq. 11 shrinks all weights toward zero; there is no mechanism that pushes opposite-sign filters to non-zero values. The 'attraction' case is at best weaker shrinkage. So the abstract's promise of attraction/repulsion is not realized by the math. The empirical comparisons also overstate: base accuracies differ across methods, some SOTA rows show negative accuracy drops (improvements), and the claim that ours 'yields superior top-1 accuracy' holds only if you ignore those rows. Alpha_e is tuned per model and dataset after seeing results, and no code or error bars are given, so the pruning numbers are not independently checkable.\n\nWho is this for? Someone thinking about physics-inspired regularizers in compression might read it as a cautionary example. The paper deserves a serious referee because the underlying question—can a simple weighted L2 penalty enable pruning without fine-tuning—is testable and the experiments could be informative if the framing were corrected. As written, the central mechanism is unsupported; I would send to review but expect major revision or rejection unless the authors fix the derivation and release code.","headline":"A Coulomb-style pruning story that fails on its own math: the derivation is invalid and the update rule is just weighted L2 shrinkage, though the experimental setup is more substantial than the theory.","tokens_in":13084,"tokens_out":3765,"would_cite":false,"duration_ms":36534,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that adding an electrostatic-force penalty to the training loss drives some convolutional filters' weights to zero and preserves others, so the network can be pruned at arbitrary ratios without retraining or architecture…","keywords":["structured pruning","electrostatic force","DCNNs","filter pruning","regularization","model compression","no fine-tuning"],"falsifier":"Train one convolutional layer on MNIST with the electrostatic regularizer and plot the fraction of repelled-filter $\\ell^1$-norms that reach machine-precision zero over epochs. Independently, differentiate $F_{e,n,l}$ with respect to a weight and compare the result with the update rule in the paper: if the analytic gradient is not proportional to the weight itself (or if repelled norms plateau far from zero), the claimed mechanism fails.","tokens_in":11934,"feed_emoji":"⚡","tokens_out":9569,"duration_ms":80767,"temperature":0.7,"pith_summary":"The paper proposes a structured-pruning method that treats each convolutional filter as a charged object. A source filter with the largest $\\ell^1$-norm creates an electrostatic field: filters with the same charge are pushed toward zero weights, while filters with the opposite charge are pulled toward non-zero weights. After training with this penalty, the zero-weight filters can be deleted at any pruning ratio, with no architectural changes and no lengthy fine-tuning. The authors report pruned accuracies competitive with existing structured-pruning methods on ResNet and VGG models over MNIST, CIFAR, and ImageNet.","feed_headline":"Electrostatic force lets neural nets be pruned without retraining","feed_subtitle":"A Coulomb-style training penalty drives redundant filters to zero, preserving accuracy on MNIST, CIFAR, and ImageNet.","key_machinery":"The central object is the electrostatic force penalty $F_{e,n,l} = k_e |q_{1,l}| |q_{n,l}| / r_{n,l}^2$, with $q_{n,l} = \\mathrm{sgn}(n,l) \\|W_{n,l}\\|_1$ and $r_{n,l} = |q_{1,l} - q_{n,l}|$. The source filter $q_1$ is the layer's largest-$\\ell^1$-norm filter. The paper's claimed mechanism is the gradient update $w \\leftarrow w - \\epsilon(\\nabla_w J + \\alpha_e k_e |q_1| r^{-2} w)$, which supposedly drives same-sign filters to zero and opposite-sign filters to larger magnitudes. This penalty is what couples weight optimization with filter-importance ranking.","core_discovery":"The paper's central claim is that a Coulomb-style penalty added to the training objective reorganizes a convolutional layer's filters into two populations: filters whose charge matches the source filter (the filter with the largest $\\ell^1$-norm) feel a repulsive force and their weights shrink toward zero, while filters of opposite charge feel an attractive force and keep non-zero weights. Concretely, the regularized loss is $\\tilde{J} = \\sum J + \\alpha_e F_e$ with $F_{e,n,l} = k_e |q_{1,l}| |q_{n,l}| / r_{n,l}^2$, where charges are the product of filter sign and $\\ell^1$-norm and $r$ is the absolute difference of charges. The authors claim that after training with this objective, deleting the repelled filters by $\\ell^1$-norm ranking yields a pruned network with minimal accuracy loss, without architecture changes and without an extensive fine-tuning stage. They present the method as the first application of electrostatic charge concepts to DCNN training for pruning, and report competitive results on MNIST, CIFAR, and ImageNet across ResNet and VGG architectures.","pith_inferences":["A natural extension is to apply the same charge-and-distance penalty to other structured units—transformer attention heads, MLP neurons, or convolutional blocks—wherever a per-unit $\\ell^1$-norm can serve as the charge magnitude.","The claimed pruning-ratio flexibility suggests a workflow in which one electrostatic-trained model is deployed at different speed/accuracy trade-offs without retraining; this could be tested by training once and evaluating a ladder of pruning ratios on a held-out set.","Because the force magnitude depends on the difference of filter charges, layers with many filters of similar magnitude may require per-layer tuning of $\\alpha_e$; an adaptive scheme that sets $\\alpha_e$ from the spread of $\\ell^1$-norms is a possible improvement the paper does not explore."],"forward_implications":["A single training run yields a model that can be pruned at multiple ratios without retraining, because the regularization already separates disposable from informative filters.","Pruning removes whole filters rather than individual weights, so the gains translate directly into fewer FLOPs and less memory with no architectural changes.","The method works from both random and pretrained initialization, with pretrained weights giving higher pruned accuracy on the reported benchmarks.","On ImageNet, the reported $1.34\\times$ speedup for ResNet-34 keeps the top-1 drop at $0.19\\%$, a level the authors say beats or matches several baselines at larger speedups."],"supporting_citations":[{"why":"The primary regularization baseline and the source of the pruning schedule used for layer-wise ratios in the comparison tables.","marker":"[15]"},{"why":"A state-of-the-art structured-pruning baseline compared at similar speedups on CIFAR.","marker":"[22]"},{"why":"The Taylor-FO importance baseline for ImageNet, used to position the method's accuracy-FLOPs trade-off.","marker":"[37]"},{"why":"The classic L1-norm filter pruning method, both a baseline in the ImageNet table and the ranking criterion the electrostatic method builds on.","marker":"[38]"},{"why":"FPGM baseline in the CIFAR-10 comparison, representing geometric-median based filter pruning.","marker":"[29]"},{"why":"Soft filter pruning baseline in the CIFAR-10 comparison, used for the same speedup brackets.","marker":"[30]"}],"fun_headline_variants":["Coulomb force prunes neural nets with no fine-tuning","Physics-based penalty trims filters, skips retraining","Electrostatic force removes redundant filters in CNNs","Charge-based pruning cuts training time","Attract and repel: physics trick prunes CNNs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pruning mechanism rests on the claim that minimizing the penalty term pushes repelled filters' weights to zero and attracted filters' weights to larger magnitudes; if the effective gradient does not have that direction, the clean separation of disposable and informative filters does not follow.","fun_headline_variants_meta":{"raw":{"variants":["Coulomb force prunes neural nets with no fine-tuning","Physics-based penalty trims filters, skips retraining","Electrostatic force removes redundant filters in CNNs","Charge-based pruning cuts training time","Attract and repel: physics trick prunes CNNs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000238,"raw_usage":{"total_tokens":1546,"prompt_tokens":1018,"completion_tokens":528,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":634,"completion_tokens_details":{"reasoning_tokens":453}},"tokens_in":634,"tokens_out":528,"duration_ms":105016,"temperature":1.0,"reasoning_tokens":453,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:55:40.187182+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train one convolutional layer on MNIST with the electrostatic regularizer and plot the fraction of repelled-filter $\\ell^1$-norms that reach machine-precision zero over epochs. Independently, differentiate $F_{e,n,l}$ with respect to a weight and compare the result with the update rule in the paper: if the analytic gradient is not proportional to the weight itself (or if repelled norms plateau far from zero), the claimed mechanism fails.","supporting_citations":[{"cited_title":"Neural pruning via growing regularization","cited_arxiv_id":null,"evidence_quote":"The primary regularization baseline and the source of the pruning schedule used for layer-wise ratios in the comparison tables."},{"cited_title":"Torque based structured pruning for deep neural network","cited_arxiv_id":null,"evidence_quote":"A state-of-the-art structured-pruning baseline compared at similar speedups on CIFAR."}],"review_version":1}