{"id":"c96b2bc3-2d28-43d7-9d43-3b1442d87160","arxiv_id":"2607.25529","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"Weight magnitude is a weak and nonlinear proxy for per-weight importance in CNNs, but the paper's quantitative claims are undermined by a mislabeled metric and an unconventional definition of 'neuron'.","lead":"The paper tests whether the largest-magnitude weights in image-classification networks are the most important ones. It finds only about 25% overlap with accuracy-critical weights, but its headline '45-80% accuracy degradation' numbers come from a misread relative metric, and its 'neurons' are actually single scalar weights.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The experiments never ablate or perturb neurons: they operate on individual scalar connection weights, so the paper's central claim about 'high-weight neurons' does not follow.","rationale":"The reader's weakest assumption is exactly the unit-of-analysis problem: the paper treats each scalar weight as a 'neuron' and measures importance by zeroing that single weight. I agree this is load-bearing because every experiment in §3.2–3.4 depends on that equation. If a 'neuron' is a filter/channel, then the overlap, perturbation, and retraining results are not about neurons at all; they describe individual connection weights. This invalidates the central claim as stated, independent of the secondary issue that the 45–80% numbers are normalized relative effects rather than measured accuracy drops. The proposed filter-level replication would directly test whether the qualitative finding survives when the object of study is changed from scalar weights to genuine neurons. Since that test is not present in the paper, and the paper itself flags the structural-unit limitation in §4.5, the rejection remains justified. No change to the reader's verdict is needed.","tokens_in":13921,"tokens_out":7711,"duration_ms":82009,"concrete_test":"Train ResNet-18 on CIFAR-10 using the §4.1 protocol. Define neurons as output channels/filters and rank filters by aggregate weight magnitude (e.g., mean absolute weight or L2 norm). Repeat the §3.2 overlap analysis by ablating entire filters (all weights in the filter set to zero) rather than single scalar weights; repeat the §3.3 perturbation analysis by perturbing all weights in a filter and recording raw accuracy changes, not the normalized R(pi|p) index. Compare top-10% overlap and bottom-10% impact at filter level against the scalar-weight results. If low-weight filters show no significant effect and top-weight filters largely overlap important filters, the claimed nonlinearity is an artifact of the scalar-weight analysis; if the pattern persists, the claim survives at the neuron level.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that a CNN neuron be identical to a single scalar weight. The method (§3.1) flattens all trainable weight matrices and sorts absolute weight values; §3.2 then describes 'setting the weights of individual neurons to zero.' But a neuron in a convolutional network is a filter/channel composed of many weights. Zeroing one scalar weight in a 3×3 kernel does not remove a feature detector—it only partially corrupts one connection, and the effect is diluted across spatial positions. Therefore the overlap sets S_w and S_a in Eq. 1, the perturbation effects in Tables 1–2, and the ablation-retraining results in §3.4 all concern individual scalar weights, not neurons. The 'top 10% of weights' is not the top 10% of neurons: a single large filter can contribute many top-10% weights. The paper itself acknowledges this limitation in §4.5, where it defers analysis 'focusing on modules or convolutional layers' to future work. Consequently, even if the tables show a nonlinear relationship, it is not established for neurons. At most, the experiments show something about per-parameter sensitivity, which cannot support the headline claim about neuron importance.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper asks whether high-magnitude weights identify functionally important neurons in image-classification CNNs. The method flattens all trainable weight matrices, sorts scalar weights by absolute value, partitions them into 10% intervals, and conducts three analyses on CIFAR-10 and Mini-ImageNet with LeNet-5, ResNet-18, and ResNet-50: (i) overlap between the top-n weight set and the top-n accuracy-impact set obtained by one-at-a-time ablation; (ii) perturbation of weights in each interval via additive/proportional operations; and (iii) ablation of top-n weights with gradient-masked retraining. The authors report that top-10% weight intervals overlap with accuracy-impacting sets by at most about 25%, that high-weight intervals dominate their normalized perturbation-effect measure while some low-weight intervals show non-negligible effects, and that top-10% ablation cannot be recovered by retraining whereas top-0.1% ablation can. They conclude that weight magnitude alone does not determine neuron importance and that low-weight neurons contribute significantly.","tokens_in":14216,"tokens_out":8239,"duration_ms":82176,"significance":"If the central claim were established at the level of neurons, the paper would be a useful challenge to magnitude-based pruning and would inform interpretability and robustness work. The paper has concrete strengths: a clear experimental protocol, multiple architectures and datasets, and a clean ablation-retraining design with gradient masking. Read as a study of individual scalar weights, the results are a plausible caution that fine-grained magnitude ranking does not align with causal impact. However, the paper's central claim is framed at the neuron level, while the experiments never manipulate a neuron (filter/channel); they manipulate single connection weights. The quantitative claims in the abstract and Section 4.3 also confuse a normalized effect share with an absolute accuracy drop. The first issue is load-bearing and cannot be repaired by editing; it requires rerunning the experiments at unit/filter granularity. The paper's contribution in its current form is therefore substantially narrower than claimed.","major_comments":[{"comment":"Throughout the method, 'neuron' is operationalized as a scalar entry in a flattened weight tensor. In the CNNs used here (LeNet-5, ResNet-18/50), the elementary functional unit is the filter/channel, not an individual connection weight. Setting one element of a 3×3 convolutional kernel to zero does not ablate a neuron; it partially corrupts one connection, and the effect is diluted across spatial positions and filters. Consequently, the overlap sets S_w and S_a in Eq. (1), the perturbation intervals in Tables 1–2, and the ablation-retraining in Algorithm 1 all concern individual scalar weights, not neurons. A filter with a large norm can contribute many top-10% scalar weights, so the top-10% scalar-weight set is not a top-10% neuron set. Section 4.5 explicitly defers module/conv-layer analysis to future work, confirming that the present experiments do not test the paper's stated unit of","section":"§3.1–§3.2, Fig. 1"},{"comment":"Eq. (5) defines E(p_i) = max(p)+ε−p_i and Eq. (6) normalizes by the sum of E over intervals. The numbers in Tables 1–2 therefore are shares of the total perturbation effect; in each row the interval entries plus the Random column sum to approximately 100%. They are not measured accuracy degradations. Yet the abstract states '45–80% accuracy degradation' and §4.3 states the 'Divide' operation 'caused 79.33% accuracy degradation' in ResNet-18 on CIFAR-10. Those sentences confuse a normalized index with an absolute drop. The qualitative finding that the top interval has a disproportionate share may survive renormalization, but the quantitative claims as written are unsupported. Also, the Mini-ImageNet 'minus' row in Table 1 lists 21.37 for three adjacent intervals, which appears to be a data-entry error rather than a real effect.","section":"§3.3, Eqs. (5)–(6), Tables 1–2"},{"comment":"The overlap analysis is the first pillar of the paper's conclusion, but Eq. (1) is not connected to the reported numbers in a reproducible way. The text describes the red line in Fig. 3 as 'the proportion of neurons that are both high-weight and important within the same interval,' while Eq. (1) defines a ratio in terms of differences of cumulative intersections divided by interval width. The symbols S_{a,m}, S_{w,m}, and N_m are not defined precisely enough to determine whether the denominator is interval size, the number of ablated neurons, or the size of the TOP-n sets. The reader cannot reconstruct the overlap percentages from the method as written.","section":"§4.2, Eq. (1)"}],"minor_comments":[{"comment":"The in-text figure references are inconsistent with the captions. §4.2 refers to Figure 2A/2C for overlap results, but Figure 2 is captioned 'Ablation and perturbation Strategies'; §4.3 refers to Figure 3A for perturbation results, but Figure 3 is captioned as overlap results; §4.4 refers to Figure 4A for ablation-retraining, but Figure 4 is the perturbation figure and Figure 5 shows the removal/retraining results. Please renumber or correct the cross-references.","section":"§4.2–§4.4, figure numbering"},{"comment":"With 1-based indexing, 'sorted indices[1:n]' omits the largest-magnitude index. This should be 'sorted indices[0:n]' (or '[:n]') if the intent is to ablate the top-n neurons/weights.","section":"Algorithm 1, line 5"},{"comment":"There are numerous typos and stylistic errors: 'T able 2' in the caption, 'Multipy' for 'Multiply', 'CIF AR-10' with an inserted space, 'Neuroncomputing' for 'Neurocomputing', and inconsistent use of curved quotes/apostrophes. The paper would benefit from a careful proofreading pass.","section":"Throughout"},{"comment":"The text attributes 'DeepMind' to reference [12], but reference [12] is Morcos et al., 'On the importance of single directions for generalization' (arXiv:1803.06959), which is not a DeepMind paper. The attribution should be corrected or removed.","section":"Introduction, reference [12]"},{"comment":"The financial-data 'generalization experiment' is disconnected from the rest of the paper: the OHLC chart dataset is not a natural-image classification task, no training protocol is given for the LeNet model on this dataset, and the appendix is never cited in the abstract, introduction, or conclusion. It appears to be a separate study and should either be integrated with full experimental details or removed.","section":"Appendix A"}],"recommendation":"reject","confidential_remarks":"The rejection is driven by the unit-of-analysis mismatch: the paper equates scalar connection weights with neurons and then draws neuron-level conclusions. This is a foundational issue that cannot be fixed by rewording; the experiments would need to be redone on filters/channels or the paper would need to be reframed as a per-parameter sensitivity study, which would substantially change its scope and contribution. The perturbation metric reporting is also misleading and would need correction in any revision. The manuscript also contains a disconnected financial-data appendix and inconsistent figure references. I see no grounds for a quick major revision; the central claim as stated is not supported by the experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read on 2607.25529. The punchline: the paper wants to say high-weight neurons aren't always important, but it never actually measures neurons. It flattens all trainable weights, sorts them by absolute value, and calls each scalar weight a 'neuron.' In a CNN, a neuron is a filter or channel with many weights. Zeroing one scalar weight in a 3×3 kernel doesn't remove a feature detector; it just corrupts one connection. The paper's own §4.5 admits that module/layer-level analysis is future work. So the central claim about 'high-weight neurons' doesn't follow from the experiments. What does follow is something about per-weight sensitivity, which is a weaker and less novel statement.\n\nSecond issue: the headline numbers. The abstract says '45–80% accuracy degradation' for top-10% weight perturbations. But Tables 1 and 2 report R(pi|p), the normalized effect index defined in Eq. 6, not measured accuracy drops. R is the share of total effect after picking the best-performing interval as baseline. So '79.33%' doesn't mean accuracy fell by 79 points. That's a significant misreport.\n\nWhat the paper does well: the experiment sweep is reasonably broad. Two datasets, three architectures, several perturbation types, and an ablation-retraining protocol that masks top weights and retrains. That protocol is a genuine attempt to test whether the network can compensate. The finding that ablating the top 0.1% allows near-full recovery while the broader top 10% doesn't is the kind of concrete result a follow-up could build on. And the paper is honest enough to include a limitations paragraph, even if the limitations undercut the title.\n\nOther soft spots: no code, no error bars, and the interval boundary choices are arbitrary. The futures-trading appendix looks like another paper pasted in; it adds noise, not evidence. I also wouldn't count this as new. The qualitative point—magnitude doesn't equal importance—was already argued by Morcos et al. and Zhou et al. This paper's contribution is at most a denser parameter scan.\n\nWho should read it: people working on pruning or weight-based importance attribution might skim the retraining section. But in current form I wouldn't rely on the numbers. It deserves a serious referee because the question is testable and the methods could be reframed and fixed, but the authors need to rework the unit of analysis and the metric reporting first.","headline":"The experiments operate on scalar weights, not CNN neurons, so the headline claim about 'high-weight neurons' doesn't follow; and the '45–80% accuracy degradation' numbers are a normalized effect index, not measured accuracy drops.","tokens_in":99,"tokens_out":3299,"would_cite":false,"duration_ms":156383,"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":"Weight magnitude alone does not determine neuron importance in image classifiers.","keywords":["neuron importance","weight pruning","weight-importance relationship","ablation study","perturbation analysis","image classification","CIFAR-10","Mini-ImageNet"],"falsifier":"Take a fixed ResNet-18 trained on CIFAR-10 and compute the overlap between the 10% largest absolute scalar weights and the 10% of weights whose individual zeroing costs the most test accuracy; the paper's claim predicts this overlap stays near 25% and not, say, above 60%. Independently, if one groups weights by convolutional filter and ranks filters by summed absolute magnitude, high-magnitude filters may overlap with accuracy-critical filters much more strongly—which would show the 'nonlinearity' is an artifact of splitting weights inside a filter rather than a property of neurons.","tokens_in":13790,"feed_emoji":"🧠","tokens_out":8711,"duration_ms":86463,"temperature":0.7,"pith_summary":"This paper tests the common assumption that neurons with the largest absolute weights are the most important units in an image classification network. Using three operational probes—overlap between high-weight and accuracy-critical neurons, perturbation sensitivity, and ablation followed by retraining—on CIFAR-10 and Mini-ImageNet, it finds that top-10% weight neurons overlap with truly impactful neurons only about a quarter of the time. Some top-decile weights are nearly inert, while some low-weight neurons degrade accuracy as much as mid-range ones when disturbed. The authors conclude that the weight–importance relationship is nonlinear and that applications such as pruning, encryption, and backdoor defense should not treat high magnitude as synonymous with high importance.","feed_headline":"In image classifiers, big weights don't make important neurons","feed_subtitle":"Three probes on standard image benchmarks find ~25% overlap between top-weight and accuracy-critical neurons.","key_machinery":"Central machinery: the sorted absolute-weight interval decomposition. Every scalar weight in the trained network is flattened into a vector, absolute-valued, sorted descending, and split into non-overlapping intervals (deciles, with the top decile subdivided into finer bands). Three operations are then applied interval-by-interval: single-weight ablation for overlap, additive/multiplicative perturbation for robustness, and masked zeroing with retraining for recoverability. Together these translate 'importance' from a magnitude claim into a behavioural claim—what actually happens to accuracy when the unit is removed, disturbed, or blocked from learning.","core_discovery":"On the paper's own terms, 'neuron' means a single scalar weight: after training, the model's weights are flattened, absolute-valued, sorted descending, and cut into 10% intervals. Three probes are then run. Per-weight ablation finds which weights, individually removed, cost the most test accuracy; overlap between that set and the largest-weight set peaks around 25% in the top decile and is worse elsewhere. Perturbation tests (adding/subtracting a constant or multiplying/dividing by a factor) show the top 10% interval is far more fragile than random selection—45–80% accuracy loss versus 3–7%—but roughly a third of top-decile weights barely matter, and some low-weight intervals lose 10–17% und","pith_inferences":["Beyond the paper: because each 'neuron' is a scalar weight rather than a convolutional filter, the result is best read as a statement about weight-level sensitivity. Repeating the three probes at filter or channel granularity would show how much of the nonlinearity survives when the unit is a real neuron.","Beyond the paper: the perturbation pattern suggests a testable 'criticality spectrum'—extreme high weights are individually fragile, mid-range weights are redundant, and very low weights matter only in groups. A direct next experiment is to ablate random groups of low-weight neurons of growing size and check whether their collective contribution is linear or sharply thresholded.","Beyond the paper: the results imply a cheap behavioural importance score—marginal accuracy loss from one-shot ablation or small perturbation—that could replace magnitude ranking in pruning and be benchmarked against magnitude pruning at matched sparsity.","Beyond the paper: the low-weight 'ecosystem' idea implies interpretability and mechanistic analyses should not focus only on high-activation or high-magnitude units; distributed low-weight codes may matter for generalization."],"forward_implications":["A pruning rule that removes the smallest absolute weights may discard functionally significant units, because some low-weight intervals affect accuracy as much as mid-range high-weight ones.","High weight alone does not justify cryptographic or defensive priority: a substantial subset of the top decile shows minimal accuracy impact under perturbation, so protecting all high-weight scalars would waste effort.","The top decile is not a single class: removing all of it prevents recovery after retraining, but finer subdivision shows many subintervals are recoverable, so irreplaceability is concentrated in a small extreme subset of weights.","Applications that rely on weight-importance equivalence—pruning, backdoor defense, interpretability—should incorporate behavioural measures rather than magnitude alone.","The same qualitative pattern appeared in a financial OHLC image generalization experiment, so the phenomenon is not unique to natural-image benchmarks."],"fun_headline_variants":["High-weight neurons aren't the key players in image models","Weight magnitude doesn't equal neuron importance in CNNs","Big weights don't equal big impact in neural nets","Top-weight neurons: only 25% overlap with critical ones","Not all heavy hitters matter: neuron importance is nonlinear"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that a single scalar weight in a flattened, absolute-sorted vector counts as a 'neuron'; if the functional unit is a filter or channel made of many weights, the measured nonlinearity applies to scalar weights, not to neurons as usually understood.","fun_headline_variants_meta":{"raw":{"variants":["High-weight neurons aren't the key players in image models","Weight magnitude doesn't equal neuron importance in CNNs","Big weights don't equal big impact in neural nets","Top-weight neurons: only 25% overlap with critical ones","Not all heavy hitters matter: neuron importance is nonlinear"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000232,"raw_usage":{"total_tokens":1357,"prompt_tokens":806,"completion_tokens":551,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":471}},"tokens_in":550,"tokens_out":551,"duration_ms":5940,"temperature":1.0,"reasoning_tokens":471,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T02:09:18.262810+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a fixed ResNet-18 trained on CIFAR-10 and compute the overlap between the 10% largest absolute scalar weights and the 10% of weights whose individual zeroing costs the most test accuracy; the paper's claim predicts this overlap stays near 25% and not, say, above 60%. Independently, if one groups weights by convolutional filter and ranks filters by summed absolute magnitude, high-magnitude filters may overlap with accuracy-critical filters much more strongly—which would show the 'nonlinearity' is an artifact of splitting weights inside a filter rather than a property of neurons.","supporting_citations":[],"review_version":1}