{"id":"5df85300-8bf9-4c24-bf12-4259ef8396d9","arxiv_id":"2501.15592","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A flow-based stopping criterion (InCoP) reduces the training epochs needed during iterative magnitude pruning while preserving final accuracy, compared with LTH and SAP.","lead":"This paper introduces a new stopping rule for iterative neural-network pruning that watches how information and gradients flow through the network, instead of waiting for the pruned network to recover its original accuracy. The authors report that their rule reaches the same final sparsity and accuracy as lottery-ticket methods while training for far fewer epochs, though the supporting mathematics has gaps.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The flow-accuracy correspondence is unsupported: the supplementary proof is logically inverted (lower bounds only) and the experiments lack an accuracy-based early-stopping baseline, so the efficiency gain may not be attributable to the flow criterion.","rationale":"The reader's weakest assumption is precisely the flow-accuracy correspondence, and the paper's own proof does not establish it; the inequalities in SM 6.1 are all lower bounds, so the implication 'small flow distance ⇒ small accuracy gap' never follows. This is the load-bearing step because Algorithm 1's stopping criterion is exactly 'stop when the flow distance is below ε', and the entire efficiency claim rests on that criterion not degrading final accuracy. The empirical section provides some evidence that InCoP matches SAP's final accuracy with fewer epochs, but without a matched-budget accuracy-based early-stopping baseline we cannot tell whether the flow metric is the cause of the savings or merely correlated with training time. The missing baseline is the most direct way to test the central claim, and adding it is a feasible revision. The reader's CONDITIONAL verdict and the specific concern about SM Eq. (8)–(13) align with this analysis, so no change to the verdict is needed; the recommendation is to require the authors to either repair the proof (establishing an upper bound) or supply the baseline experiment, ideally both.","tokens_in":13214,"tokens_out":5737,"duration_ms":47190,"concrete_test":"Run the same SAP pruning schedule and hyperparameters on ResNet18/FashionMNIST and VGG16/MNIST, but replace the flow-based stopping criterion with a validation-accuracy-based early stopping rule: after each retraining epoch, stop if validation accuracy exceeds (Acc∗ − δ) for a small δ, or if validation accuracy has not improved by a threshold over the last few epochs, using the same maximum epoch budget (e.g., 20). Compare final test accuracy and total training epochs across InCoP-IF, InCoP-GF, and this accuracy-based early stopping. If the accuracy-based rule attains equal or better final accuracy with no more epochs, the flow criterion is not necessary for the efficiency benefit; if InCoP clearly improves on it, the flow-accuracy correspondence gains empirical support.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that keeping the layer-wise information (or gradient) flow close to that of the fully trained dense network, within epsilon, is sufficient to preserve final test accuracy, enabling early stopping in iterative magnitude pruning. The theoretical argument in Supplementary Section 6.1 does not establish this. Equation (8) gives a lower bound on the accuracy gap in terms of weight distance: |Acce−Acc∗| ≥ (a/2) E[Y λmin ||we−w∗||²]. Equation (12) gives a lower bound on weight distance in terms of flow gap: E[Cσ ||we−w∗||²] ≥ ε − L Cσ. Combining these (and the unproven inequality (13)) yields |Acce−Acc∗| ≥ (a/2)(ε − L Cσ), which is still a lower bound. It shows that a large flow gap forces a large accuracy gap, but it does not show that a small flow gap forces a small accuracy gap; the desired implication requires an upper bound such as |Acce−Acc∗| ≤ C · flow_gap. The proof is therefore inverted relative to the claim. Moreover, the experiments compare InCoP only against SAP and LTH that train a fixed number of epochs; there is no control where the same early-stopping schedule is driven by validation accuracy. Thus the reported epoch savings could be achieved by any sensible early-stopping rule, and the specific role of the flow metrics in preserving accuracy is not isolated. Without a validated flow-accuracy correspondence or this baseline, the efficiency improvement is not convincingly attributable to information/gradient flow.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes InCoP, an early-stopping criterion for iterative magnitude pruning (IMP). After fully training a dense reference network, InCoP stores its layer-wise information flow (correlation between consecutive-layer activations) and gradient flow; during each pruning iteration it retrains only until the corresponding flow distance to the reference is below a threshold epsilon, then uses SAP's adaptive pruning ratio to remove weights. Experiments on MNIST, Fashion-MNIST, and CIFAR-10 with VGG16, ResNet18, and ResNet50 report the same final accuracy as SAP, while using substantially fewer training epochs. A supplementary section (Section 6.1) attempts to prove that small flow distance implies small accuracy gap.","tokens_in":13605,"tokens_out":8052,"duration_ms":73854,"significance":"If the empirical claims hold, InCoP would be a practically useful technique that reduces the dominant training cost of IMP while preserving final sparsity and accuracy. The paper has several strengths: the method is simple, it is evaluated across several dataset/architecture combinations, the code is publicly available, and the authors report wall-clock comparisons rather than only epoch counts. However, the theoretical support in Section 6.1 of the supplementary material is not valid as written, and the experiments do not isolate the flow criterion from generic early stopping, so the central claim is not yet convincingly established.","major_comments":[{"comment":"The proof of the Analytical Hypothesis is logically inverted. Eq. (6) applies a reverse Jensen bound to obtain a lower bound on |Acc_e - Acc*| without justifying that the required convexity and positivity conditions hold for the expression Y * |l(w_e) - l(w*)|. Eq. (8) treats lambda_min as if it were the smallest eigenvalue while calling it the maximum eigenvalue, and the second-order Taylor argument requires a positive-definite Hessian. More importantly, Eqs. (8) and (12) are both lower bounds on the accuracy gap and on the weight distance, respectively, and Eq. (13) is asserted without derivation and with an inequality direction that is not justified: taking eC >= C_sigma makes the left-hand side larger than E[C_sigma ||w_e - w*||^2], so it cannot provide the stated upper connection. The final conclusion |Acc_e - Acc*| >= epsilon - L*C_sigma is only a lower bound and becomes trivial as epsilon -> 0; it cannot establish the claimed implication 'small flow gap implies small accuracy gap', which would require an upper bound such as |Acc_e - Acc*| <= C * flow_gap. The theoretical claim in the abstract and Section 2.3 is therefore unsupported by the presented mathematics.","section":"Supplementary Material, Section 6.1, Eqs. (5)-(13)"},{"comment":"The experimental section compares InCoP only with SAP and LTH, both of which train a fixed number of epochs per iteration. There is no baseline in which the same per-iteration training budget is terminated by a conventional accuracy-based or validation-loss-based early-stopping rule. Without such a control, the observed epoch savings cannot be attributed specifically to the information/gradient-flow criterion, because any sensible validation-based early-stopping heuristic might yield similar savings. Please add such a baseline and report the resulting accuracy and training time, so that the role of the flow metrics is isolated.","section":"Section 3.1 and 3.2, Figs. 3-5"},{"comment":"The pseudocode resets the remaining weights to w_t, the weights at the beginning of the iteration, rather than to the trained weights w_e: 'Reset the remaining weights to w_t by w_{t+1} = m_t ⊙ w_t'. If this is taken literally, the retraining epochs do not update the weights that are carried into the next iteration, which is inconsistent with the described IMP procedure and with the accuracy results reported in Figs. 3 and 4. Please either correct the pseudocode to w_{t+1} = m_t ⊙ w_e, or explain the intended rewind step and why it still produces the reported accuracies.","section":"Algorithm 1, lines 12-13"},{"comment":"The main experiments rely on the threshold epsilon as the stopping tolerance, but the paper does not report how epsilon was selected for the reported runs; the ablation study in Supplementary Section 6.2 only shows qualitative behavior for various values. Please specify the chosen epsilon values for each dataset/architecture combination and the selection rule (for example, using a validation set). This is necessary for reproducibility because epsilon directly controls the trade-off between epoch savings and final accuracy.","section":"Section 3 and Supplementary Section 6.2"}],"minor_comments":[{"comment":"The text refers to 'Table 5' twice when describing the execution-time comparison, but the comparison is presented in Figure 5.","section":"Section 3.2"},{"comment":"The hyperparameter study refers to the two variants as 'CIAP' and 'GIAP'; these should be InCoP-IF and InCoP-GF for consistency with the main text.","section":"Supplementary Section 6.2"},{"comment":"The pseudocode does not explicitly define how w_t is initialized for t = 1; the reader must infer that w_0 equals the trained reference w*.","section":"Algorithm 1"},{"comment":"Equation (8) is missing an absolute value or a sign condition on Y * lambda_min; as written the right-hand side can be negative, making the lower bound vacuous or invalid.","section":"Supplementary Material, Eq. (8)"},{"comment":"The complexity expression O(T * E * N * L * M_l * M_{l+1}) should be written as a sum over layers; with the displayed notation it appears to depend on a layer index l outside the O, which is not a well-formed single bound.","section":"Section 3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is not ready for publication in its current form. The core empirical idea is interesting and the code availability helps, but the theoretical section in SM 6.1 must either be repaired or explicitly reduced to a heuristic argument, and the experiments need an accuracy-based early-stopping control. I recommend major revision rather than rejection because the algorithmic idea is testable and the empirical trend, if reproducible under the added baseline, would be a useful contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The quick take: InCoP is a genuinely new stopping criterion for iterative magnitude pruning—watch the distance between the pruned network's layer-wise information or gradient flow and that of the fully trained dense network, stop when it's within epsilon. The reported epoch savings are large and the accuracy is preserved. That is worth paying attention to.\n\nWhat's solid: the experimental design has the right comparison for pruning schedule—the remaining weight curves are identical to SAP, so the only difference is the stopping rule. The code is public. The epsilon ablation is useful and shows the expected trade-off. If the empirical claim reproduces, this could save real compute in IMP pipelines.\n\nThe soft spots are real, though. The supplementary proof is upside down. Eq (6) uses a reverse Jensen bound that gives a lower bound on the accuracy gap; Eq (12) gives a lower bound on weight distance; combining them still gives a lower bound. That shows a large flow gap forces a large accuracy gap, but it does not show that a small flow gap keeps the accuracy close. The premise of early stopping needs an upper bound on accuracy gap from the flow gap, and the paper never provides one. The asserted constant in Eq (13) is just pulled out of the air. So the theory should be rewritten as a heuristic justification or given a real argument.\n\nMore damaging is the missing baseline. The paper compares against SAP and LTH with fixed epoch budgets. There is no experiment where the same early-stopping schedule is driven by validation accuracy. Until that control exists, we cannot attribute the saving to the flow metrics. You could get the same curve by stopping on any reasonable proxy that correlates with accuracy.\n\nMinor things: the supplement says ResNet18 but the figure is labeled ResNet50; the ablation text refers to 'CIAP and GIAP' instead of InCoP-IF/GF; the time-complexity claim treats M_l as constant, but it shrinks during pruning.\n\nNet: this deserves referee time. The idea is novel, the evidence is suggestive but incomplete, and the theory is currently misleading. I'd recommend sending it out with the clear expectation that the authors either repair or drop the proof, add an accuracy-based early-stopping control, and clean up the details. A reader could reproduce the core experiment and check for themselves. That is more than many papers offer.","headline":"The flow-based stopping criterion for IMP is a novel idea that could save real compute, but the current proof only gives lower bounds and the experiments lack the accuracy-based early-stopping control needed to isolate the flow metric's effect.","tokens_in":14113,"tokens_out":3730,"would_cite":true,"duration_ms":34043,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Flow proximity, not full retraining, preserves pruned-network accuracy","keywords":["iterative magnitude pruning","lottery ticket hypothesis","information flow","gradient flow","stopping criterion","sparse neural networks","training efficiency","network pruning"],"falsifier":"A concrete falsifier: run InCoP on MNIST with ResNet18 and a small $\\epsilon$, log $\\|\\Phi_e - \\Phi^*\\|$ and test accuracy at the end of every retraining epoch, and check whether training ever stops while the accuracy gap to the dense network remains larger than the gap the accuracy-based stopping rule would allow; if such an epoch exists, the flow-accuracy premise fails.","tokens_in":13047,"feed_emoji":"⚡","tokens_out":12220,"duration_ms":98498,"temperature":0.7,"pith_summary":"Information Consistent Pruning (InCoP) claims that iterative magnitude pruning can stop retraining long before a network recovers its original accuracy, as long as the network's layer-wise information flow or gradient flow has returned within a small tolerance of the fully trained dense network. The paper argues that accuracy-based stopping is too coarse because it only inspects the final output layer, while flow-based stopping tracks how each layer relates to its neighbors or to the gradient. Across Fashion-MNIST, MNIST, and CIFAR-10 with several ResNet and VGG architectures, InCoP reaches the same final sparsity and accuracy as the SAP baseline while training for far fewer epochs; for InCoP-IF with $(p,q)=(0.5,1)$ on one ResNet18 setting, the paper reports 13% of SAP's execution time. If the flow-accuracy correspondence holds, this turns the most time-consuming part of lottery-ticket pruning into a cheap early-stopping check.","feed_headline":"Flow check cuts pruning retraining to 13% of run time","feed_subtitle":"InCoP stops retraining once layer-wise information or gradient flow matches the fully trained dense network.","key_machinery":"The load-bearing object is the flow vector $\\Phi$, either the information-flow connectivity matrix $\\Delta_w(f^{(l)}, f^{(l+1)}) = \\mathbb{E}_{(X,Y)\\sim D}[f_i^{(l)} f_j^{(l+1)}]$ between consecutive layers, or the gradient flow $g_i = \\partial \\ell / \\partial w^{(i)} \\odot m^{(i)}$ per layer. The stopping rule compares these to the corresponding quantities $\\Phi^*$ of the fully trained dense network and halts training at epoch $e$ once $\\|\\Phi_e - \\Phi^*\\| \\le \\epsilon$. This turns the stopping criterion from a global accuracy comparison into a layer-wise proximity check, and it is what lets Algorithm 1 cut training epochs while keeping the same sparsity schedule as SAP.","core_discovery":"The paper's discovery claim is that the expensive retrain-to-full-accuracy step in iterative magnitude pruning is unnecessary: a pruned network can be considered ready for the next pruning round once its information flow, the expected product of activations in consecutive layers, or its gradient flow, the masked gradients per layer, is within $\\epsilon$ of the corresponding flow of the fully trained dense network. InCoP implements this by computing $\\Phi^*$ once on the optimal dense network and then stopping each retraining phase at the first epoch where $\\|\\Phi_e - \\Phi^*\\| \\le \\epsilon$. The authors report that this reaches the same final accuracy as SAP and LTH on Fashion-MNIST, MNIST, and CIFAR-10 with VGG16, ResNet18, and ResNet50 while using substantially fewer epochs, and in one configuration uses 13% of SAP's execution time. The supplementary material argues, using a reverse Jensen inequality and Lipschitz continuity of activations, that small connectivity distance implies small accuracy distance.","pith_inferences":["An implication the paper leaves implicit is that the same flow-matching stopping rule could be attached to any iterative pruning schedule with a reference dense network, including continual-learning and transfer settings.","The supplementary proof bounds the accuracy gap from below, not above, so a rigorous guarantee for the stopping rule would need a separate argument showing small flow distance forces high accuracy; until then the method rests on the empirical flow-accuracy correspondence.","The reported 13% runtime figure is for one configuration; on very wide or deep networks the per-epoch cost of computing connectivity could outweigh the epoch savings, so the speedup has to be re-measured per architecture.","A testable extension is whether one $\\epsilon$ value transfers across datasets and architectures or must be retuned, since the paper's ablation shows the tradeoff shifts with $\\epsilon$."],"forward_implications":["The retraining phase of an iterative magnitude pruning loop, usually a fixed number of epochs per iteration, can be replaced by an adaptive number of epochs determined by flow proximity to the dense reference.","Because InCoP inherits SAP's pruning rates, the final sparsity for a given $(p,q)$ is the same as SAP; only the stopping point changes.","The information-flow and gradient-flow stopping rules are independent of the pruning method, so they could be attached to other magnitude- or sensitivity-based pruning procedures.","The choice of $\\epsilon$ controls the tradeoff: larger values stop earlier and behave more conservatively, while very small values risk stopping before the network has recovered enough for the next pruning step.","InCoP-IF is more sensitive to aggressive pruning because connectivity between consecutive layers drops on the order of $M_l^2$ as neurons are removed, while InCoP-GF stays stable across pruning ratios."],"supporting_citations":[{"why":"Defines the lottery-ticket iterative pruning-and-retraining loop that InCoP inherits and shortens.","marker":"[9]"},{"why":"Supplies the Sparsity-Informed Adaptive Pruning schedule and PQ-Index that InCoP uses for the pruning step.","marker":"[6]"},{"why":"Introduces effective gradient flow as a proxy for sparse-network performance, the basis of InCoP-GF.","marker":"[36]"},{"why":"Shows that preserving gradient flow can identify winning tickets before training, motivating the flow-based stopping view.","marker":"[37]"},{"why":"Establishes that pruning degrades gradient flow and that retraining restores it, the process InCoP monitors.","marker":"[8]"},{"why":"Provides the connectivity matrix between consecutive layers that InCoP uses to define information flow.","marker":"[12]"},{"why":"Supplies the inverse Jensen inequality used in the paper's supplementary proof connecting flow distance to accuracy distance.","marker":"[39]"}],"fun_headline_variants":["Flow match stops pruning retrain early, keeping final accuracy","InCoP: halt retraining when flows align, save 87% time","Prune with flow consistency: 13% runtime, same accuracy","Retrain only until flows match, then prune again efficiently","Stop retraining at flow proximity for faster network pruning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a network whose layer-wise information or gradient flow is close to the optimal dense network's flow will also have test accuracy close to the optimal dense network's accuracy; if flow closeness and accuracy closeness can diverge, stopping early on flow alone may leave the pruned network underperforming.","fun_headline_variants_meta":{"raw":{"variants":["Flow match stops pruning retrain early, keeping final accuracy","InCoP: halt retraining when flows align, save 87% time","Prune with flow consistency: 13% runtime, same accuracy","Retrain only until flows match, then prune again efficiently","Stop retraining at flow proximity for faster network pruning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000461,"raw_usage":{"total_tokens":2304,"prompt_tokens":938,"completion_tokens":1366,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":1279}},"tokens_in":554,"tokens_out":1366,"duration_ms":11945,"temperature":1.0,"reasoning_tokens":1279,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:09:15.632339+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete falsifier: run InCoP on MNIST with ResNet18 and a small $\\epsilon$, log $\\|\\Phi_e - \\Phi^*\\|$ and test accuracy at the end of every retraining epoch, and check whether training ever stops while the accuracy gap to the dense network remains larger than the gap the accuracy-based stopping rule would allow; if such an epoch exists, the flow-accuracy premise fails.","supporting_citations":[{"cited_title":"Slimming Neural Networks using Adaptive Connectivity Scores","cited_arxiv_id":"2006.12463","evidence_quote":"Provides the connectivity matrix between consecutive layers that InCoP uses to define information flow."},{"cited_title":"In: 2021 IEEE Information Theory Workshop (ITW)","cited_arxiv_id":null,"evidence_quote":"Supplies the inverse Jensen inequality used in the paper's supplementary proof connecting flow distance to accuracy distance."},{"cited_title":"CoRR (2023),http://arxiv.org/abs/2302","cited_arxiv_id":null,"evidence_quote":"Supplies the Sparsity-Informed Adaptive Pruning schedule and PQ-Index that InCoP uses for the pruning step."},{"cited_title":"Keep the Gradients Flowing: Using Gradient Flow to Study Sparse Network Optimization","cited_arxiv_id":"2102.01670","evidence_quote":"Introduces effective gradient flow as a proxy for sparse-network performance, the basis of InCoP-GF."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes that pruning degrades gradient flow and that retraining restores it, the process InCoP monitors."}],"review_version":1}