{"id":"900e1131-4f43-4407-a11b-7a049665fafa","arxiv_id":"2505.11569","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A master's thesis that prunes CNNs into smaller subnetworks and then rebuilds them by reinserting pruned filters, but it does not test the claimed runtime adaptivity.","lead":"A prune-and-rebuild method embeds smaller versions of a CNN inside the full model, aiming for runtime size adjustment. A generalist should read it because the gap between the abstract's promise and the experiments tests how honestly elastic-network claims are validated.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The core subnetwork is never extracted or tested after rebuilding; freezing weights does not freeze BatchNorm running statistics, so runtime switching without retraining is unsupported.","rationale":"The reader's REJECT verdict is well supported. My stress-test identifies the same load-bearing weakness but sharpens it: preserving core weights is necessary but not sufficient for a switchable nested subnetwork. BatchNorm running statistics and downstream shared weights are part of the model's function; freezing convolutional weights alone does not freeze them. The manuscript's own limitation statements (§5.1, §6.3) concede that dynamic runtime adaptation was not tested. This is not an external-consensus disagreement; it is an internal mismatch between the paper's central claim and its evidence. The one-shot and iterative pruning results may be useful empirical observations about prune/rebuild pipelines, but they do not establish an elastic model. No code or reproducible artifact is provided, and there are internal inconsistencies (Imagenette in the abstract but not in the body; duplicated Table 5.4/5.7 captions; manual pruning issues in ResNet-20), which further reduce confidence. A single additional experiment—extracting the core from the rebuilt model and comparing its accuracy and BN statistics before/after—would decisively test whether the central claim survives; until then, rejection remains appropriate.","tokens_in":21725,"tokens_out":4156,"duration_ms":44816,"concrete_test":"On ResNet-20 with 50% one-shot pruning, save the fine-tuned pruned core including per-channel BatchNorm running statistics. Rebuild and fine-tune per §3.3–3.4. Then run the rebuilt model in 'compact mode' by masking the reinserted channels (or extracting the core via the stored pruning indices). Compare (a) compact-mode test accuracy against the saved pruned core's accuracy; (b) per-channel running_mean/running_var of core channels before and after rebuild; (c) maximum absolute feature-map or logit difference on a fixed test batch. Also run two consecutive forward passes alternating between compact and full mode to confirm no state contamination. If compact-mode accuracy drops by more than 1–2 points or BN statistics diverge, the claim of switching without retraining is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract and §1.0.2 claim a single CNN contains nested subnetworks switchable at runtime without retraining. For that claim, the rebuilt model must contain the fine-tuned pruned core as a usable subnetwork. The paper reports accuracy of the standalone pruned model before rebuilding and of the full rebuilt model after fine-tuning, but never runs the core inside the rebuilt model, never masks the reinserted channels, and never compares the embedded core's outputs. §5.1 explicitly says adaptivity is not tested dynamically and §6.3 says dynamic runtime adaptation is beyond scope, so the headline claim is not supported by the experiments. The freeze step in §3.3 and §3.4 only zeroes gradients of retained weights; BatchNorm running mean and variance are updated during training-mode fine-tuning unless explicitly frozen. If the core channels' BN statistics move, the compact configuration embedded in the rebuilt model no longer computes the same function as the reported pruned core, even with identical convolutional weights. No evidence is given that this drift is absent, and no extraction test is reported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript, a master's thesis posted on arXiv, proposes a prune-and-grow approach for CNNs. It applies dependency-aware structured pruning (via DepGraph/Torch-Pruning) to obtain a compact 'core' model, then rebuilds the full architecture by reinserting the previously pruned filters at their original positions with their original weights. During post-rebuild fine-tuning, the retained core channels are frozen (gradients zeroed) while the reinserted channels are updated. The paper claims this procedure creates nested subnetworks within a single CNN, allowing runtime switching between compact and full-sized configurations without retraining. Experiments are reported on CIFAR-10 with VGG-16, AlexNet, ResNet-20, and ResNet-56, comparing one-shot and iterative pruning across magnitude, Taylor, Hessian, manual, and soft pruning criteria, plus an alternate-layer pruning ablation. The paper's own evaluation and limitations sections state that dynamic runtime adaptivity was not tested and is beyond the scope of the work.","tokens_in":21935,"tokens_out":4671,"duration_ms":45668,"significance":"If the central claim were established, the paper would offer a practical recipe for embedding multiple capacity levels in a single CNN, which is a relevant deployment problem. The work has some strengths: it systematically compares several importance criteria under a common dependency-aware pruning framework, reports parameter and size reductions, and includes an interesting ablation on alternate-layer pruning. The use of DepGraph is appropriate and makes the pruning structurally sound. However, the distinctive claim of the paper—runtime switching without retraining—is not tested in any experiment and is explicitly declared out of scope in §5.1 and §6.3. The reported experiments are a pruning-and-recovery study: they show that reinserting known pretrained weights and fine-tuning restores accuracy, which is largely expected. The empirical base is further weakened by mislabeled tables and a mismatch between the abstract's claimed datasets and the experiments actually described. As a result, the paper's significance for the runtime-adaptivity claim is low, and the evidence does not support the headline contribution.","major_comments":[{"comment":"The central claim that the network can 'dynamically switch between compact and full-sized configurations without retraining' is not supported by the reported experiments. §5.1 states that 'the evaluation does not explicitly test adaptivity in dynamic runtime environments,' and §6.3 states that dynamic runtime adaptation is 'beyond the scope of this work.' No experiment extracts or masks the core subnetwork from the rebuilt model, measures its accuracy, or measures switching overhead. The paper's own limitations therefore contradict the abstract's headline claim.","section":"Abstract; §1.0.2; §5.1; §6.3"},{"comment":"The claim that freezing retained channels preserves the embedded core subnetwork is not substantiated. Freezing as described only zeroes the gradients of the retained weights during backpropagation; it does not freeze BatchNorm running statistics or other shared state unless explicitly done. After fine-tuning the rebuilt model in training mode, the BatchNorm statistics of core channels can drift, so the embedded core may no longer compute the same function as the reported pruned core even if convolutional weights are unchanged. No extraction test or output-comparison test is reported to rule out this drift, and this evidence is load-bearing for the runtime-switching claim.","section":"§3.3, step 4; §3.4"},{"comment":"Table 5.4 is labeled 'Results from Iterative Pruning and Rebuilding on ResNet-20' but its baseline row (86.50% accuracy, 0.86M parameters, 3.37MB) matches the ResNet-56 baseline in Table 5.1, and the text immediately above the table says 'Table 5.4 shows the changes ... for ResNet-56.' Table 5.7 repeats the ResNet-20 label but reports a different baseline (91.5%, 0.27M, 1.08MB). This internal inconsistency makes the iterative results impossible to interpret and suggests the labels or numbers were swapped.","section":"Table 5.4; Table 5.7"},{"comment":"The abstract claims experiments on both CIFAR-10 and Imagenette datasets, but §4.1 states that 'all experiments in this study are conducted on the CIFAR-10 dataset,' and no Imagenette results appear anywhere in Section 5. This is not a mere typo: it overstates the empirical scope of the paper and should be corrected.","section":"Abstract; §4.1; §5"},{"comment":"All accuracy claims are based on single runs with no error bars, confidence intervals, or seeds. For example, in Table 5.2, the magnitude-pruned VGG-16 at 30% reports pruned accuracy 88.56 and rebuilt accuracy 88.80, a difference of 0.24 percentage points; such small differences cannot be interpreted without variance information. The repeated claims that 'rebuilt models consistently achieved higher accuracy than their corresponding pruned counterparts' (§5.4) are therefore not statistically supported.","section":"§5.3–§5.5; Tables 5.2, 5.4, 5.5, 5.7"}],"minor_comments":[{"comment":"The 'This thesis is organized as follows' paragraph says Chapter 4 presents experimental results and Chapter 5 evaluates the model, but in the actual manuscript Chapter 4 is the experimental setup and Chapter 5 contains the results; the organization description should be corrected.","section":"§1.0.2"},{"comment":"Section 4.2 says 'we experimented with three popular CNNs (VGG-16, AlexNet, ResNet-20 and ResNet-56)' while listing four architectures, and §6.3 says experiments were conducted only on three architectures; the count should be made consistent.","section":"§4.2; §6.3"},{"comment":"The Taylor / ResNet-20 / 30% rebuilt cell contains a trailing comma ('82.89,') that appears to be a typographical artifact.","section":"Table 5.2"},{"comment":"Figures 3.1 and 3.5 appear to be the same overview figure with different captions, and Figure 3.1 carries a source citation [1] while Figure 3.5 does not; the duplication and attribution should be clarified.","section":"Figures 3.1 and 3.5"},{"comment":"The text says Table 5.3 'presents the numbers from Manual Pruning,' but the table caption and surrounding discussion refer to pruning results generically; the method used for the numbers in Table 5.3 should be stated explicitly.","section":"§5.3.2, Table 5.3"}],"recommendation":"reject","confidential_remarks":"This appears to be a master's thesis submitted as an arXiv paper. The central claim—runtime switching without retraining—is explicitly untested in the manuscript and is declared out of scope in the limitations section. The table mislabeling (Table 5.4 labeled ResNet-20 but containing ResNet-56 numbers) and the abstract's Imagenette claim are additional signs that the manuscript has not been carefully checked before submission. Even a major revision would require substantial new experiments (e.g., extracting and evaluating the embedded core, checking BatchNorm drift, adding variance estimates) that go beyond the current scope. If the authors are willing to reposition the paper as a pruning-and-recovery study and remove the runtime-adaptivity claims, a future submission might be viable, but in its present form the paper does not support its advertised contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nMy take on arXiv:2505.11569: the headline claim—a single CNN can switch between compact and full configurations at runtime without retraining—is not supported by the experiments. The paper never tests runtime switching; the evaluation section says so explicitly, and §6.3 lists dynamic runtime adaptation as out of scope. The rebuilt model is never decomposed back into the core subnetwork and checked for functional equivalence.\n\nWhat the paper does well is a fairly systematic comparison of structured pruning criteria (magnitude, Taylor, Hessian, soft, manual) on VGG-16, AlexNet, ResNet-20/56, with an iterative prune-and-rebuild loop. Rebuilding with cached weights does recover most of the lost accuracy, and the ablation on pruning only alternate layers is a reasonable experiment. The authors are honest about many limitations, which is more than some theses manage.\n\nThe soft spots are serious and mostly about the gap between abstract and evidence. The two \"adaptivity\" chapters are about rebuilding offline, not runtime elasticity. The stress-test point about BatchNorm is correct: freezing the gradients of retained convolutional weights does not freeze BatchNorm running statistics, which update during fine-tuning in train mode. Since the core is never extracted and evaluated inside the rebuilt model, the claim that the core is \"preserved\" is unverified. Table 5.4 is mislabeled (caption says ResNet-20, numbers are ResNet-56). The abstract promises Imagenette; the body only uses CIFAR-10. No error bars, no code, and some numbers look inconsistent between tables (e.g., ResNet-20 baseline is 92.60 in Table 5.1 and 91.5 in Table 5.2/5.7).\n\nThe novelty is modest. NestDNN already does freeze-and-grow nested subnets, and the paper cites it as the adopted strategy. The contribution here is essentially an implementation using DepGraph metadata plus a few ablations. That is useful as an engineering exercise, not as a new result.\n\nWho is this for? Someone working through pruning toolchains may find the methodology notes useful, but I would not send it to a serious venue as-is. If it were submitted as a paper, my recommendation is desk reject: the central claim is untested and the paper contradicts itself on the scope.","headline":"The runtime-adaptivity claim is unsupported: the paper measures offline pruning and rebuilding, never tests switching, and never verifies the embedded core.","tokens_in":22450,"tokens_out":2154,"would_cite":false,"duration_ms":21778,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A CNN can be pruned into a compact core and then grown back, so one model holds nested sizes switchable without retraining.","keywords":["model elasticity","structured pruning","dependency graphs","network rebuilding","nested subnetworks","adaptive CNNs","runtime adaptivity","CIFAR-10"],"falsifier":"Take a rebuilt model and run inference using only its core filters, then compare accuracy against the same core model fine-tuned in isolation. If the two accuracies differ substantially, or if the internal normalization statistics of the core have shifted during post-rebuild fine-tuning, the claim that switching between sizes requires no retraining would be refuted.","tokens_in":21530,"feed_emoji":"🔀","tokens_out":9117,"duration_ms":87183,"temperature":0.7,"pith_summary":"This paper tries to establish that a single convolutional neural network can be made elastic: it can contain a compact version of itself nested inside the full-size network and switch between the two at runtime without retraining. The proposed route is structured pruning, which removes entire filters to form a small core model, followed by rebuilding the original architecture by reinserting the removed filters with their original weights and fine-tuning only the new parts. The payoff, if the claim holds, is one deployed model that can serve both low-resource edge devices and high-capacity servers without storing or paging multiple separate models. Experiments on VGG-16, AlexNet, ResNet-20, and ResNet-56 over CIFAR-10 show rebuilt models recovering most or all of the original accuracy, with iterative pruning giving smoother recovery than one-shot pruning.","feed_headline":"Prune-and-grow packs a smaller twin model into one CNN","feed_subtitle":"Nested subnetworks let one deployed CNN handle both low-power and full-power workloads.","key_machinery":"The load-bearing mechanism is a prune-and-grow cycle built on dependency-aware structured pruning. A dependency graph tracks which layers are structurally coupled, so pruning a filter automatically prunes the matching channels in skip connections, batch-norm layers, and downstream convolutions; this keeps the shrunken model valid. The grow half reinserts previously removed filters into a full-sized copy of the original network and restores their original weights. During post-rebuild fine-tuning, the core (unpruned) filters are frozen and gradients are zeroed for them, so only the reinserted filters are updated; this freezing step is what is supposed to preserve the compact subnetwork inside the rebuilt model and make later runtime switching training-free.","core_discovery":"The central claim is that pruning and rebuilding can be reversed without losing the pruned model. Concretely, a trained CNN is structurally pruned with dependency-aware filter pruning, producing a smaller core model; the removed filters are not discarded but logged with their positions and weights. A full-sized copy of the original network is then created, the removed filters are reinserted in their original places, and the model is fine-tuned with the core filters frozen so that only the reintroduced filters change. The intended outcome is a single architecture containing nested capacity levels, so the compact core and the full model coexist in one set of weights and a deployment can move between them without retraining. Across the architectures tested, rebuilt models generally recovered the bulk of the lost accuracy and sometimes exceeded the original baseline.","pith_inferences":["Inference: if the frozen-core premise holds, the scheme could be applied recursively: each new build level freezes everything below it, yielding a ladder of nested models with more than two operating points from a single training run.","Inference: the same machinery could be pointed at modern architectures such as vision transformers, but the dependency graph would need to cover attention heads and layer-norm couplings; the paper does not test this.","Inference: the decisive missing experiment is to extract the compact core from a rebuilt model at inference time and compare it against an independently fine-tuned core; internal normalization statistics or shared layers could shift during post-rebuild fine-tuning and break the no-retraining switch."],"forward_implications":["A single deployed model could serve devices with different compute budgets: run the core filters when resources are tight, run the full network when they are not, with no retraining step in between.","Restoring a larger configuration costs only a short fine-tuning pass over the reinserted filters, not training from scratch, because the original weights are reused.","Iterative prune-and-grow yields several intermediate capacity levels in one model, so adaptivity can step through sizes rather than jump between two extremes.","Selective pruning choices, such as leaving alternate layers untouched, can retain more accuracy than uniform pruning, so the pruning pattern itself becomes a design lever for adaptivity.","Rebuilt models sometimes match or exceed baseline accuracy, implying the pipeline can act as a cheap capacity-increasing fine-tune rather than just a recovery step."],"supporting_citations":[{"why":"Supplies the filter-pruning formulation and L1-norm importance criterion used to shrink the model into a core.","marker":"[1]"},{"why":"Provides the dependency-graph pruning backbone that keeps pruned and rebuilt structures valid across coupled layers.","marker":"[2]"},{"why":"Introduces the freeze-and-grow multi-capacity model that this paper's prune-and-rebuild adaptivity extends.","marker":"[4]"},{"why":"Demonstrates block-grained runtime scaling and motivates rebuilding a single model from reusable parts.","marker":"[5]"},{"why":"Shows a single network can run at multiple widths, the conceptual basis for nested configurations without retraining.","marker":"[6]"},{"why":"Establishes the supernetwork idea of many configurations supported by one model, background for no-retraining adaptivity.","marker":"[12]"},{"why":"Supports the claim that sparse subnetworks can match full-model performance, underpinning iterative pruning.","marker":"[15]"}],"fun_headline_variants":["One CNN, two sizes: toggle between compact and full at runtime","Nested subnetworks let a single CNN adapt to hardware anytime","Prune-and-grow builds one model that runs small or large on demand","Same weights, two configurations: adaptive CNN without retraining","Compact and full modes coexist in one CNN for dynamic deployment"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Freezing the core filters while fine-tuning the rebuilt network preserves the compact subnetwork's behavior closely enough that a user can switch to it later without retraining — a premise the paper assumes but never directly measures.","fun_headline_variants_meta":{"raw":{"variants":["One CNN, two sizes: toggle between compact and full at runtime","Nested subnetworks let a single CNN adapt to hardware anytime","Prune-and-grow builds one model that runs small or large on demand","Same weights, two configurations: adaptive CNN without retraining","Compact and full modes coexist in one CNN for dynamic deployment"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00025,"raw_usage":{"total_tokens":1523,"prompt_tokens":883,"completion_tokens":640,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":499,"completion_tokens_details":{"reasoning_tokens":552}},"tokens_in":499,"tokens_out":640,"duration_ms":6352,"temperature":1.0,"reasoning_tokens":552,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:57:29.283531+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a rebuilt model and run inference using only its core filters, then compare accuracy against the same core model fine-tuned in isolation. If the two accuracies differ substantially, or if the internal normalization statistics of the core have shifted during post-rebuild fine-tuning, the claim that switching between sizes requires no retraining would be refuted.","supporting_citations":[{"cited_title":"write newline","cited_arxiv_id":null,"evidence_quote":"Supplies the filter-pruning formulation and L1-norm importance criterion used to shrink the model into a core."},{"cited_title":"write newline","cited_arxiv_id":null,"evidence_quote":"Provides the dependency-graph pruning backbone that keeps pruned and rebuilt structures valid across coupled layers."},{"cited_title":"write newline","cited_arxiv_id":null,"evidence_quote":"Introduces the freeze-and-grow multi-capacity model that this paper's prune-and-rebuild adaptivity extends."},{"cited_title":", \" * write output.state after.block = add.period write newline","cited_arxiv_id":null,"evidence_quote":"Shows a single network can run at multiple widths, the conceptual basis for nested configurations without retraining."}],"review_version":1}