{"id":"3c1a81c9-cba8-4357-8370-f4a1a05d2679","arxiv_id":"1909.01861","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A width-search method built on function-preserving widening finds CIFAR channel patterns that reduce parameters and slightly lower test error relative to standard network schedules.","lead":"This paper searches the number of channels in convolutional networks by repeatedly widening layers with function-preserving transformations and evolving the width pattern. It reports small accuracy gains with far fewer parameters on CIFAR-10 and CIFAR-100 using only 0.4 to 1.3 GPU-days of search.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The search selector's 15-epoch validation fitness (Sec. 4.3) is never checked against the 511-epoch converged test error (Sec. 4.4); if that ranking does not transfer, the Table 1 improvements may be selection noise.","rationale":"The reader correctly identifies the 15-epoch fitness proxy as the weakest assumption. I agree because the selection step is the only bridge between the cheap search and the converged accuracy numbers in Table 1. If the proxy ranking is uncorrelated with final performance, the reported improvements could be selection noise, especially for networks searched only once. I also considered the paper's silence about BatchNorm in the function-preserving widening equations (Sec. 3.1); with BN present, replicating only convolution weights does not preserve the network function unless BN parameters and statistics are replicated too. That is a real technical gap, but it primarily affects the search efficiency claim, whereas the proxy validation directly underwrites the accuracy claims. The proposed test would settle the proxy concern; if it passes, the paper's central assertion is substantially supported, and if it fails, the reported gains would need re-analysis with a validated proxy or a random-selection baseline. The current evidence is promising but incomplete, so the reader's CONDITIONAL verdict remains appropriate.","tokens_in":12425,"tokens_out":6339,"duration_ms":66788,"concrete_test":"Using the released code, rerun one ResNet-18 CIFAR-10 search to completion; record the 15-epoch validation accuracy for every individual in the final population. Retrain the top-5, bottom-5, and 10 randomly selected individuals to convergence with the exact Sec. 4.4 protocol (Cutout, SGDR, 511 epochs, full training set), plus the original ResNet-18 baseline, with at least 3 seeds each. Compute the Spearman rank correlation between 15-epoch fitness and converged test error, and the difference between the best-selected and baseline errors with confidence intervals. If the correlation is not significantly positive or the best-selected gain is within run-to-run noise, the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim (Table 1) rests entirely on selecting the width configuration with the highest validation accuracy after only 15 epochs of training on a mutated network (Sec. 4.3). The selected configuration is then retrained from scratch for 511 epochs with a different protocol: full training set, Cutout, lmax=0.1, SGDR (Sec. 4.4). The paper never shows that the 15-epoch validation ranking correlates with converged test error. Because the search deliberately stops when parameter counts are near the baseline, the parameter reduction is enforced by the stopping rule; what must be shown is that the accuracy gain is not an artifact of choosing among many random mutations. In fact, SE-ResNet-50 and VGG-16 were searched only once, so the reported gains (0.43% and 0.52%) are single draws. Without a proxy validation or multiple seeds, the 'discovered more efficient rules' claim is unsubstantiated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an efficient evolutionary search method, termed functionally incremental search, for discovering the number of channels at each convolutional layer of a network. The method uses function-preserving mutations (Net2Net-style) to explore a family of hand-designed width-profiles (exponential, linear, step-shaped, and their symmetric variants) while keeping the parameter count close to the original network. On CIFAR-10 and CIFAR-100, the authors report that the modified ResNet-18, ResNet-34, VGG-16, and SE-ResNet-50 achieve lower test error with fewer parameters than the original baselines, at a search cost of 0.4 to 1.3 GPU-days. Supplementary experiments on ResNet-18 and PyramidNet-110 suggest that 'fluctuating' channel widths can outperform both the conventional sharp increase at downsampling and the gradual pyramid-shaped increase. The paper claims this as evidence for rethinking the design rule for the number of channels.","tokens_in":12624,"tokens_out":7207,"duration_ms":69479,"significance":"If the reported results hold, the paper makes a useful empirical contribution: it demonstrates that a function-preserving evolutionary search over channel widths can improve the accuracy of several standard CNNs on CIFAR-10/100 while reducing parameter counts, with a modest search budget. The PyramidNet-110 transfer experiment is genuine independent support for the claim that fluctuating widths can generalize beyond the searched ResNet-18. The authors also provide code for the search and for some discovered architectures, which is a credit to reproducibility. However, the strength of the conclusions is limited by the lack of a check that the short-episode validation fitness used for selection correlates with final converged accuracy, by the absence of error bars for two of the searched networks, and by the omission of FLOPs, which are needed to support the 'efficiency' framing.","major_comments":[{"comment":"The search selects the individual with the highest validation accuracy after 15 epochs of training on a mutated network (§4.3), but the final comparison is obtained by retraining the selected architecture from scratch for 511 epochs with Cutout and lmax=0.1 (§4.4). The paper does not verify that the 15-epoch validation ranking correlates with converged test accuracy. Since the search intentionally stops when the parameter count approaches the baseline, the accuracy improvements in Table 1 could partly arise from selecting among many random mutations rather than from the width rule. Please provide a correlation analysis (e.g., a scatter plot of 15-epoch validation accuracy versus 511-epoch test error for a sample of visited architectures) and report error bars across multiple independent search runs.","section":"§4.3–4.4"},{"comment":"For VGG-16 and SE-ResNet-50, the text states that the search was performed only once, so the reported improvements of 0.52% and 0.43% are single draws. Without multiple seeds or a variance estimate, these gains cannot be distinguished from selection noise. The three rows for ResNet-18 provide some evidence of reproducibility, but the paper should either add multiple search runs for the other two networks or explicitly temper the claim that the method 'can discover more efficient rules.'","section":"Table 1 and §4.3"},{"comment":"The paper's efficiency claim rests on parameter counts and search time, but inference FLOPs are not reported. The modified ResNet-18 in Table 4 has roughly 200 channels in the first layer, compared with 64 in the original, so the computational cost of early layers could be substantially higher even though the total parameter count is lower. Since the abstract and conclusions describe the discovered widths as 'more efficient' and the authors themselves acknowledge in §4.5 that FLOPs are important, please report FLOPs for all architectures in Table 1 (and Table 3 if possible), or restrict the claim to parameter efficiency explicitly.","section":"Table 1 and §4.5"}],"minor_comments":[{"comment":"The stopping criterion 'until the number of parameters of the individual with the highest fitness is similar to the comparison network' is not quantified; please state the tolerance used.","section":"§4.3"},{"comment":"The parameter count for the original ResNet-18 is 11.18M in Table 2 but 11.54M in Table 1; please reconcile this discrepancy or clarify that the two tables use different ResNet-18 variants.","section":"Table 2"},{"comment":"The row labels 'Increasing number of channels (Modified ResNet-18)' and 'Decreasing number of channels (Modified ResNet-18)' are ambiguous; the text says the opposite order was used, so state explicitly which row corresponds to the searched order and which to its reverse.","section":"Table 2 and §4.5"},{"comment":"Please specify the initial channel count used for each network (ResNet-34, VGG-16, SE-ResNet-50), not only for ResNet-18, to clarify the search-space construction.","section":"§4.1"},{"comment":"There are several typos: 'slop' should be 'slope', 'convoluaional' in Section 4.1 should be 'convolutional', and 'filed' in Section 2 should be 'field'. Please also check the author list for spacing issues.","section":"Section 3.2"},{"comment":"The paper does not compare the proposed search to prior automatic width-search methods such as DenseNAS, EfficientNet, or FBNet; a brief comparison or discussion would help position the contribution.","section":"Related Work"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the core idea is reasonable, but the empirical validation needs strengthening along the lines of the major comments. The self-citation to EENA is understandable given the methodological lineage, but the novelty over EENA should be stated more crisply in the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a legitimate, fairly cheap width-search method that consistently finds lower-parameter networks with small accuracy gains on CIFAR-10/100, and it offers a non-obvious qualitative rule — fluctuating channel widths beat both sharp increases at downsampling and gradual increases everywhere. The main weakness is exactly what you flagged: the search selects on validation accuracy after only 15 epochs of training on mutated nets, while the final comparison trains from scratch for 511 epochs with a different protocol. The paper never checks that the 15-epoch ranking transfers to converged accuracy, so some of the Table 1 improvements could be selection noise. On top of that, most networks were searched only once (SE-ResNet-50 and VGG-16 are single draws), there are no error bars, and FLOPs are not reported. These are real problems, but they are addressable rather than fatal.\n\nWhat is actually new: the specialization of function-preserving Net2Net widening to channel counts, with a hand-designed family of increment functions and an evolutionary search that controls parameter count. The appendix publishes the discovered channel configurations, and the code is available, so the work is reproducible. The PyramidNet-110 transfer experiment is the most convincing piece — a fluctuating schedule beats the original additive pyramid with similar parameters, which is independent support for the rule and not just an artifact of the search protocol. The consistency of improvements across four different architectures also makes it unlikely that all of the gains are noise.\n\nWhere it falls short: the fitness-proxy issue is central, not minor. The stopping rule enforces that parameter counts end up near baseline, so the parameter reduction is partly by construction; what needs demonstration is that the accuracy gain is not just picking a lucky mutation from a noisy pool. Single runs for most networks mean the reported gains of 0.43–0.52% on CIFAR-10 could easily be within run-to-run variation. The author's own note that decreasing channels can look good only when FLOPs are ignored is an honest caveat, but it also exposes that the claimed efficiency is about parameters, not compute. The 2.33% gain on CIFAR-100 for SE-ResNet-50 is the headline number and it comes from a single search; that needs a seed study before I'd trust it.\n\nWho it's for: people working on cheap NAS, channel pruning, or efficient CNN design. It deserves a serious referee — the idea is plausible, the code and configurations are shared, and the PyramidNet result is a genuine hook. The referee should ask for proxy validation (correlate 15-epoch fitness with converged error), a few repeats to get error bars, and a FLOPs comparison. With those additions the paper could be solid; without them, treat Table 1 as suggestive, not conclusive.","headline":"A cheap channel-width search that finds lower-parameter networks with small CIFAR gains, plus a genuinely interesting qualitative rule (fluctuating widths beat monotone schedules), but the unvalidated 15-epoch fitness proxy and single search runs leave the central claim under-supported.","tokens_in":13157,"tokens_out":1581,"would_cite":true,"duration_ms":17881,"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":"Width-only search lifts CNN accuracy up to 2.33 percent while cutting parameters","keywords":["neural architecture search","channel width search","function-preserving transformation","evolutionary algorithm","convolutional neural networks","CIFAR-10","CIFAR-100","network morphism"],"falsifier":"Take the mutated networks evaluated during one search run, record their validation accuracies at 15 epochs, train each one from scratch to full convergence, and measure the rank correlation between the two orderings. If the correlation is low, or if a control network with the same parameter count and a monotonic width profile converges to equal or better accuracy, the claim that the discovered width rules explain the gains is not supported.","tokens_in":12236,"feed_emoji":"🧠","tokens_out":7021,"duration_ms":69940,"temperature":0.7,"pith_summary":"This paper claims that the number of channels in a convolutional network can be searched automatically with very little compute, and that the channel patterns found this way outperform both the sharp channel doublings used in classic networks and the gradual pyramid increases used in others. The search starts from a small uniform-width version of the target network and repeatedly applies function-preserving mutations that widen individual layers along a family of curve shapes, with an evolutionary selector keeping the mutations with the highest validation accuracy. The paper reports improved test accuracy of about 0.5 percent on CIFAR-10 and up to 2.33 percent on CIFAR-100 while using fewer parameters, with total search cost of 0.4 to 1.3 GPU-days. If the claim holds, width is a cheap, largely independent design axis that can be tuned without a large search budget.","feed_headline":"Width-only search lifts CNN accuracy up to 2.33 percent","feed_subtitle":"Four classical networks gained accuracy with fewer parameters in 0.4 to 1.3 GPU-days of search.","key_machinery":"The machinery is a function-preserving layer-widening operation: to add channels to a convolutional layer, existing filters are replicated and the incoming weights of the next layer are divided by the replication count, with a small random noise added, so the network's input-output function is unchanged. Search wraps this operation in a set of width-curve mutation functions, including increasing, constant, and decreasing slopes, step shapes at downsampling points, symmetric counterparts, and a uniform addition, applied cumulatively to an initial uniform-width model. Because each mutation preserves the function, mutated networks can be evaluated after only 15 epochs of training, which is what makes the search cheap.","core_discovery":"The central discovery is that the widths of a convolutional network can be optimized separately from its connectivity, and that the winning width profiles are neither the classic double-at-every-downsampling pattern nor the continuous pyramid pattern; they are irregular, fluctuating allocations that keep total parameters below the baseline while lowering error. The paper establishes this by running its functionally incremental search on several classical network families, selecting the highest-fitness individual, and retraining it from scratch with the same recipe as the original network. In every reported case the modified network has fewer parameters and lower test error, with the largest reported gain being 2.33 percent on CIFAR-100.","pith_inferences":["Inference: if the 15-epoch fitness ranking carries over to converged accuracy on larger datasets, the same set of width-curve mutations could be applied to ImageNet-scale backbones, where full architecture search is usually unaffordable; the paper does not test this.","Inference: the fluctuating width profiles suggest a separation between parameter efficiency and FLOPs efficiency, so a width profile that wins on accuracy-per-parameter may not win on accuracy-per-FLOP, which matters for deployment.","Inference: because the mutation machinery acts layer by layer, width profiles discovered for one backbone could be transplanted onto a differently connected backbone to test whether the gains transfer with the widths alone.","Inference: adding parameter count and FLOPs to the fitness function, as the paper says it plans to do, would let the search target a specific compute budget and would likely change which width curves win."],"forward_implications":["The best width profiles improve CIFAR-10 accuracy by about 0.5 percent and CIFAR-100 accuracy by up to 2.33 percent while using fewer parameters than the original networks.","Classic width rules, namely sharp doublings at downsampling and continuous pyramid growth, are outperformed by fluctuating profiles at lower parameter counts in the reported settings.","The search is cheap enough to run per network and per dataset: 0.4 to 1.3 GPU-days on a single GPU.","The same function-preserving search can be applied to almost any convolutional architecture, since it only needs the layer structure as input."],"supporting_citations":[{"why":"Supplies the function-preserving widening operation that lets mutated networks be evaluated after only 15 epochs of training.","marker":"Chen, Goodfellow, and Shlens 2016"},{"why":"Provides the ResNet baselines and initialization scheme used in the main comparisons and initial models.","marker":"He et al. 2016a"},{"why":"Provides prior evidence that width is an important design axis and a baseline for width-based improvements.","marker":"Zagoruyko and Komodakis 2016"},{"why":"Supplies the PyramidNet architecture and the gradual-increase width rule that the paper compares against in supplementary experiments.","marker":"Han, Kim, and Kim 2017"},{"why":"Provides the SGDR training schedule used for both the search-phase training and the final converged training.","marker":"Loshchilov and Hutter 2017"}],"fun_headline_variants":["Irregular CNN channel counts boost accuracy up to 2.33%","Fast width search finds better CNN channels with fewer parameters","CNN width optimization improves accuracy, cuts parameters","Function-preserving search tunes CNN widths for accuracy","Width-only neural architecture search improves CNN accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The search selects networks by validation accuracy after only 15 epochs of continued training, and the paper assumes that this quick ranking predicts test accuracy after the final training from scratch to convergence; if that correlation fails, the reported improvements could be selection noise rather than consequences of the width rules.","fun_headline_variants_meta":{"raw":{"variants":["Irregular CNN channel counts boost accuracy up to 2.33%","Fast width search finds better CNN channels with fewer parameters","CNN width optimization improves accuracy, cuts parameters","Function-preserving search tunes CNN widths for accuracy","Width-only neural architecture search improves CNN accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000822,"raw_usage":{"total_tokens":3539,"prompt_tokens":833,"completion_tokens":2706,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":449,"completion_tokens_details":{"reasoning_tokens":2631}},"tokens_in":449,"tokens_out":2706,"duration_ms":19157,"temperature":1.0,"reasoning_tokens":2631,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:06:04.577277+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the mutated networks evaluated during one search run, record their validation accuracies at 15 epochs, train each one from scratch to full convergence, and measure the rank correlation between the two orderings. If the correlation is low, or if a control network with the same parameter count and a monotonic width profile converges to equal or better accuracy, the claim that the discovered width rules explain the gains is not supported.","supporting_citations":[{"cited_title":"J.; and Shlens, J","cited_arxiv_id":null,"evidence_quote":"Supplies the function-preserving widening operation that lets mutated networks be evaluated after only 15 epochs of training."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides prior evidence that width is an important design axis and a baseline for width-based improvements."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the PyramidNet architecture and the gradual-increase width rule that the paper compares against in supplementary experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the SGDR training schedule used for both the search-phase training and the final converged training."}],"review_version":1}