{"id":"6c675053-c6ec-4df6-b673-e8609cee4e29","arxiv_id":"2501.16917","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Bayesian optimization of per-layer-group pruning rates consistently improves the size-accuracy trade-off of FPGM-pruned lightweight face detectors on WIDER FACE.","lead":"This paper combines Bayesian optimization with FPGM and soft filter pruning to automatically choose per-layer pruning rates for lightweight face detectors. The resulting pruned EResFD and EXTD models achieve better accuracy at smaller sizes than uniform pruning, which matters for deploying face detection on edge devices.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 3 contradicts the 'consistently superior' claim: at T=10% B-FPGM has lower mAP than uniform FPGM on Easy/Medium/Hard while being 10.24% vs 5.25% pruned; all comparisons use unmatched actual sparsities.","rationale":"The reader's conditional verdict is reasonable, but the single most load-bearing concern is not the one-epoch Bayesian-optimization proxy. The reported numbers in Table 3 already contain an internal counterexample to the strongest claim: at T=10%, B-FPGM has worse mAP on all WIDER FACE subsets than uniform FPGM while being more aggressively pruned. Moreover, the paper compares nominal targets rather than matched actual sparsities, and B-FPGM consistently ends up with a larger pruning budget at every target. That does not invalidate the method's apparent advantage at higher rates, but it does mean the headline 'consistently superior' claim is overbroad and the comparison is not confound-controlled. The proxy concern is real but speculative; it should be checked separately, for example by correlating Eq. 1 values with final mAP. My recommendation is to keep the CONDITIONAL verdict: the paper should qualify the claim and add size-matched comparisons before acceptance.","tokens_in":557,"tokens_out":8519,"duration_ms":157350,"concrete_test":"Rerun the uniform FPGM baseline at the actual sparsities achieved by B-FPGM on EResFD (10.24%, 22.27%, 31.59%, 40.02%, 50.37%, 59.87%) instead of the nominal targets, and plot mAP against actual parameter count. If uniform FPGM at 10.24% sparsity matches or exceeds B-FPGM's T=10% mAP (0.8622/0.8506/0.7636), the 'consistently superior' claim fails at low sparsity; if uniform FPGM loses at every matched size, the qualitative conclusion survives at all but requires the T=10% caveat.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that B-FPGM 'consistently' achieves a superior size-to-performance trade-off compared with uniform FPGM on EResFD and EXTD. Table 3 falsifies this at the T=10% row: uniform FPGM scores 0.8728/0.8582/0.7757 on Easy/Medium/Hard at 5.25% actual sparsity, whereas B-FPGM scores 0.8622/0.8506/0.7636 at 10.24% actual sparsity. B-FPGM is both smaller and worse on every subset, so the Pareto trade-off is not superior at low sparsity. The same row also reveals a systematic mismatch: at the same nominal T, B-FPGM's actual sparsity is consistently higher than uniform FPGM's (10: 10.24 vs 5.25; 20: 22.27 vs 16.84; 30: 31.59 vs 24.36; 40: 40.02 vs 35.95; 60: 59.87 vs 54.05). Thus Fig. 3 compares B-FPGM with a larger effective pruning budget. Even if B-FPGM wins at higher T, the claim must be restricted to moderate/high pruning rates and supported by size-matched comparisons.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes B-FPGM, a structured pruning pipeline for lightweight face detectors that combines FPGM filter pruning, Soft Filter Pruning (SFP), and Bayesian optimization. Instead of applying a uniform pruning rate to all layers, the method groups layers (N=6 by default), uses Bayesian optimization to select per-group pruning rates for a target sparsity T, and then applies the optimized rates within the SFP/FPGM training scheme of Gkrispanis et al. Experiments on the WIDER FACE benchmark compare B-FPGM with uniform FPGM pruning on the EResFD detector across targets from 10% to 60%, plus a smaller EXTD ablation and grouping ablations. The paper claims that B-FPGM consistently achieves a superior size-to-performance trade-off, with code and trained models publicly available.","tokens_in":12783,"tokens_out":4577,"duration_ms":41320,"significance":"If the empirical claims hold, the paper offers a practical contribution: automated per-group pruning-rate selection for small face detectors, with source code and trained models, a two-architecture demonstration, and a five-seed stability check at T=20%. The method is simple and reproducible in principle. However, the evidence as presented is not yet sufficient to support the central claim of consistent superiority: the comparison is made at mismatched actual sparsity levels, the low-sparsity regime contradicts the claim, and only a single pruning baseline is used. The core idea is plausible and the high-sparsity gains are large, but those gains need to be separated from the confound of larger effective pruning budgets.","major_comments":[{"comment":"The stated central claim that B-FPGM 'consistently' achieves a superior size-to-performance trade-off is directly contradicted by the T=10% row. At this target, uniform FPGM reaches mAP 0.8728/0.8582/0.7757 on Easy/Medium/Hard at 5.25% actual sparsity, while B-FPGM reaches only 0.8622/0.8506/0.7636 at 10.24% actual sparsity. B-FPGM is therefore both smaller and worse on every subset at low sparsity. The claim must be restricted to moderate-to-high pruning rates, or a low-sparsity regime must be added where B-FPGM is not worse.","section":"Table 3, T=10% row"},{"comment":"The comparison is confounded by systematically unmatched actual sparsity. For every target T, B-FPGM's actual sparsity is higher than uniform FPGM's (e.g., T=20: 22.27% vs 16.84%; T=60: 59.87% vs 54.05%). Fig. 3 plots nominal target sparsity, not actual sparsity, so the visible gains at high T may partly reflect that B-FPGM is being compared at a larger effective pruning budget. The authors should provide a size-matched comparison, for example by plotting mAP against actual sparsity on the horizontal axis or by running uniform FPGM at the same actual sparsity as each B-FPGM model, and should report the actual sparsity values on the plots.","section":"Tables 3 and 4, Fig. 3"},{"comment":"The entire Bayesian optimization step rests on the assumption that the validation loss after a trial soft-pruning and one epoch of training is a reliable proxy for the final mAP after the full soft-pruning/retraining and hard-pruning/fine-tuning pipeline. The paper provides no evidence for this correlation, such as a scatter plot of proxy values versus final performance for a sample of pruning-rate vectors. If this one-epoch proxy is noisy or misleading, the optimized rates could be worse than uniform rates, which would invalidate the central comparison. To make the method trustworthy, the authors should validate the proxy empirically or provide a theoretical/empirical argument for why one epoch suffices.","section":"Section 3.3, Eq. (1)"},{"comment":"The Abstract and Section 5 claim superiority over 'existing approaches' and 'SoA' in balancing size and performance, but the pruning comparison in Tables 3 and 4 is only against uniform FPGM pruning. Fig. 4 compares against other face detectors, not other pruning methods. To support the broader claim, the authors should either add at least one additional pruning baseline (e.g., L1-norm pruning, AMC-style automatic pruning, or a magnitude-based non-uniform baseline) or explicitly restrict the claim to 'superior to uniform FPGM at moderate-to-high sparsity on EResFD and EXTD.'","section":"Section 4.2 and Fig. 4"}],"minor_comments":[{"comment":"The pseudocode loop 'while i <= I' never increments i, so the algorithm as written would run forever; add an explicit 'i ← i + 1' at the end of the loop body.","section":"Algorithm 1"},{"comment":"The acquisition function is called 'UPC' but is described as the Upper Confidence Bound; it should be 'UCB' to match standard terminology.","section":"Section 3.3"},{"comment":"The first row is labeled 'EXTD (orig.) [23]' but EXTD is reference [56], not [23]; the citation should be corrected.","section":"Table 4"},{"comment":"The notation for the bounds is unclear: the text says the bounds are '[0, T+ bound of f set]' and Table 2 lists 'T + 0.04' and 'bound of f set 0.2' without a clear formula. Please define the bound explicitly, e.g., upper bound = T + offset, and state the offset value in a consistent notation.","section":"Table 2 and Section 3.3"},{"comment":"The condition 'T − T + ≤ S(θϕi ) ≤ T + T +' is visually ambiguous because 'T+' is used both as a threshold symbol and an addition operator; use an explicit threshold variable, such as τ, to avoid confusion.","section":"Eq. (1)"},{"comment":"The main results tables and figures do not report error bars or multiple-seed variability, even though Table 6 shows that B-FPGM has moderate variance at T=20% (std up to 0.0066 on Hard). Reporting the five-seed mean and standard deviation for all main results would strengthen the comparison.","section":"Tables 3, 4, and Fig. 3"},{"comment":"References [59] and [60] appear to be the same paper, but they are listed as separate entries and cited differently (Eagle Eye [59] in the text vs [60] in Fig. 4); this should be cleaned up.","section":"References [59] and [60]"}],"recommendation":"major_revision","confidential_remarks":"The central idea is reasonable and the high-sparsity results are promising, but the manuscript needs a size-matched comparison and a qualified claim before publication. The T=10% contradiction and the unmatched actual sparsity are load-bearing issues, not cosmetic ones. The related-work claim that this is the 'first application of Bayesian optimization to structured pruning' also seems overly strong, since automatic per-layer pruning-rate search by other means exists (e.g., AMC); the authors should verify this claim carefully."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper ships a clean, reproducible engineering result—Bayesian-optimized per-group pruning rates beat uniform FPGM at moderate-to-high pruning for two lightweight face detectors—but the 'consistently superior' claim and the 'first application of Bayesian optimization to structured pruning' claim both overreach. The T=10% row of Table 3 undercuts the consistency language, and the actual-sparsity mismatch makes the comparison noisier than the text admits.\n\nWhat's new: replacing the uniform pruning vector in FPGM/SFP pruning of EResFD with per-group rates found by Bayesian optimization. The experiments are honest in that they report per-group rates, target vs actual sparsity, parameter counts, and an ablation over layer grouping; they also include a five-seed run at T=20. That is more than many pruning papers provide. Applied to EResFD and EXTD, the optimized rates do produce better mAP at equal-or-larger actual sparsity for T=20 through T=60, with the largest gains at 50–60% pruning. The grouping ablation shows N=6 is a sensible default. The GitHub with code and trained models is a creditable part of the submission.\n\nSoft spots, in order. First, Table 3 at T=10% directly contradicts 'consistently': B-FPGM scores 0.8622/0.8506/0.7636 at 10.24% actual sparsity, while uniform FPGM scores 0.8728/0.8582/0.7757 at 5.25%. B-FPGM is smaller and worse on every subset—that is a worse Pareto point, not a better trade-off. Second, actual sparsity is not matched between methods at any row; B-FPGM consistently prunes more than uniform at the same nominal T. So Fig. 3 is partly comparing different sparsity levels. The conclusions should be restricted to moderate-to-high pruning rates and supported by size-matched comparisons. Third, the objective is a one-epoch validation loss after a trial soft prune; the whole search leans on that proxy. The authors give no evidence that the proxy correlates with final mAP. It may be fine, but it is load-bearing and untested. Fourth, the 'first application' claim is not substantiated and is very likely false; the related work cites BO for unstructured pruning but does not survey BO with structured pruning. That claim should be dropped. Fifth, error bars are only for T=20; the rest of Tables 3–5 are single runs.\n\nWho this is for: practitioners working on tiny face detectors for edge deployment, and people comparing pruning-rate search strategies. A serious referee should engage; the method is plausible, reproducible, and the empirical result at high pruning rates is nontrivial. With the consistency claim fixed, the unmatched-sparsity comparisons addressed, and the 'first' claim removed, I'd be happy to see it accepted.","headline":"Useful BO-per-group pruning result for tiny face detectors, but the 'consistently superior' and 'first BO structured pruning' claims both overreach; worth a serious referee after fixes.","tokens_in":13268,"tokens_out":2545,"would_cite":false,"duration_ms":22765,"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":"This paper claims that Bayesian optimization of per-layer-group pruning rates, combined with geometric-median soft pruning, yields a better size-versus-accuracy trade-off for tiny face detectors than applying a single uniform pruning rate…","keywords":["face detection","network pruning","Bayesian optimization","filter pruning","geometric median","soft filter pruning","lightweight models","WIDER FACE"],"falsifier":"Take a fixed target sparsity, for example 20%, and generate many random pruning-rate vectors; for each, record the one-epoch validation-loss proxy and then run the complete soft-prune, retrain, hard-prune, and fine-tune pipeline to get final WIDER FACE mAP. If the rank correlation between the proxy and the final mAP is weak, the optimization objective is not carrying the argument. A direct version is to find a random vector that scores worse on the proxy but, after full training, beats the B-FPGM rates at matched sparsity.","tokens_in":12232,"feed_emoji":"🎯","tokens_out":5556,"duration_ms":46076,"temperature":0.7,"pith_summary":"This paper tries to show that the per-layer pruning-rate decisions in structured pruning of face detectors should not be uniform, and that Bayesian optimization can discover better non-uniform rates automatically. The authors combine geometric-median filter pruning (FPGM) with soft filter pruning (SFP) and a Bayesian search over layer-group pruning rates, then apply the result to two of the smallest competitive face detectors, EResFD and EXTD, evaluating on WIDER FACE. Across target sparsities from 10% to 60%, the Bayesian-optimized versions consistently keep more detection accuracy (mAP) than the same pipeline pruned uniformly at the same or even smaller model size. The payoff is that edge-device face detectors can be compressed substantially, up to roughly 60% of parameters, with a much gentler accuracy decline than uniform pruning delivers, and the work is the first to apply Bayesian optimization to structured pruning for face detection.","feed_headline":"Bayesian pruning keeps tiny face detectors accurate at 50% sparsity","feed_subtitle":"Per-layer rate search beats uniform pruning, holding face-detection mAP high while cutting parameters nearly in half.","key_machinery":"The machinery is a three-stage pipeline. First, the network is partitioned into $N$ layer groups, with a default of 6, to keep the search space small. Second, Bayesian optimization with an upper-confidence-bound acquisition function searches for the vector of per-group pruning rates $\\phi^*$, guided by an objective that combines the validation loss after a single-epoch soft-pruned training run with a penalty for missing the target sparsity. Third, the optimized rates are applied through the FPGM/SFP procedure: filters closest to the geometric median of each layer's filter set are iteratively soft-pruned, meaning zeroed but still updated, during retraining, then hard-pruned and fine-tuned. The geometric median acts as a proxy for redundant filters, SFP lets the network recover from pruning during training, and Bayesian optimization removes the need to hand-engineer layer-wise rates.","core_discovery":"The central claim is that replacing the uniform pruning vector $\\phi=[T,\\ldots,T]$ with a per-group vector $\\phi^*$ found by Bayesian optimization changes the size–performance trade-off of FPGM-based soft pruning in a way that dominates uniform pruning for lightweight face detectors. On EResFD at target sparsities 10–60%, B-FPGM achieves mAP on WIDER FACE that is close to the original model at sparsities up to 50%, while uniform FPGM collapses at 50% and 60%, with Hard-subset mAP falling from 0.7757 to 0.5254 and 0.2936, versus B-FPGM's 0.6993 and 0.6051 at slightly higher actual sparsity. The authors attribute the advantage to the optimizer concentrating pruning in later layer groups, Groups 3–6 of EResFD, where redundancy is higher, while sparing early layers. A smaller ablation on EXTD and a grouping ablation support the claim that the method transfers across architectures.","pith_inferences":["The one-epoch surrogate may favor conservative pruning in early layers because loss after one epoch is dominated by layers that are hard to recover, so a multi-epoch or learned surrogate could discover even better allocations.","The method's reliance on hand-defined layer groups leaves on the table some of the benefit of full per-layer optimization, and automatic grouping by sensitivity could push the trade-off further.","Because the Bayesian optimizer only sees a single epoch, rates that look good early may be suboptimal after the full 200-epoch soft-pruning schedule, and ablating the number of surrogate epochs would test this.","Applying the same optimization to detection heads or to latency-aware objectives such as FLOPs or inference time, rather than parameter count alone, could yield different pruning patterns."],"forward_implications":["Tiny face detectors can be compressed to around half their parameters while keeping practical accuracy, making them more deployable on phones, drones, and other edge devices.","The improvement is largest at high sparsities of 50–60%, where uniform pruning fails but B-FPGM retains usable detection, so the method extends the usable compression range.","Because the optimizer automatically chooses which layers to prune, engineering effort shifts from manual rate tuning to defining layer groups and a reliable validation-loss proxy.","The authors show transfer to a second architecture, EXTD, suggesting the method does not depend on EResFD-specific details."],"supporting_citations":[{"why":"Provides the uniform-FPGM pruning pipeline and baseline results that B-FPGM is compared against.","marker":"[12]"},{"why":"Defines the geometric-median filter pruning criterion used to select redundant filters.","marker":"[19]"},{"why":"Introduces soft filter pruning, the training scheme that lets zeroed filters be updated and recover.","marker":"[17]"},{"why":"Supplies the EResFD architecture and pre-training setup that the main experiments prune.","marker":"[23]"},{"why":"Supplies the WIDER FACE benchmark and the Easy/Medium/Hard evaluation protocol.","marker":"[55]"},{"why":"Supplies the EXTD architecture used in the transferability ablation.","marker":"[56]"},{"why":"Provides the Bayesian optimization framework and rationale for using it with expensive objective functions.","marker":"[11]"}],"fun_headline_variants":["Bayesian-tuned pruning beats uniform for tiny face detectors","Per-layer pruning search keeps face detection accurate at 50% sparsity","Smart pruning rates preserve face detector mAP at high sparsity","Optimal per-layer pruning trumps uniform for small face detectors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire rate search stands on the assumption that the validation loss measured after one short training epoch on a trial-pruned network reliably predicts how accurate the fully pruned and fine-tuned model will end up being.","fun_headline_variants_meta":{"raw":{"variants":["Bayesian-tuned pruning beats uniform for tiny face detectors","Per-layer pruning search keeps face detection accurate at 50% sparsity","Smart pruning rates preserve face detector mAP at high sparsity","Optimal per-layer pruning trumps uniform for small face detectors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000552,"raw_usage":{"total_tokens":2665,"prompt_tokens":1013,"completion_tokens":1652,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":629,"completion_tokens_details":{"reasoning_tokens":1580}},"tokens_in":629,"tokens_out":1652,"duration_ms":11218,"temperature":1.0,"reasoning_tokens":1580,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T05:36:44.297385+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a fixed target sparsity, for example 20%, and generate many random pruning-rate vectors; for each, record the one-epoch validation-loss proxy and then run the complete soft-prune, retrain, hard-prune, and fine-tune pipeline to get final WIDER FACE mAP. If the rank correlation between the proxy and the final mAP is weak, the optimization objective is not carrying the argument. A direct version is to find a random vector that scores worse on the proxy but, after full training, beats the B-FPGM rates at matched sparsity.","supporting_citations":[{"cited_title":"Filter-pruning of lightweight face detectors using a geometric median criterion","cited_arxiv_id":null,"evidence_quote":"Provides the uniform-FPGM pruning pipeline and baseline results that B-FPGM is compared against."},{"cited_title":"Filter pruning via Geometric Median for deep convolutional neural networks acceleration","cited_arxiv_id":null,"evidence_quote":"Defines the geometric-median filter pruning criterion used to select redundant filters."},{"cited_title":"Soft filter pruning for accelerating deep convolutional neural networks","cited_arxiv_id":null,"evidence_quote":"Introduces soft filter pruning, the training scheme that lets zeroed filters be updated and recover."},{"cited_title":"EResFD: Rediscovery of the effectiveness of standard convolution for lightweight face detection","cited_arxiv_id":null,"evidence_quote":"Supplies the EResFD architecture and pre-training setup that the main experiments prune."},{"cited_title":"WIDER FACE: A face detection benchmark","cited_arxiv_id":null,"evidence_quote":"Supplies the WIDER FACE benchmark and the Easy/Medium/Hard evaluation protocol."},{"cited_title":"EXTD: Extremely Tiny Face Detector via Iterative Filter Reuse","cited_arxiv_id":"1906.06579","evidence_quote":"Supplies the EXTD architecture used in the transferability ablation."}],"review_version":1}