{"id":"a96edca5-e3ea-4cfd-b58d-8c01143355ca","arxiv_id":"2509.04244","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Simultaneous or sequential integration of geometric-median filter pruning with 4-bit additive-power-of-two quantization compresses ResNet and VGG models on CIFAR-10 by about 15x with modest accuracy loss.","lead":"This paper combines two existing tricks for shrinking neural networks, filter pruning and low-bit quantization, in two different training schedules. The authors report up to 15x smaller models with accuracy losses around 1-2% on CIFAR-10 image classification.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Residual/skip-connection handling is absent: 30% per-conv filter pruning on ResNets requires coordinated channel removal across shortcuts, but Algorithms 1–2 and §III-A never describe it; without it the claimed ×15 size/BOP reductions are not reproducible.","rationale":"The central claim requires that integrating GM pruning and APoT quantization yields roughly ×15 compression with minimal accuracy loss. The paper's tables make this outcome plausible, but the described algorithms omit essential ResNet-specific channel bookkeeping: structured filter pruning in residual networks requires coordinated channel selection across conv branches and skip connections, and no such rule appears in Algorithms 1–2 or Section IV. This is more fundamental than the reader's concern about GM being computed on full-precision weights under quantization: even if the GM ranking is perfect, the skip-connection coordination is a precondition for the reported model-size reduction to be implementable. The GM/quantization mismatch is a secondary concern that mainly affects SPQ's pruning-criterion justification; PPQ, which prunes before quantization, is less exposed to it. The missing residual handling affects both methods on all ResNet architectures, which carry most of the empirical support. I recommend CONDITIONAL rather than REJECT because the gap is an omission that could be resolved by releasing code and specifying the shortcut-handling rule; the numerical results are not inherently implausible. Additional minor issues include the absence of error bars, unspecified training epochs for each architecture, and an overclaim in the conclusion that PPQ 'consistently achieves superior accuracy' when SPQ is better on VGG-16.","tokens_in":17929,"tokens_out":12250,"duration_ms":126546,"concrete_test":"Reproduce ResNet-20 on CIFAR-10 from Algorithms 1 and 2 alone, with p=30% per convolutional layer. (a) Without adding any shortcut-coordination rule, attempt to prune the final conv of every residual block and record whether the residual addition has matching channel counts. (b) If the only way to make it run is to zero-mask filters without removing channels, recompute model size with Eq. (5) and check whether the reported ×15.78 is achieved. A pass requires code or instrumentation showing the exact per-block channel-selection strategy for both conv and shortcut paths, and a final stored-weight count equal to the reported 0.55×10^6 bits for ResNet-20.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"To obtain the claimed ×15.78–×15.94 model-size reductions (Tables II, Eq. 5), pruned filters must be physically removed, not merely zero-masked, and removing an output filter from a conv layer also removes the corresponding input channels of the following conv (§III-A). In ResNet basic blocks, the final conv of each block is followed by a residual addition with a shortcut. If 30% of filters are removed from both convs in a block, the number of output channels of the block no longer matches the shortcut channels; if only the first conv is pruned, the stated 'all convolutional layers, 30%' and the computed compression ratios change. A correct ResNet filter-pruning scheme must select a common channel subset for both conv branches and for the shortcut path (and for downsampling shortcut convs), or else must explicitly retain pruned channels through the skip connection. Neither Algorithms 1–2 nor Section IV states such a rule. If filters are only zero-masked while all channels are retained in the dataflow, the forward pass is dimensionally valid, but Eq. (5) then counts the zeroed filters as stored weights, so the ×15 model-size reductions do not follow. The paper therefore does not specify a reproducible procedure that produces the headline numbers on any ResNet. Since ResNet-20/32/56/110 are the primary experimental support for the central claim, this omission is the most load-bearing weakness.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two pipelines that combine GM-based structured filter pruning with APoT low-bit quantization: SPQ, which applies pruning and quantization simultaneously during training, and PPQ, which first prunes and then performs quantization-aware training. Experiments on CIFAR-10 with ResNet-20/32/56/110 and VGG-16 report roughly ×15.8–×15.95 model-size reduction and ×115–×126 BOPs reduction with small accuracy drops, and the authors claim PPQ achieves the best accuracy on ResNet-32 and ResNet-110 while SPQ is best on VGG-16. The paper is an empirical integration of existing techniques; it provides no new theoretical derivation.","tokens_in":18363,"tokens_out":4962,"duration_ms":49607,"significance":"If the reported results are reproducible, the two pipelines would be practically useful: combining 30% structured pruning with 4-bit APoT quantization yields roughly a further 2× compression over quantization alone, with reported accuracy drops mostly between 0.12% and 2.21%. The manuscript is strengthened by reporting comparisons against several SOTA methods on public benchmarks and by defining the compression metrics explicitly. However, the contribution is primarily incremental—GM pruning and APoT quantization are both published methods—and several load-bearing experimental details are missing. The paper also ships no code or machine-checked artifacts, so the credibility of the headline numbers rests entirely on the textual description, which is currently incomplete.","major_comments":[{"comment":"The pruning procedure is not specified for residual networks. Both algorithms select filters per convolutional layer and zero-mask them, but neither describes how pruning interacts with ResNet skip connections. If 30% of filters are removed from both conv layers in a basic block, the block's output channel count no longer matches the shortcut; if only the first conv is pruned, the stated 'all convolutional layers, 30%' policy and the computed compression ratios change. If the filters are only zero-masked and the channels are kept, then Eq. (5) counts those zeroed weights, so the reported ×15.78–×15.94 model-size reductions (and corresponding BOPs reductions) do not follow. The authors must state the exact channel-removal rule for ResNet blocks, including downsampling shortcuts, and report the compression metrics for the actually deployed architecture.","section":"Section III-A, Algorithms 1–2, Eq. (5), Table II"},{"comment":"Key hyperparameters are never reported. The number of training epochs n, the number of QAT epochs m, the number of pruning stages s, the stage-wise pruning schedule {p_i}, and the APoT parameters k and α are all essential for reproducibility, but Section IV only gives learning-rate schedules, batch size, and the global 30% pruning ratio. For example, Algorithm 2's pruning condition depends on n and s, and the quantization levels in Eq. (4) are determined by k and α. Please provide a complete hyperparameter table for every architecture and both methods.","section":"Section IV-A, Algorithms 1–2, Eq. (4)"},{"comment":"Several headline claims rest on accuracy differences that are within plausible run-to-run variation of a single CIFAR-10 training run. For example, PPQ on ResNet-110 is reported as 94.56% versus a 94.50% baseline, and PPQ versus SPQ on VGG-16 differs by 0.35 percentage points. No error bars, number of seeds, or statistical significance tests are reported. Since the central claim is that PPQ/SPQ match or exceed SOTA accuracy while achieving much larger compression, at least 3–5 independent runs with mean ± std (and, ideally, paired significance tests) are needed to support the comparison.","section":"Tables II–III, Fig. 8"},{"comment":"The pruning criterion is computed on full-precision weights even though the forward pass in SPQ uses 4-bit APoT-quantized weights. The paper explicitly says 'GM is computed based on full-precision weights' but does not justify why the geometric center of full-precision filters is the right redundancy signal for a low-bit quantized model. If filter redundancy changes after quantization, the pruning decisions—and therefore the reported accuracy/compression trade-off—may not generalize. Please provide an ablation comparing GM computed on full-precision weights against GM computed on quantized weights, or otherwise justify the mismatch.","section":"Section III-C, Fig. 4"}],"minor_comments":[{"comment":"The first author's name is spelled 'Sara Makenali' in the abstract and 'Sara Mekenali' in the author block; please unify.","section":"Author block / Abstract"},{"comment":"Typo: 'ReNet-32' should be 'ResNet-32'.","section":"Section IV, first paragraph"},{"comment":"Typo in the figure: 'Quatization' should be 'Quantization'.","section":"Fig. 1"},{"comment":"The loop header 'For epoch = 1 to m ⌈|D|/BS⌉' is malformed; presumably the quantization phase should iterate over epochs and, inside each epoch, over mini-batches.","section":"Algorithm 2, line 23"},{"comment":"The x-axis labels are garbled/unreadable in the submitted PDF; please replace with clean text.","section":"Fig. 8"},{"comment":"The APoT notation is introduced but the specific values of k, α, and n used in the 4-bit experiments are not stated; please add them in Section IV.","section":"Section III-B, Eq. (4)"}],"recommendation":"major_revision","confidential_remarks":"The main issue is reproducibility rather than novelty or correctness of the underlying ideas. The missing residual-connection handling is the most serious gap because it directly affects the validity of the headline compression numbers for all ResNet experiments. I would encourage the editor to ask for the hyperparameter table, the ResNet channel-alignment rule, and at least a small number of repeated runs before considering the paper for publication. The paper fits the journal's scope, and the integration result is useful if the details are supplied."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a clean engineering study combining two established techniques — FPGM filter pruning and APoT quantization — under two schedules (simultaneous SPQ, sequential PPQ). The reported accuracy/compression trade-offs are plausible and the comparison tables are broad. But the central result is not reproducible as written, and the reason is structural: neither Algorithm 1 nor 2, nor Section III-A, says what happens to ResNet residual/shortcut paths when every convolutional layer is pruned by 30%. Filter pruning changes output channel counts, and in a residual block those must match the shortcut at the addition. The paper states the usual rule — removing filters in a layer removes the corresponding input channels in the next layer — but never extends it to skip connections, downsampling shortcuts, or the last conv of a block. If filters are only zero-masked, the forward pass is dimensionally fine but the claimed ×15.8 model-size reduction does not follow from Eq. (5), since zeroed weights are still stored. If filters are physically removed, a coordinated channel-selection rule is needed that the paper never gives. The stress-test note lands.\n\nThat said, there are real strengths. The two integration schedules are clearly distinct, and the description of SPQ — pruning at the end of each epoch while updating pruned weights so they can revive — is easy to follow. The model-size and BOPs accounting is internally consistent, and comparing against quantization-only and pruning-only baselines is the right way to frame a combination paper. Credit is given to FPGM and APoT, and the authors’ survey citation [16] is contextual, not load-bearing.\n\nSmaller soft spots: no error bars; n, m, s, and APoT’s k and α are not reported; and the conclusion overclaims “PPQ consistently superior” when SPQ beats it on VGG-16. The pruning-schedule omission is the serious one.\n\nThis paper is for practitioners wanting a concrete ×15 recipe for shrinking CIFAR-scale CNNs. Right now the recipe is incomplete. It deserves a serious referee because the combination question is legitimate and the results are worth checking, but acceptance should require a major revision: specify the channel-alignment rule for residual networks (or explicitly state which layers are pruned), disclose all hyperparameters, and correct the conclusion.","headline":"Honest combination of two known techniques with plausible CIFAR-10 numbers, but the paper never explains how 30% per-layer pruning handles ResNet skip connections, so the headline 15x compression is not reproducible as written.","tokens_in":18790,"tokens_out":4259,"would_cite":false,"duration_ms":41331,"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":"Combining geometric-median pruning with APoT quantization compresses CIFAR-10 CNNs by roughly 15x while keeping accuracy within about two points of the full-precision baseline.","keywords":["model compression","filter pruning","geometric median","APoT quantization","power-of-two quantization","quantization-aware training","CIFAR-10","deep neural networks"],"falsifier":"Run both pipelines on CIFAR-10 with the same 30% pruning rate and the same 4-bit APoT settings, but replace the GM importance score with random filter selection or with smallest-L2-norm filters. If random pruning reproduces the reported accuracy and compression trade-offs, the geometric-median criterion is not load-bearing; if GM-selected pruning consistently beats the controls across ResNet and VGG, the criterion is doing the claimed work.","tokens_in":17892,"feed_emoji":"🧠","tokens_out":8772,"duration_ms":83527,"temperature":0.7,"pith_summary":"The paper sets out to show that two compression techniques usually applied separately—similarity-based filter pruning and low-bit quantization—work better when deliberately integrated. It proposes two integration recipes: SPQ, which prunes and quantizes together during every training epoch, and PPQ, which first prunes incrementally and only then applies 4-bit quantization-aware training. On CIFAR-10 ResNet-20/32/56/110 and VGG-16, the authors report roughly a 15x cut in model size and more than a 100x cut in bit-operations, with accuracy losses around 0.1 to 2.2 percentage points, and with PPQ slightly exceeding the full-precision baseline on ResNet-110. If these results hold, the recipes are practical ways to deploy accurate image classifiers on devices with limited memory and compute.","feed_headline":"Two pipelines shrink deep networks 15x with little accuracy loss","feed_subtitle":"Combining geometric-median pruning with 4-bit power-of-two quantization on CIFAR-10 models keeps accuracy within about two points.","key_machinery":"Geometric-median filter pruning: in each convolutional layer, the GM is the point minimizing total Euclidean distance to all filters; the filters nearest it are judged redundant and zero-masked, because their represented information is assumed to overlap with the remaining filters. Additive Powers-of-Two (APoT) quantization: quantization levels are sums of n powers of two, matching the bell-shaped weight and activation distribution; at 4-bit precision, MAC multiplications can be replaced by bit-shift operations. First and last layers are kept at 8-bit. The two integration schemes—SPQ's simultaneous application each epoch and PPQ's staged pruning followed by quantization-aware training—carry","core_discovery":"The central claim is that combining GM-based filter pruning with APoT quantization yields effective compression without the accuracy collapse usually feared from stacking two lossy operations. SPQ quantizes weights and activations to 4-bit APoT in every epoch and zero-masks the 30% of filters closest to each layer's geometric median at the end of each epoch, while continuing to update the pruned filters. PPQ trains a full-precision network, prunes in two incremental stages (30% total) using the same GM criterion, then performs quantization-aware training for about 50 epochs. The reported outcome is that PPQ is the more accurate pipeline on all four ResNets—best among compared methods on ResN","pith_inferences":["The paper pairs one redundancy-based pruning score with one low-bit quantizer, but the same integration recipe could be tested with other similarity-based pruning criteria and other non-uniform quantizers; the paper only demonstrates the GM-plus-APoT combination.","The SPQ-versus-PPQ accuracy pattern hints at a rule the authors do not state: sequential compression protects representation in deep residual networks with narrow bottlenecks, while simultaneous compression exploits redundancy in heavily overparameterized networks like VGG-16.","BOPs counts assume bit-width-proportional arithmetic cost; actual latency and energy gains on a given chip also depend on memory traffic and whether shift-add units are implemented, so the reported x115–126 reductions are not automatically the same as wall-clock speedups.","Uniform 30% pruning and uniform 4-bit precision were fixed across all layers; layer-wise per-sensitivity pruning rates and bit-widths, mentioned only as future work, are the natural next lever and could push compression beyond x15 at the same accuracy."],"forward_implications":["PPQ on ResNet-110 reaches 94.56% accuracy, 0.06 points above the full-precision baseline, while cutting model size by x15.94 and BOPs by x120.95.","On all tested ResNets, PPQ stays within 0.77 points of baseline on ResNet-20 and often beats pruning-only or quantization-only SOTA methods, making sequential integration the safer default for deep residual networks.","SPQ achieves the best accuracy among all compared methods on VGG-16 (93.81%, 0.24-point drop) at a lower training cost than PPQ, making simultaneous integration attractive for wide, overparameterized networks.","Because both weights and activations are APoT-quantized, inference can replace floating-point multiplications with bit shifts, so the savings go beyond model size.","Structured filter pruning also removes the corresponding input channels in the next layer, keeping the compressed model hardware-friendly."],"supporting_citations":[{"why":"Supplies the geometric-median filter pruning criterion that selects which filters to remove.","marker":"[20]"},{"why":"Supplies the APoT quantization method, including the bit-shift multiplication rationale.","marker":"[4]"},{"why":"Defines the BOPs metric and motivates structured pruning for hardware deployment.","marker":"[21]"},{"why":"Prior hardware-friendly combination of channel pruning and power-of-two quantization; the closest integration baseline to beat.","marker":"[17]"}],"fun_headline_variants":["Prune then quantize: best accuracy for 15x smaller DNNs","Combining filter pruning and 4-bit quantization cuts DNNs 15x","GM pruning + APoT quantization: 15x compression, accuracy intact","Two-step method prunes then quantizes neural nets for big savings"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that a filter's distance from the geometric median of full-precision weights still identifies redundant filters after weights and activations are quantized to 4-bit, even though the forward pass never sees the full-precision weights.","fun_headline_variants_meta":{"raw":{"variants":["Prune then quantize: best accuracy for 15x smaller DNNs","Combining filter pruning and 4-bit quantization cuts DNNs 15x","GM pruning + APoT quantization: 15x compression, accuracy intact","Two-step method prunes then quantizes neural nets for big savings"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001062,"raw_usage":{"total_tokens":4288,"prompt_tokens":739,"completion_tokens":3549,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":483,"completion_tokens_details":{"reasoning_tokens":3479}},"tokens_in":483,"tokens_out":3549,"duration_ms":25443,"temperature":1.0,"reasoning_tokens":3479,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T10:13:09.163823+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run both pipelines on CIFAR-10 with the same 30% pruning rate and the same 4-bit APoT settings, but replace the GM importance score with random filter selection or with smallest-L2-norm filters. If random pruning reproduces the reported accuracy and compression trade-offs, the geometric-median criterion is not load-bearing; if GM-selected pruning consistently beats the controls across ResNet and VGG, the criterion is doing the claimed work.","supporting_citations":[{"cited_title":"Filter pruning via geometric median for deep convolutional neural networks ac celeration,","cited_arxiv_id":null,"evidence_quote":"Supplies the geometric-median filter pruning criterion that selects which filters to remove."},{"cited_title":"Differentiable joi nt pruning and quantization for hardware efficiency,","cited_arxiv_id":null,"evidence_quote":"Defines the BOPs metric and motivates structured pruning for hardware deployment."},{"cited_title":"Hfpq: deep neural network com pression by hardware-friendly pruning-quantization,","cited_arxiv_id":null,"evidence_quote":"Prior hardware-friendly combination of channel pruning and power-of-two quantization; the closest integration baseline to beat."}],"review_version":1}