{"id":"cfa34a9e-14aa-4f25-a0bc-31e4a30a67cb","arxiv_id":"1909.01026","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Using depthwise convolution to expand channels yields networks with about 60 percent of MobileNetV2's parameters and comparable CIFAR accuracy.","lead":"The paper proposes replacing 1x1 convolutions with depthwise convolutions to expand channel counts in bottleneck blocks, creating two lighter network designs called PSDNet and DPDNet. It reports fewer parameters and less compute than ResNet and MobileNetV2 with roughly similar or better accuracy on small image datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DWC-expansion claim is confounded: replacing the expansion 1x1 with DWC also changes block information flow, so the CIFAR accuracy gains do not isolate the expansion mechanism.","rationale":"The reader's conditional verdict already captures the main risk: the accuracy comparison between DPDNet and MobileNetV2 may be confounded by block-structure differences beyond the DWC-versus-PWC expansion. My read agrees with that concern and sharpens it: even the seemingly closer PSDNet comparison confounds the expansion mechanism with a structural change in information flow (no cross-channel mixing in the DWC expansion, followed by residual addition). This is not an internal inconsistency or a fatal flaw; the parameter and FLOP arithmetic is correct, and the CIFAR results are plausible. However, because the paper's own ImageNet section is explicitly incomplete and the CIFAR accuracy gaps are often within single-run noise, the evidence does not support the strong, general conclusion that DWC expansion is more efficient than PWC expansion. The proposed ablation directly isolates the expansion operation while holding all other block structure fixed, which would settle the causal question. Since the reader already assigned CONDITIONAL, I do not move the verdict; the concern reinforces the need for the stated conditions (error bars, controlled ablations, completed ImageNet evaluation).","tokens_in":12409,"tokens_out":9005,"duration_ms":93177,"concrete_test":"On CIFAR-10 and CIFAR-100, take the DPDNet architecture (m=5, alpha=1) and construct a matched control that replaces only the first DWC(m>1) expansion layer in every block with a PWC (1x1, m>1) expansion, keeping the subsequent PWC and DWC layers, strides, and all training hyperparameters identical. Train both variants with the paper's exact 300-epoch schedule for at least 5 seeds, reporting mean +/- std accuracy, parameter count, and FLOPs. Run the analogous swap in the PSD block (last DWC replaced by a 1x1 expansion). If the PWC-expansion control matches or exceeds DWC accuracy within noise, the paper's central efficiency claim is not established; if DWC wins consistently despite fewer parameters, the claim is supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim rests on comparing blocks that differ by more than the expansion operation. In the PSD block, the final 1x1 expansion (C to mC) is replaced by DWC(m>1). The parameter/FLOP arithmetic in Eq. (1)-(2) is correct, but the two operations are not functionally equivalent: a 1x1 expansion mixes all C input channels into every output channel, whereas a DWC(m>1) output channel depends on exactly one input channel. After the residual addition in the PSD block, each output channel is therefore updated from a single input-derived spatial feature, not from a combination of channels. The measured accuracy is thus not a clean comparison of 'DWC expansion vs PWC expansion' at equal representational behavior. In DPDNet versus MobileNetV2, the blocks differ in layer ordering (DWC-PWC-DWC vs PWC-DWC-PWC), in which layer performs expansion, in the number of DWC versus PWC layers, and possibly in residual/linear-bottleneck details; the accuracy differences in Tables 5-6 are mostly under 1% (MobileNetV2 is actually better on CIFAR-100 at m=3-6), with no error bars or multiple seeds. The manuscript's own placeholders in the ImageNet section ('M parameters', 'Table x', an empty DPDNet row in Table 8) and the concluding admission that ImageNet evaluation is future work further limit the scope of the claim. Therefore, the causal statement that DWC channel expansion is more efficient than PWC channel expansion is not isolated by the current experiments.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes replacing the pointwise (1×1) convolution used for channel expansion in bottleneck blocks with a depthwise convolution (DWC) whose channel multiplier m is greater than 1. The authors derive parameter and FLOP ratios in Eqs. (1)–(2) and show that DWC expansion is cheaper than PWC expansion by a factor of C/k². Based on this idea, they introduce PSDNet, a ResNet variant that replaces the final 1×1 expansion layer with a DWC layer, and DPDNet, a lightweight network built from a Depthwise-Pointwise-Depthwise (DPD) inverted bottleneck block. Experiments on CIFAR-10, CIFAR-100, and CINIC-10 report parameter and FLOP reductions with comparable or better accuracy relative to ResNet and MobileNetV2, along with sweeps over the hyperparameters m and α. An ImageNet evaluation is described in the text but not actually reported.","tokens_in":12716,"tokens_out":5575,"duration_ms":56144,"significance":"The arithmetic in Eqs. (1) and (2) is correct, and the parameter and FLOP counts in Tables 4–6 are internally consistent, giving a concrete efficiency comparison for the proposed blocks. The two hyperparameters m and α provide a useful accuracy–cost trade-off, and the CINIC-10 results at matched parameter counts are a genuine empirical contribution. The central proposal—using DWC for channel expansion—is simple and plausibly effective at reducing parameter cost. However, the experiments do not isolate the proposed mechanism: the compared blocks differ in multiple structural ways, accuracy differences are often below 1% with no error bars, and the promised ImageNet results are absent. The significance of the claim that DWC expansion is both cheaper and more accurate than PWC expansion is therefore not yet established.","major_comments":[{"comment":"The parameter and FLOP comparison treats DWC(m>1) and PWC as equivalent channel-expansion operators, but they are not functionally equivalent. Each output channel of a DWC(m>1) layer is derived from a single input channel with m filters, whereas every PWC output channel mixes all C input channels. Replacing a PWC expansion layer with a DWC therefore changes the information flow of the block, not merely its parameter cost. Consequently, the accuracy differences between PSDNet and ResNet in Table 4 (0.9% on CIFAR-10, 0.5% on CIFAR-100) conflate the efficiency gain with a structural change. To support the central claim, the paper needs an ablation that holds the block structure as constant as possible and varies only the expansion operation, with results averaged over multiple seeds.","section":"Section 3.1, Eqs. (1)–(2)"},{"comment":"The DPDNet-versus-MobileNetV2 comparison changes several architectural factors simultaneously: layer order (DWC-PWC-DWC versus PWC-DWC-PWC), the layer that performs expansion, the number of DWC versus PWC layers, and stride handling. The reported accuracy differences are frequently below 1% and sometimes favor MobileNetV2 (e.g., CIFAR-100 for m=3–6 in Table 5; CIFAR-10 for α=2.5, 3.0, 4.0 in Table 6). Because the blocks differ in more than the expansion operation, these results do not isolate the effect of using DWC for expansion. The paper should provide matched comparisons in which only the expansion mechanism differs, or explicitly control for the other structural variables, and should report multiple runs with error bars.","section":"Section 3.3, Tables 5 and 6"},{"comment":"The ImageNet evaluation is incomplete: the text contains placeholders such as \"M parameters\" and \"Table x\", the DPDNet row in Table 8 is empty, and the conclusion states that evaluation on ImageNet is future work. The abstract and introduction claim that DPDNet is more efficient than MobileNetV2 without dataset qualification. The paper must either provide the ImageNet results or explicitly and prominently scope the claims to the CIFAR and CINIC-10 datasets; as written, the general efficiency claim is unsupported.","section":"Section 3.2.2 and Table 8"},{"comment":"All accuracy numbers are reported for a single run, with no error bars, multiple seeds, or statistical significance tests. Many differences are smaller than 1% (e.g., Table 5, CIFAR-10 m=3–5; Table 6, CIFAR-10 α=1.25–2.0), so without variance estimates the \"higher accuracy\" and \"approximate accuracy\" claims are not robustly supported. The authors should run each configuration at least three times and report mean and standard deviation.","section":"Tables 4–6"}],"minor_comments":[{"comment":"The section headings \"3.2. Results of PSDNet\" and \"3.3. Results of DPDNet\" should be renumbered to reflect the current section (e.g., 4.2 and 4.3), since Section 4 contains the datasets and results.","section":"Section 4.1"},{"comment":"Table numbering is duplicated: there are two tables labeled \"Table 3\" (DPD block versus DPDNet body architecture) and two labeled \"Table 4\" (ImageNet architecture versus PSDNet results). Renumber all tables sequentially.","section":"Tables 1–8"},{"comment":"\"ResNet50 [20]\" should cite the original ResNet architecture [8] rather than the identity-mapping paper [20], since the architecture used is the standard ResNet50.","section":"Section 3.2.1"},{"comment":"Several typographical errors should be corrected: \"concolutional\" and \"netwok\" in the abstract, \"PDP inversed bottleneck block\" in the conclusion, and \"NerurIPS\" in references [4] and [23] should be \"NeurIPS\".","section":"Abstract and Conclusion"},{"comment":"The subfigure labels (a)–(d) are not explicitly referenced in the caption text beyond the order of the displayed blocks; please add a sentence identifying each block, and clarify whether linear bottlenecks (no ReLU after the PWC layer) are used in the DPD block as they are in MobileNetV2.","section":"Figure 2"},{"comment":"The note \"The number of parameters is the model for CIFAR-10 and CINIC-10 datasets\" is ambiguous; for CIFAR-100 the FC layer has more output neurons, so the parameter count differs. Please report all three parameter counts explicitly or state the convolutional parameter count separately.","section":"Table 5 and 6"}],"recommendation":"major_revision","confidential_remarks":"The manuscript resembles an extended abstract or a work in progress: the ImageNet results are placeholder-filled, the table numbering is inconsistent, and the experimental design does not isolate the paper's central mechanism. These issues are fixable, but they are substantial enough that the paper is not currently suitable for acceptance. I would also note that the novelty over MobileNetV2/Xception is incremental—using DWC for expansion is a natural extension of depthwise separable convolutions—so the authors should strengthen the empirical isolation and significance testing to justify journal publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is an unpolished draft with a small but real parameter-efficiency result, and the authors oversell the mechanism. The central arithmetic is correct, and the CIFAR tables support the parameter savings. What's actually new is using depthwise convolution with channel multiplier m>1 in place of 1x1 pointwise convolution for channel expansion, in both a ResNet-style bottleneck (PSD block) and an inverted bottleneck (DPD block). That is a legitimate architecture variant, even if the idea is a reframing of known depthwise operations.\n\nThe paper does solid things. Eq (1)-(2) are correct: PWC parameters are C*mC, DWC are k*k*mC, ratio C/k^2. For C much larger than 9, DWC is cheaper. The parameter and FLOP tables for PSDNet vs compressed ResNet50 and DPDNet vs own-implemented MobileNetV2 are reproducible from the stated architecture tables. On CIFAR-10/100, DPDNet stays within about 1% of MobileNetV2 while using 60% of the parameters. That is worth knowing.\n\nSoft spots are in proportion. First, no error bars or multiple seeds; differences of 0.5% are noise-prone. Second, the DPDNet vs MobileNetV2 comparison changes block ordering and layer composition simultaneously, so the reported accuracy differences do not isolate \"DWC vs PWC expansion.\" The authors should do a controlled ablation: same block structure, swap only the expansion layer. Third, the ImageNet section is placeholder text: \"M parameters\", \"Table x\", and an empty DPDNet row. The conclusion that spatial features matter more than channel combination is an overgeneralization from small datasets and should be softened. Also, the manuscript has structural errors (two Table 3s, missing equation numbers in text).\n\nThe stress-test concern about information flow is valid: a DWC expansion channel depends on one input channel, so the block's representational behavior changes beyond parameter count. That doesn't make the efficiency claim false, but it means the accuracy comparison is not a clean test of the mechanism.\n\nWho it's for: people building lightweight CNNs for mobile or embedded deployment will find the parameter/FLOP trade-off useful. It deserves a serious referee for a workshop or a short paper, not a desk reject, because the arithmetic is honest and the small-dataset experiments support the parameter-efficiency claim. But it needs ImageNet numbers, multiple seeds, and a proper ablation before publication. My recommendation: treat as a borderline-major-revision paper; do not accept as is.","headline":"A rough draft with a real parameter-efficiency trick, no ImageNet results, and an overreaching mechanism claim; worth a serious look for the small but honest CIFAR evidence.","tokens_in":13289,"tokens_out":2209,"would_cite":false,"duration_ms":21682,"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 argues that depthwise convolution can replace 1x1 convolution for expanding channel counts in bottleneck blocks, yielding networks with fewer parameters and FLOPs and comparable or better accuracy.","keywords":["depthwise convolution","channel expansion","bottleneck block","inverted bottleneck","lightweight CNN","PSDNet","DPDNet","parameter efficiency"],"falsifier":"Measure wall-clock latency and energy on a representative mobile or edge accelerator for DPDNet and MobileNetV2 at matched parameter budgets: if DWC-based models are not faster despite fewer FLOPs, the central efficiency claim is false in practice. Alternatively, train MobileNetV2 with the DPD block's layer ordering but identical channel counts and training settings; if the accuracy gap vanishes, the paper's attribution to DWC expansion is not supported.","tokens_in":12238,"feed_emoji":"📉","tokens_out":4807,"duration_ms":39156,"temperature":0.7,"pith_summary":"This paper argues that the common practice of using 1x1 pointwise convolutions to expand channels in bottleneck blocks is less efficient than using depthwise convolutions (DWCs) for the same job. It derives a simple ratio showing that, for a kernel size k and C input channels, DWC beats PWC by a factor of C/$k^{2}$ in both parameters and computation. Based on this, it builds PSDNet, a ResNet-style network with DWC expansion, and DPDNet, an inverted-bottleneck network using a depthwise-pointwise-depthwise (DPD) block. On CIFAR-10 and CIFAR-100, PSDNet improves on the compressed ResNet50 baseline while cutting parameters from 2.0M to 1.6M, and DPDNet matches MobileNetV2 accuracy at roughly 60% of its parameters.","feed_headline":"Cheaper CNN blocks swap 1x1 expansion for depthwise layers","feed_subtitle":"Bottleneck blocks built with depthwise channel expansion match or beat ResNet and MobileNetV2 accuracy with fewer parameters.","key_machinery":"The load-bearing identity is the parameter and cost ratio C/$k^{2}$ between pointwise and depthwise channel expansion (Eq. 1 and Eq. 2): a depthwise convolution with kernel size k and channel multiplier m needs k*k*mC parameters and W*H*k*k*mC operations, versus C*mC for a 1x1 convolution, so the DWC wins whenever the channel count C exceeds $k^{2}$. The second machinery is the DPD block, an inverted bottleneck made of an expanding DWC (m>1), a channel-mixing PWC, and a per-channel DWC (m=1), which the paper stacks to form DPDNet and tunes with width multiplier $\\alpha$ and channel multiplier m.","core_discovery":"The central discovery is that depthwise convolution, normally used with channel multiplier m=1 so input and output channel counts match, can instead be used with m>1 to expand channels in a bottleneck block. The paper proves that for expanding a feature map from C to mC channels, a k x k DWC has C/$k^{2}$ times fewer parameters and FLOPs than a 1x1 PWC, since the PWC costs C*mC while the DWC costs k*k*mC. Networks built on DWC expansion, PSDNet and DPDNet, report fewer parameters, lower FLOPs, and accuracy equal to or better than their PWC-based counterparts. The authors further report that models with more DWC layers outperform models with more PWC layers, interpreting this as evidence that extracting spatial features matters more than combining channel information.","pith_inferences":["The paper's efficiency metric is theoretical FLOPs; depthwise convolutions are memory-bound on many real accelerators, so the reported parameter and FLOP savings may not translate to proportional latency gains without explicit hardware benchmarking.","The accuracy comparison with MobileNetV2 is confounded by different block structure and channel counts; a cleaner test would hold block order constant and vary only the expansion layer type.","The C/k^2 ratio suggests the advantage of DWC expansion grows with channel count, so the benefit should be largest in the last layers of deep networks; measuring layer-wise sensitivity would test this.","If the spatial-feature hypothesis holds, DWC expansion should show larger gains on spatially demanding tasks like object detection or segmentation than on classification."],"forward_implications":["PSDNet50 cuts parameters from 2.0M to 1.6M and raises CIFAR-10 accuracy from 92.95% to 93.87% versus the compressed ResNet50 baseline.","DPDNet uses about 60% of MobileNetV2's parameters at the same m and alpha, with accuracy within about a percentage point on CIFAR-10/100 across most settings.","Two hyperparameters, alpha and m, give a smooth accuracy-versus-cost trade-off, letting one network family cover a range of deployment budgets.","The DWC-first ordering in DPD blocks appears to outperform the PWC-first ordering of MobileNetV2 when parameter counts are matched, suggesting spatial feature extraction is more valuable per parameter than channel combination."],"supporting_citations":[{"why":"Supplies the inverted-bottleneck baseline architecture that DPDNet modifies and is compared against.","marker":"[22]"},{"why":"Defines the ResNet bottleneck block with two 1x1 convolutions that PSDNet replaces with a DWC expansion.","marker":"[20]"},{"why":"Introduces depthwise separable convolution, the decomposition into DWC and PWC that the paper's cost comparison relies on.","marker":"[18]"},{"why":"Provides prior evidence that depthwise separable convolutions can replace standard convolutions while maintaining accuracy.","marker":"[34]"},{"why":"Supplies the CINIC-10 dataset and the reported MobileNetV2 accuracy that DPDNet is compared against on that benchmark.","marker":"[47]"}],"fun_headline_variants":["Depthwise conv beats 1x1 for channel expansion","Bottleneck blocks with depthwise expansion cut parameters","PSDNet and DPDNet: depthwise channel expansion, fewer FLOPs","Depthwise expansion in bottlenecks: fewer params, similar accuracy","Swap 1x1 for depthwise in bottlenecks: cheaper and accurate"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes that parameter count and FLOPs are the right measures of efficiency and that the reported accuracy gaps are caused by the DWC-versus-PWC block ordering rather than by other architectural differences.","fun_headline_variants_meta":{"raw":{"variants":["Depthwise conv beats 1x1 for channel expansion","Bottleneck blocks with depthwise expansion cut parameters","PSDNet and DPDNet: depthwise channel expansion, fewer FLOPs","Depthwise expansion in bottlenecks: fewer params, similar accuracy","Swap 1x1 for depthwise in bottlenecks: cheaper and accurate"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000569,"raw_usage":{"total_tokens":2702,"prompt_tokens":960,"completion_tokens":1742,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":576,"completion_tokens_details":{"reasoning_tokens":1653}},"tokens_in":576,"tokens_out":1742,"duration_ms":12879,"temperature":1.0,"reasoning_tokens":1653,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:28:19.377401+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure wall-clock latency and energy on a representative mobile or edge accelerator for DPDNet and MobileNetV2 at matched parameter budgets: if DWC-based models are not faster despite fewer FLOPs, the central efficiency claim is false in practice. Alternatively, train MobileNetV2 with the DPD block's layer ordering but identical channel counts and training settings; if the accuracy gap vanishes, the paper's attribution to DWC expansion is not supported.","supporting_citations":[{"cited_title":"Sandler, A","cited_arxiv_id":null,"evidence_quote":"Supplies the inverted-bottleneck baseline architecture that DPDNet modifies and is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the ResNet bottleneck block with two 1x1 convolutions that PSDNet replaces with a DWC expansion."},{"cited_title":"Chollet, Xception: Deep learning with depthwise separable convolutions, in: IEEE conference on computer vision and pattern recognition (CVPR), 2017","cited_arxiv_id":null,"evidence_quote":"Provides prior evidence that depthwise separable convolutions can replace standard convolutions while maintaining accuracy."}],"review_version":1}