{"id":"25b102d1-e33f-480b-82db-271c800f3ad6","arxiv_id":"2502.01303","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A conv-attention hybrid block (PATConv) plus an adaptive channel-split scheme (DPConv) yields the PartialNet models, which report accuracy and throughput gains over FasterNet on ImageNet-1K and COCO.","lead":"The paper introduces a family of image-recognition networks that split each layer's channels into two parts, running cheap attention on one part and convolution on the other, and reports higher accuracy with fewer FLOPs than FasterNet on ImageNet and COCO. It matters because it offers a practical way to make vision models faster and more accurate on existing hardware.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported ImageNet superiority may be a training-recipe artifact: same-recipe reproduction shrinks the FasterNet gap to 0.1–0.3 points at M/L scale.","rationale":"After reading in good faith, the most load-bearing concern is not the DPConv equation error (Eq. 7), which affects an auxiliary component not used in the final models, nor the fixed split ratio, which is an engineering choice. It is the validity of the headline empirical comparisons. The architecture-level evidence (Tab. 3, 4, 6) is internally consistent and supports PATConv as an efficient block, but the paper's strongest claim is a network-level superiority statement quantified against external models. That statement is only as strong as the least controlled comparison. The paper itself provides evidence of confounding: its reproduced FasterNet numbers (§Tab. 13) differ from the original Table 1 values, and including those reproductions reduces the claimed gains, especially at M and L where the gap is 0.1–0.3. For a 300-epoch ImageNet training, such differences are not reliable without confidence intervals. The other baselines are not reproduced at all. The concrete test would resolve this by adding matched-recipe runs or error bars. I therefore keep the reader's CONDITIONAL status: the paper should be accepted only after the comparisons are made fair, either by reproducing all baselines or by reporting uncertainty and limiting the claim to the scales where the advantage is robust.","tokens_in":21673,"tokens_out":8134,"duration_ms":71047,"concrete_test":"Run 5 independent ImageNet-1K trainings of PartialNet-M and FasterNet-M (or, if compute is limited, use the Tab. 13 reproduced FasterNet-M* values) under the exact Tab. 8 recipe and compare mean±std top-1 accuracy. If the 0.1-point gap is within one standard deviation, the central 'superior accuracy' claim must be qualified to small variants or abandoned at medium/large scale.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim depends on Table 1 showing PartialNet beats SOTA models in top-1 accuracy. That table mixes baselines from heterogeneous training recipes. The supplementary (§7) concedes some baselines use different augmentation/regularization, and Tab. 13 reproduces only FasterNet under the authors' exact recipe. Using those reproduced numbers, the PartialNet advantage over FasterNet drops from 1.3 to 1.0 points at T2, and at M and L it becomes 0.1 and 0.3 points (83.1 vs 83.0 and 83.9 vs 83.6, respectively). These margins are within typical ImageNet run-to-run variance, so 'superior accuracy' is not established at larger scales. The remaining baselines (Swin, ConvNeXt, PoolFormer, MobileViTv2) are taken from their original papers, so the comparison is not matched. If the same-recipe reproduction had been used throughout, the headline conclusion would be weaker or reversed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a partial channel mechanism (PCM) that splits feature-map channels and applies different operations to each part. On this basis it introduces Partial Attention Convolution (PATConv), three concrete blocks (PAT_ch, PAT_sp, PAT_sf), a Dynamic Partial Convolution (DPConv) intended to learn split ratios, and a PartialNet family. The central claims are that PATConv can replace both regular convolution and regular visual attention, and that PartialNet achieves superior top-1 accuracy and inference speed compared with state-of-the-art efficient models on ImageNet-1K, with strong transfer to COCO detection and segmentation. Experiments include classification tables, COCO results, ablations on attention type and convolution type, and a supplementary same-recipe reproduction of FasterNet.","tokens_in":21889,"tokens_out":8098,"duration_ms":72032,"significance":"If the main claims held, the partial-channel combination of convolution and attention would be a practically valuable efficient operator, because it offers a simple, hardware-friendly way to reduce FLOPs and parameters while preserving accuracy. The paper has real strengths: it provides a code link, extensive ImageNet and COCO results, a same-recipe reproduction of FasterNet in the supplement (Tab. 13), and ablation studies that isolate the contribution of each PAT block. However, the current evidence does not establish the headline 'superior accuracy' claim at the larger model scales, and the DPConv derivation contains demonstrable errors. The proposed mechanism may still be a useful engineering contribution, but the manuscript needs substantial correction and re-benchmarking before the central claims are supported.","major_comments":[{"comment":"Table 1's headline comparison mixes baselines trained under different recipes, as the supplement itself concedes ('For other models without ADT, such as ShuffleNetV2, MobileNetV2, and GhostNet, although the comparison is not entirely fair, we include them for reference'). Table 13 provides a same-recipe reproduction of FasterNet; comparing against those reproduced numbers, the PartialNet margins become 2.9/1.6/1.0/0.6/0.1/0.3 points for T0/T1/T2/S/M/L. At M and L the advantage is only 0.1 and 0.3 points, which is within typical ImageNet run-to-run variance. The abstract's claim of 'superior top-1 accuracy' is therefore not established at the larger scales. Please provide same-recipe numbers for all compared backbones, or restrict the claim to the scales where the margin is robust.","section":"Supplementary §7, Tab. 13 vs Tab. 1"},{"comment":"The paper states that Eq. (7) counts the number of non-zero elements in the mask U of the l-th DPConv layer. But the expression (row-sum) times (column-sum) is not that count in general. For example, with K=2 and g=(0,0), U is the 4x4 identity matrix, which has four non-zero entries, whereas the formula gives 1. The regularization term ζ and the Pareto objective in Eq. (9) therefore do not measure the stated computational complexity. Please rederive Eq. (7) and the associated κ/ζ constraint.","section":"§3.2, Eq. (7)"},{"comment":"The mask definition in Eq. (5) is not well-formed. Since r_p is a ratio, 1/r_p is not a channel index; as written, the condition i < 1/r_p would select only the first three channels when r_p=1/4, independent of c_in. The connection between the Kronecker-gate representation and the actual channel mask needs a clean formal statement before the adaptive split ratio can be evaluated.","section":"§3.2, Eq. (5)"},{"comment":"The abstract and Section 3.2 present DPConv as a mechanism that 'adaptively learn[s] the proportion of split channels', but Tab. 9 fixes r_p=1/4 for every PAT block in every PartialNet variant, and no reported model in Tabs. 1-2 is trained with learned ratios. Fig. 7 shows learned ratios only for T0 and does not report top-1 accuracy for those configurations. Moreover, the constraint target κ uses θ=4, explicitly chosen to match FasterNet's r_p=1/4 complexity, so the regularizer is enforcing a hand-picked target rather than deriving the split from an independent principle. Either train and report final models with learned r_p, or present DPConv as a separate analysis tool rather than as a component of the evaluated models.","section":"§3.2 vs Tab. 9"}],"minor_comments":[{"comment":"The notation F^{c_in × r_p} is not a standard way to denote a channel subset; please use a mask or channel-range notation so the split is unambiguous.","section":"§3.1, Eq. (1)"},{"comment":"The symbol U denotes both the binary mask matrix in DPConv and the concatenation operation in the PAT block figures; please rename one of them to avoid confusion.","section":"Fig. 3 and Fig. 4"},{"comment":"In the FasterNet-S rows, one row appears to be a reproduced result but does not carry the '*' marker; please clarify which entries are original and which are reproduced under the authors' setup.","section":"Supplementary Tab. 13"},{"comment":"The text says the authors use 'the same regularization and augmentation techniques as FasterNet', but Tab. 8 shows different mixup, cutmix, and stochastic-depth values across variants; please state that the training recipe is of the same family rather than identical.","section":"§4.1 and Supplementary Tab. 8"},{"comment":"The summation index i in Eq. (8) is inconsistent with the gate subscript k used in the rest of the equation; please align the notation.","section":"§3.2, Eq. (8)"},{"comment":"The θ sweep reports only learned ratios and not the resulting top-1 accuracies; including accuracy values would let the reader judge whether the adapted ratios actually improve the trade-off.","section":"Fig. 7"}],"recommendation":"major_revision","confidential_remarks":"The paper is an architecture-engineering contribution whose value depends on fair benchmarking and on a correct formalization of the learned split mechanism. The same-recipe reproduction in Tab. 13 substantially weakens the 'superior accuracy' claim at M/L scales, and the DPConv equations contain errors that prevent the adaptive mechanism from being assessed as presented. I would ask the editor to weigh whether the authors can fix these issues with additional experiments; if the same-recipe baselines cannot be extended to all competitors, the claims will need to be narrowed considerably."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe takeaway: this is a legitimate engineering paper with a transferable module, but its headline \"superior accuracy vs SOTA\" does not survive a matched-recipe check. At M/L scale the advantage over FasterNet shrinks to 0.1–0.3 points, which is noise.\n\nWhat is actually new: PATConv splits the feature channels and runs convolution on one part and attention on the other. That specific combination is not in FasterNet or ShuffleNetV2, and the three attention variants (channel, spatial, self) are reasonable designs. The ablation work is the strongest part. Table 3 shows partial attention matching or beating full attention while cutting FLOPs; Table 5 shows the module transferred to ResNet50, MobileNetV2, and ConvNeXt-tiny with gains of 1.5–2.6 points. That is real evidence the mechanism does something.\n\nSoft spots, in order. First, the DPConv derivation is algebraically wrong: Eq 7 equates the nonzero count to the product of row and column sums, which undercounts by a factor of 2^K when gates are zero, and Eq 5's mask formula is malformed. Second, the paper claims adaptive split ratios but the final architecture fixes rp=1/4 everywhere and the reported models do not use DPConv. The Fig 7 experiments are exploratory, not load-bearing. Third, the main comparison table mixes training recipes. The supplementary reproduces FasterNet under the authors' recipe, which is good, but those same-recipe numbers reduce the advantage to 1.0 points at T2, 0.6 at S, 0.1 at M, and 0.3 at L. The large-scale margins are within typical ImageNet variance. To the paper's credit, §7 admits that some baselines are not trained with the same augmentation, but the abstract's \"superior to SOTA\" claim is stronger than the evidence.\n\nWho this is for: researchers building efficient backbones or hybrid conv-attention blocks. The PATConv module is worth trying; the DPConv regularizer needs a corrected derivation and a demonstration on the actual models before it is credible.\n\nRecommendation: send to peer review, but require matched training recipes for all baselines, a corrected DPConv derivation, and scaled-back accuracy claims at larger model sizes.","headline":"Useful partial-conv-plus-attention module with transferable gains, but headline accuracy claims rest on mixed recipes and a wrong complexity formula.","tokens_in":22429,"tokens_out":5172,"would_cite":true,"duration_ms":42160,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single module that splits feature channels into a convolution branch and an attention branch can replace both regular convolution and regular visual attention while cutting parameters and FLOPs.","keywords":["partial channel mechanism","partial attention convolution","efficient neural networks","visual attention","ImageNet-1K classification","COCO detection and segmentation","dynamic partial convolution","hybrid vision backbone"],"falsifier":"Train PartialNet-T2 twice under the paper's ImageNet-1K settings: once with the attention branch inside PAT_ch and once with that branch replaced by a second 3x3 convolution of matched parameter cost. If top-1 accuracy stays within the noise level, the reported gains come from the extra parameters and channel split, not from partial attention; if accuracy drops clearly, partial attention is doing the work.","tokens_in":21456,"feed_emoji":"⚡","tokens_out":6473,"duration_ms":54544,"temperature":0.7,"pith_summary":"The paper proposes a partial channel mechanism (PCM): split a feature map into channel groups and run a different operation on each group in parallel. Its main module, PATConv, sends one fraction of channels through a 3x3 convolution and the other fraction through a visual attention block, so attention is computed only on a subset of channels. The claim is that this single module can replace both regular convolution and regular visual attention, reducing parameters and FLOPs while improving or matching accuracy. Building on it, PartialNet gives a family of hybrid backbones that the authors report outperform efficient baselines on ImageNet-1K and COCO. If correct, the result means a cheap channel-split design, not full attention, is enough to capture the global information that attention is supposed to provide.","feed_headline":"Split channels let one block replace convolution and attention","feed_subtitle":"The PartialNet recipe stays accurate at half the FLOPs of FasterNet, its closest partial-convolution baseline.","key_machinery":"The central object is the partial channel mechanism: a split operation divides the input feature map into two channel groups, each handled by a different parallel operator, followed by concatenation. The specific engine is PATConv, whose novelty is running convolution and attention side by side on disjoint channel subsets rather than serially. Three instantiations carry the accuracy gains: PAT_ch pairs a 3x3 convolution with an enhanced Gaussian channel-attention (mean and standard deviation squeeze), PAT_sp pairs a 1x1 convolution with a spatial-attention map and can be folded into adjacent MLP layers, and PAT_sf pairs a 3x3 convolution with self-attention with relative position encoding in the last stage. DPConv supplies the learnable split ratio through a Kronecker product of 2x2 binary gates trained with a straight-through estimator.","core_discovery":"The paper's central claim is that a single module, PATConv, can completely replace both a regular convolution and a regular visual attention layer while using fewer parameters and FLOPs. Formally, with input $F \\in \\mathbb{R}^{h \\times w \\times c_{\\text{in}}}$ and output $O$, PATConv computes $O = \\mathrm{Conv}(F_{c_{\\text{in}} r_p}) \\cup \\mathrm{Atten}(F_{c_{\\text{in}}(1-r_p)})$: a convolution branch processes the fraction $r_p$ of channels and an attention branch processes the rest in parallel, then the two outputs are concatenated. The paper develops three attention variants for that branch—channel attention (PAT_ch), spatial attention (PAT_sp), and self-attention (PAT_sf)—and a learnable dynamic partial convolution that adapts $r_p$ layer by layer. The reported experiments show the resulting PartialNet family beating FasterNet at lower FLOPs on ImageNet-1K and improving COCO detection and segmentation over FasterNet backbones.","pith_inferences":["The paper leaves the attention-versus-split attribution open; a diagnostic that would separate the two effects is to replace the attention branch with a second convolution of the same cost and compare top-1 accuracy.","The learned split-ratio pattern—first and last layers stay denser while middle layers become sparse—mirrors results from quantization studies; the same ordering could be tested as a prior for pruning or mixed-precision allocation.","Since PATConv is defined for any attention module, the three PAT blocks are only a sample of the recipe; applying the identical split-and-run-in-parallel pattern to other modalities, such as video or point cloud backbones, is a direct transfer test."],"forward_implications":["A network built from PATConv no longer needs a separate attention module: the three PAT blocks cover channel, spatial, and self-attention, so the same module set serves both local and global interaction.","Because PATConv is a convolution replacement, it can be inserted into existing CNN backbones; the paper's transfer experiments report accuracy gains on ResNet50, MobileNetV2, and ConvNeXt-tiny with higher throughput.","The adaptive DPConv ratio means the split between cheap convolution and expensive attention can be tuned per layer to meet a specified complexity budget, avoiding manual search over $r_p$.","At matched model sizes, the reported ImageNet-1K numbers imply the accuracy-FLOPs frontier is shifted: PartialNet-T2 reaches 80.2 percent top-1 at 1.03 G FLOPs, above FasterNet-T2's 78.9 percent at 1.91 G FLOPs."],"supporting_citations":[{"why":"Provides the partial convolution (PConv) baseline, the FasterNet architecture that PartialNet compares against, and the training setup reused for fair comparison.","marker":"[3]"},{"why":"Introduces the channel split operation that leaves part of the feature map untouched, the direct precursor to the partial channel mechanism.","marker":"[20]"},{"why":"Establishes redundancy among feature map channels, the premise for operating on a subset of channels instead of all of them.","marker":"[8]"},{"why":"Supplies the MetaFormer-style self-attention paradigm used in the PAT_sf block and serves as a hybrid baseline on ImageNet-1K.","marker":"[38]"},{"why":"Provides the channel-attention mechanism (SE-Net) that PAT_ch enhances with Gaussian statistics by adding standard deviation to the mean squeeze.","marker":"[14]"},{"why":"Supplies the Kronecker-product binary gate mechanism that DPConv adapts to learn per-layer split ratios.","marker":"[39]"},{"why":"Represents an efficient linear-attention hybrid backbone whose approach of combining cheap operations with attention is extended by PATConv.","marker":"[2]"},{"why":"Offers a separable self-attention hybrid baseline that also blends convolution and attention, used for comparison in the ImageNet-1K tables.","marker":"[22]"}],"fun_headline_variants":["Partial channels let one module replace conv and attention","Split channels: cut FLOPs, keep accuracy, beat FasterNet","PATConv: convolution and attention in a single block","PartialNet halves FLOPs, replaces conv and attention","One split channel trick: fewer params, better performance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that attention applied to only a fraction of the channels captures most of the benefit of full attention, so the remaining channels do not need their own attention.","fun_headline_variants_meta":{"raw":{"variants":["Partial channels let one module replace conv and attention","Split channels: cut FLOPs, keep accuracy, beat FasterNet","PATConv: convolution and attention in a single block","PartialNet halves FLOPs, replaces conv and attention","One split channel trick: fewer params, better performance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0002,"raw_usage":{"total_tokens":1416,"prompt_tokens":1024,"completion_tokens":392,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":640,"completion_tokens_details":{"reasoning_tokens":312}},"tokens_in":640,"tokens_out":392,"duration_ms":4162,"temperature":1.0,"reasoning_tokens":312,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T15:44:37.527537+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train PartialNet-T2 twice under the paper's ImageNet-1K settings: once with the attention branch inside PAT_ch and once with that branch replaced by a second 3x3 convolution of matched parameter cost. If top-1 accuracy stays within the noise level, the reported gains come from the extra parameters and channel split, not from partial attention; if accuracy drops clearly, partial attention is doing the work.","supporting_citations":[{"cited_title":"Run, don’t walk: Chasing higher flops for faster neural networks","cited_arxiv_id":null,"evidence_quote":"Provides the partial convolution (PConv) baseline, the FasterNet architecture that PartialNet compares against, and the training setup reused for fair comparison."},{"cited_title":"Shufflenet v2: Practical guidelines for efficient cnn architec- ture design","cited_arxiv_id":null,"evidence_quote":"Introduces the channel split operation that leaves part of the feature map untouched, the direct precursor to the partial channel mechanism."},{"cited_title":"Ghostnet: More features from cheap operations","cited_arxiv_id":null,"evidence_quote":"Establishes redundancy among feature map channels, the premise for operating on a subset of channels instead of all of them."},{"cited_title":"Metaformer is actually what you need for vision","cited_arxiv_id":null,"evidence_quote":"Supplies the MetaFormer-style self-attention paradigm used in the PAT_sf block and serves as a hybrid baseline on ImageNet-1K."},{"cited_title":"Squeeze-and-excitation net- works","cited_arxiv_id":null,"evidence_quote":"Provides the channel-attention mechanism (SE-Net) that PAT_ch enhances with Gaussian statistics by adding standard deviation to the mean squeeze."},{"cited_title":"Differ- entiable learning-to-group channels via groupable convolu- tional neural networks","cited_arxiv_id":null,"evidence_quote":"Supplies the Kronecker-product binary gate mechanism that DPConv adapts to learn per-layer split ratios."},{"cited_title":"Efficientvit: Lightweight multi-scale attention for high- resolution dense prediction","cited_arxiv_id":null,"evidence_quote":"Represents an efficient linear-attention hybrid backbone whose approach of combining cheap operations with attention is extended by PATConv."}],"review_version":1}