{"id":"3d37deb1-6eb8-492b-a623-4abd434fd43a","arxiv_id":"1908.05867","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A Kronecker-product binary mask makes the number of groups in each convolution layer learnable end-to-end, giving 0.5 to 1.7 percent top-1 ImageNet gains over ResNet and ResNeXt at comparable cost.","lead":"This paper introduces a convolution operation that learns its own channel-group structure during training, instead of fixing the number of groups by hand. Read it to see a straightforward knob that trades computation for accuracy automatically inside standard convolutional networks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing random-baseline control: learned per-layer group counts are only compared to constant 32×4d, so the gain could stem from non-uniformity rather than from DGConv's differentiable learning.","rationale":"The reader's weakest assumption concerns whether the straight-through estimator (STE) guides the binary gates to useful structures. This is a mechanism-level concern, and the paper partially addresses it: Table 6 shows that the final learned group numbers, when transferred to a standard ResNeXt and retrained from scratch, reproduce the accuracy of G-ResNeXt. Even if STE is imperfect, the transferability of the final numbers suggests they are meaningful. The more fundamental gap is that the paper never compares the learned per-layer group numbers to random or hand-crafted non-uniform schedules. The treatment differs from the constant baseline in two ways: the group numbers are non-uniform, and their specific values are learned. The design cannot separate these two effects. If random non-uniform schedules achieve the same accuracy, then the improvement is due to the search space (non-uniform cardinalities) rather than to DGConv's optimization, undermining the paper's core claim of 'differentiable learning-to-group channels.' This is a standard control in architecture search (random architecture baseline) and is directly testable. The paper's Figure 5 shows a simple monotonic trend, and Table 5 indicates a flat optimum for the learned schedules, both of which make the missing control more pressing. Since the empirical results are otherwise well-controlled, the appropriate verdict remains CONDITIONAL, but the required condition should include this random-baseline experiment rather than only STE analysis or error bars.","tokens_in":13406,"tokens_out":21612,"duration_ms":215060,"concrete_test":"Train three variants of ResNeXt101 with non-uniform per-layer group numbers at the same total GConv FLOPs as G-ResNeXt101 (b=32): (1) a random permutation of the learned group numbers across layers, (2) a random schedule sampled from a distribution matched to the learned cardinality histogram, and (3) a monotonic schedule (e.g., groups increasing with depth) designed to mimic the trend in Figure 5. Compare top-1 accuracy to ResNeXt* learned (79.8 in Table 6). If any variant reaches within 0.2% of 79.8, the learned schedule is not necessary for the gain; if all variants are at least 0.5% worse, the learning matters.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central empirical claim is that the cardinalities learned by DGConv are a genuine improvement over a fixed cardinality. The supporting evidence is Table 6, which transfers the learned group numbers into a standard ResNeXt and retrains from scratch. This demonstrates that the learned numbers can be reused and that the mask pattern is not the source of the gain. However, it does not establish that the learned numbers are better than other non-uniform assignments. The only comparison is to ResNeXt 32×4d, a constant-cardinality baseline. If any non-uniform schedule with the same total complexity (e.g., a random assignment or a simple heuristic like monotonically increasing group numbers) achieves similar accuracy, then the improvement is caused by relaxing the uniform constraint, not by the differentiable optimization. Note that Figure 5 shows lower layers consistently use fewer groups, which is a simple, hand-craftable pattern; and Table 5 shows different runs yield different schedules with nearly identical accuracy, suggesting a flat optimum. Without a random or heuristic non-uniform control, the reported gain cannot be attributed to DGConv's learning mechanism.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a dynamic grouping convolution (DGConv) that learns per-layer group cardinalities end-to-end. DGConv parameterizes a binary channel-connection mask as a Kronecker product of 2x2 blocks controlled by a small set of continuous gates, uses a straight-through estimator for the sign function, and adds a complexity regularizer. The authors build Groupable-ResNeXt by replacing group convolutions in ResNeXt with DGConv and report ImageNet results: G-ResNeXt50/101 outperform their manual ResNet/ResNeXt counterparts at similar or lower complexity, learned group assignments transfer to standard ResNeXt by retraining from scratch (Table 6), and three runs with different seeds produce comparable accuracy (Table 5).","tokens_in":13670,"tokens_out":4271,"duration_ms":43847,"significance":"If the central claim holds, the paper makes a useful contribution: it replaces a manually fixed hyperparameter (group count) with a cheap differentiable mechanism, shows that the learned per-layer cardinalities transfer to standard group convolution, and demonstrates a favorable complexity-accuracy trade-off. Strengths include the explicit transfer experiment in Table 6, the reproducibility runs in Table 5, the small parameter overhead (log of the channel count), and the systematic sweep over complexity budgets in Table 3. The main weakness is that the experimental design does not yet isolate whether the gain comes from the differentiable learning mechanism or simply from allowing non-uniform group counts; this needs a non-uniform control before the central attribution can be accepted.","major_comments":[{"comment":"The central claim that DGConv's differentiable learning discovers better group numbers is not fully supported, because every comparison is against ResNeXt with a constant cardinality (32x4d). The authors never test a non-uniform group schedule with the same total FLOPs that is not produced by the differentiable optimization, such as a random assignment of group counts or a hand-designed depth-increasing schedule. Figure 5 indicates that the learned pattern is simple (lower layers use fewer groups), so a hand-craftable or random non-uniform baseline could plausibly match the reported accuracy. Please add at least one such control, trained from scratch with the same complexity budget, and report whether the learned schedules outperform it. Without this control, the improvement could be attributed to relaxing the uniform-cardinality constraint rather than to the learning mechanism.","section":"Section 4, Tables 2, 5, and 6"},{"comment":"The G-ResNeXt101 rows for b=96 and b=128 both report 0.22x GConv FLOPs. This is internally inconsistent with the stated budget mapping, where a larger b should correspond to a smaller complexity budget, and it undercuts the complexity-accuracy trade-off conclusions drawn from the table. Please correct the values or explain why two different budgets produce identical FLOPs; also verify the surrounding text statement that b=64 corresponds to 0.5x complexity, which the table lists as 0.47x.","section":"Table 3"},{"comment":"The optimization of the binary gates relies on a straight-through estimator, but the paper does not analyze how this estimator behaves on the Kronecker-product mask space. The cited reference [5] addresses activation quantization, which is a different optimization setting. Because Figure 6 shows the gates saturating during training, the learned schedules could be influenced by STE dynamics rather than by the loss landscape. Please provide supporting evidence that the differentiable optimization is what drives the gains, for example an ablation with fixed random masks at matched complexity or a comparison with a continuous relaxation of the gates. This directly affects the paper's claim that group counts are \"learned\" in a meaningful sense.","section":"Section 3.2, Eq. (6)"}],"minor_comments":[{"comment":"The text says G-ResNeXt50(b=32) and G-ResNeXt101(b=32) correspond to Fig. 5(d) and Fig. 5(a), but the Fig. 5 caption assigns (c) to G-ResNeXt50 b=32 and (d) to G-ResNeXt50 b=96. Please correct the panel reference.","section":"Section 4, Figure 5 cross-reference"},{"comment":"All reported accuracies are single runs except the reproducibility study in Table 5, and no error bars or standard deviations are given for the key comparisons. Please report at least mean values over multiple seeds for the main baselines and the proposed models, or state explicitly that the reproducibility study is the only multi-seed evidence.","section":"Section 4 and Tables 2-5"},{"comment":"Equation (9) writes the total complexity as a sum over ζ_k, while Eq. (8) defines layer-wise terms as ζ_ℓ. Please unify the notation to avoid confusion.","section":"Eq. (9)"},{"comment":"There are several typos and reference-format issues, including \"signle\" in Section 4, \"The Univesity of Hong Kong\" in the author affiliation, and the incomplete citation \"Yin Penghang et al.\" for reference [5]. Please proofread and complete the references.","section":"Throughout"},{"comment":"The statement that a=0 when ζ≤o and a=α otherwise, with α=-0.02, makes the penalty term [o/ζ]^a discontinuous in the exponent; the paper should clarify how this is implemented in the autodiff computation graph and whether the discontinuity affects training.","section":"Section 3.3"}],"recommendation":"major_revision","confidential_remarks":"The paper makes a strong \"first\" claim in the introduction about differentiable group-number learning; the authors should double-check that CondenseNet, FLGC, and recent differentiable NAS works do not already cover this capability. The missing non-uniform control is the main substantive issue; if the authors can show that a random or hand-designed non-uniform schedule does not match the learned one, the central claim would be much stronger. The duplicate FLOPs row in Table 3 also needs correction before the paper is publishable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper does something genuinely new: it makes the per-layer group count of a group convolution learnable in an end-to-end fashion, using a Kronecker product of 2×2 binary gates that costs only O(log C) parameters per layer. That is a clean and useful idea, and the authors do more than just report a better number. The Table 6 experiment, where the learned cardinalities are transferred into a standard ResNeXt and retrained from scratch, is the right control: it shows the gain comes from the learned group counts, not from the specific mask pattern of DGConv. That is real evidence.\n\nThe empirical work is generally solid. On ImageNet, the gains over re-implemented ResNet and ResNeXt baselines are consistent across 50- and 101-layer networks and across multiple complexity budgets. The re-implemented baselines land near the published numbers, which adds credibility. Table 5 shows that different random seeds give similar accuracy with slightly different group schedules, suggesting the method is not riding on a single lucky initialization.\n\nThe soft spot is one the stress-test note identified: there is no non-uniform but non-learned control. The learned schedules consistently give lower layers fewer groups (more computation), and Table 5 shows fairly flat optima. It is entirely possible that a simple hand-designed schedule, or even a random non-uniform assignment with the same total FLOPs, would recover most of the gain. The paper compares only to the constant 32×4d baseline, so the specific claim that the differentiable optimization is what helps is not yet proven. That is the main scientific gap.\n\nOther issues are minor and fixable: no error bars on the main comparisons, a duplicated 0.22× row in Table 3, a confusing cross-reference to Figure 5 in the text, and a hand-wavy citation for the straight-through estimator instead of any analysis of how well STE works on this mask space. No code is released, which makes replication harder but not impossible given the detail in the paper.\n\nNone of these problems are fatal. The central result—that learned per-layer cardinalities beat a fixed cardinality at matched or lower FLOPs—is supported by the transfer experiment and the reproducibility runs. This paper deserves a serious referee. I would ask the authors to add a random or heuristic non-uniform control, run each setting a few times to get error bars, and clean up the presentation.\n\nI would bring it to a reading group: the method is simple to explain, the transfer experiment is a good example of how to test architecture search results, and the missing control is a useful discussion point.","headline":"Learned per-layer group counts transfer to standard ResNeXt and beat the constant 32×4d baseline; a missing random non-uniform control keeps the mechanism claim conditional.","tokens_in":14161,"tokens_out":1405,"would_cite":true,"duration_ms":15012,"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":"The paper shows that convolutional networks can learn per-layer channel-group counts, and that doing so beats fixed-group ResNeXt at matched or lower cost.","keywords":["group convolution","dynamic grouping convolution","learned cardinality","ResNeXt","straight-through estimator","Kronecker product","resource-constrained training","ImageNet"],"falsifier":"One decisive test: train two Groupable ResNeXts of the same budget with all gates initialized positive versus all negative. If their learned per-layer group counts differ substantially while validation accuracy matches, the grouping strategy is determined more by optimization path than by the data, and the claimed gain is not attributable to the learned structure. A simpler check is to replace the learned group counts in a trained model with a constant cardinality of matched FLOPs, retrain from scratch, and see whether accuracy falls.","tokens_in":13234,"feed_emoji":"🧩","tokens_out":8002,"duration_ms":72646,"temperature":0.7,"pith_summary":"Group convolution is a standard way to cut cost in ConvNets, but the number of groups is usually a hand-set constant that is the same in every layer. This paper argues that constant cardinality is sub-optimal and proposes Dynamic Grouping Convolution (DGConv), which learns per-layer grouping strategies end-to-end. The authors build Groupable ResNeXt by replacing every group convolution with DGConv, and report that it outperforms re-implemented ResNet and ResNeXt on ImageNet with equal or fewer parameters and fewer group-convolution FLOPs. The learned per-layer group counts also transfer to ordinary group convolution, so the final model needs no special inference code. If correct, the result is a new, cheaply optimizable architectural dimension: where and how to group channels.","feed_headline":"GroupNet learns group counts per layer and outperforms ResNeXt","feed_subtitle":"On ImageNet, G-ResNeXt101 reaches 79.9% top-1 with fewer parameters and 70% of group-conv FLOPs.","key_machinery":"The central mechanism is a Kronecker-structured binary relationship matrix $U$ that masks the convolution kernel. For a layer with $C$ input and output channels, $U$ is built as $U = U_1 \\otimes \\cdots \\otimes U_K$ with $K = \\log_2 C$ factors, each $U_k = g_k \\mathbf{1} + (1-g_k) I$, where $g_k = \\operatorname{sign}(\\tilde{g}_k)$ and $\\tilde{g}_k$ is a continuous learnable gate. This construction uses only $\\log_2 C$ parameters per layer, always yields a valid group structure (adjacent or non-adjacent channels), and covers regular, group, and depthwise convolution as special cases. The sign function is trained with a straight-through estimator, and the complexity penalty $\\zeta = \\sum_l C_l \\prod_k (1 + g^l_k)$ lets the user cap the total number of active kernel connections. That combination is what makes the group count itself a differentiable, resource-controllable parameter.","core_discovery":"The paper claims that the optimal grouping strategy is layer-dependent and learnable, and that replacing a fixed cardinality like ResNeXt's 32 groups with learned per-layer group counts improves accuracy at matched or lower cost. In their ImageNet experiments, G-ResNeXt101 with budget $b=32$ reaches 79.9% top-1 accuracy versus 79.1% for their re-implemented ResNeXt101 32x4d, with slightly fewer parameters and roughly 0.70x the group-convolution FLOPs. When the learned group numbers are copied back into a standard ResNeXt and trained from scratch, accuracy stays nearly identical (79.8%), showing that the learned grouping strategy, not DGConv's masking, carries the gain. Under tighter budgets the method still holds accuracy: at 0.22x group-convolution FLOPs, G-ResNeXt101 reaches 79.5%. The paper also reports that lower layers tend to learn fewer groups, i.e., more capacity, which it reads as evidence that early layers carry more representational weight.","pith_inferences":["If the learned grouping strategies transfer across datasets, per-layer group counts could be treated as a reusable architectural prior rather than something re-learned per task; the paper does not test cross-dataset transfer.","DGConv's grouping is a structured form of sparsity, so it could be combined with channel pruning: the learned group boundaries identify which channel bundles are computationally separable, potentially guiding pruning decisions.","The straight-through estimator is one route to binarizing the gates; alternative relaxations such as Gumbel-softmax might find different (possibly better) group structures, which the paper does not explore.","The observation that early layers favor fewer groups suggests an efficiency guideline for future architectures: keep high-resolution early layers dense or wide, and concentrate grouping in deeper layers."],"forward_implications":["Per-layer cardinality becomes a tunable architectural choice: at a fixed parameter or FLOP budget, a network can spend capacity where it helps instead of spreading it uniformly.","A model trained with DGConv can be converted into an ordinary group-convolution network, so the learned structure ships with no runtime overhead and no custom kernels.","The observed pattern that early layers prefer fewer groups (more channels per group) provides a concrete design rule for hand-designed efficient networks.","The resource-constrained training with budget $b$ gives a direct control knob for the accuracy-complexity trade-off, with reported retention of accuracy at a fraction of group-convolution cost."],"supporting_citations":[{"why":"Defines ResNeXt and the fixed 32x4d baseline whose accuracy and complexity the paper compares against.","marker":"[33]"},{"why":"Supplies the residual building blocks and backbone that Groupable ResNeXt modifies.","marker":"[7]"},{"why":"Cites convergence of straight-through estimators, the training mechanism for the binary gates.","marker":"[5]"},{"why":"Provides the weighted Pareto-optimal resource-constraint formulation adapted in the complexity penalty.","marker":"[30]"},{"why":"Prior work that learns group connections but leaves group number a hyper-parameter, motivating the contribution.","marker":"[10]"},{"why":"Another prior learned-connection method with fixed group number, used to position DGConv's novelty.","marker":"[31]"},{"why":"Gives the Kronecker product decomposition used to build the large relationship matrix from small submatrices.","marker":"[2]"},{"why":"Defines depthwise convolution, which DGConv covers as a special case.","marker":"[9]"}],"fun_headline_variants":["Learn per-layer group counts, beat ResNeXt","GroupNet: adaptive channel groups outperform fixed ResNeXt","Learn optimal group splits per layer, beat ResNeXt","Differentiable grouping learns per-layer channels, beats ResNeXt","GroupNet: learned group counts per layer improve over ResNeXt"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the straight-through estimator used to train the binary gates genuinely guides them to useful grouping structures, rather than merely saturating to signs chosen by initialization and gradient noise.","fun_headline_variants_meta":{"raw":{"variants":["Learn per-layer group counts, beat ResNeXt","GroupNet: adaptive channel groups outperform fixed ResNeXt","Learn optimal group splits per layer, beat ResNeXt","Differentiable grouping learns per-layer channels, beats ResNeXt","GroupNet: learned group counts per layer improve over ResNeXt"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000373,"raw_usage":{"total_tokens":2012,"prompt_tokens":981,"completion_tokens":1031,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":597,"completion_tokens_details":{"reasoning_tokens":948}},"tokens_in":597,"tokens_out":1031,"duration_ms":7449,"temperature":1.0,"reasoning_tokens":948,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:03:27.155979+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"One decisive test: train two Groupable ResNeXts of the same budget with all gates initialized positive versus all negative. If their learned per-layer group counts differ substantially while validation accuracy matches, the grouping strategy is determined more by optimization path than by the data, and the claimed gain is not attributable to the learned structure. A simpler check is to replace the learned group counts in a trained model with a constant cardinality of matched FLOPs, retrain from scratch, and see whether accuracy falls.","supporting_citations":[{"cited_title":"Aggregated residual transformations for deep neural networks","cited_arxiv_id":null,"evidence_quote":"Defines ResNeXt and the fixed 32x4d baseline whose accuracy and complexity the paper compares against."},{"cited_title":"Understanding straight-through estima- tor in training activation quantized neural nets","cited_arxiv_id":null,"evidence_quote":"Cites convergence of straight-through estimators, the training mechanism for the binary gates."},{"cited_title":"Fully learnable group convolution for acceleration of deep neural networks","cited_arxiv_id":null,"evidence_quote":"Another prior learned-connection method with fixed group number, used to position DGConv's novelty."},{"cited_title":"A constructive arbitrary- degree kronecker product decomposition of tensors","cited_arxiv_id":null,"evidence_quote":"Gives the Kronecker product decomposition used to build the large relationship matrix from small submatrices."}],"review_version":1}