{"id":"f3da9625-3dbd-4c94-90ea-9ed98d57e910","arxiv_id":"1908.11036","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"DWnet, a deep plus broad network for skeleton action recognition, reaches accuracy near HCN on SBU and NTU datasets while reducing test time per sample by about 100 times.","lead":"This paper builds a network for 3D action recognition by combining a deep convolutional feature extractor with a broad learning system. The authors report that the combined model matches the accuracy of a strong baseline on two skeleton datasets while cutting per-sample test time by roughly two orders of magnitude.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported ~100x testing-time speedup is not yet supported: DWnet retains HCN's convolution stack, so the timing comparison appears to measure only the BLS readout rather than the full DWnet pipeline end-to-end.","rationale":"Reader's weakest assumption concerned whether a randomly expanded Fc6 feature can match HCN's classifier, together with test-set selection of enhancement-node counts. That is a valid concern about the accuracy half of the claim. However, the single most load-bearing vulnerability is the speed half: the paper's own complexity analysis covers only the removed layers, and it cannot explain a roughly 190x total-latency reduction because the retained convolutional stack is identical in both models. If the DWnet timing omits the PruHCN forward pass, the central contribution of saving testing time collapses even if accuracy is accepted. The concrete timing check should therefore precede acceptance. Since the paper was already CONDITIONAL and this concern reinforces that condition rather than overturning the verdict, I keep the verdict unchanged while disagreeing with the reader's identification of the load-bearing concern.","tokens_in":12602,"tokens_out":7853,"duration_ms":74849,"concrete_test":"Re-time HCN and DWnet in the same framework and on the same GPU, with DWnet implemented as the full end-to-end pipeline: encoded skeleton to PruHCN (Conv1-5 plus Fc6) to BLS enhancement nodes to output, and HCN implemented as its complete architecture including Dropout and Fc7. Measure wall-clock time per sample on identical SBU test batches without caching features, and report a per-layer timing breakdown. If the full DWnet time is not about 100x below HCN, the headline speed claim should be dropped or substantially weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III.C justifies the speed advantage by comparing only the pruned Dropout and Fc7 layers with the BLS readout, and Table II reports HCN at 4.12e-2 s/sample versus DWnet at 2.17e-4 s/sample on SBU. This comparison is the core of the paper's headline claim, but it is not yet supported. In PruHCN, all convolutional layers and Fc6 are retained (Fig. 2), and on the SBU input (16x15x3) those layers account for virtually all multiply-adds; the removed Dropout and Fc7 layers are a Bernoulli mask plus a 256x8 dense product. Merely substituting the BLS readout cannot make the full DWnet about 190x faster than HCN if HCN's timing includes the same convolutional and Fc6 forward pass. Either the HCN timing is dominated by an implementation artifact such as unoptimized framework code, CPU inference, or data-loading overhead, or the DWnet timing excludes the PruHCN forward pass and is not an end-to-end measurement. The paper does not state which quantity was timed, and no code is released. The complexity analysis in Section III.C therefore supports a claim about the deleted layers, not about total latency. The accuracy comparison may also be weakened by selecting enhancement-node counts from test-set curves (Figs. 5-7), but that concern is secondary to the missing full-pipeline timing evidence.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DWnet, which combines a pruned Hierarchical Co-occurrence Network (PruHCN) with a Broad Learning System (BLS) readout for 3D skeleton-based action recognition. The deep part is HCN with the final Dropout and Fc7 layers removed, and the BLS readout consists of randomly generated enhancement nodes followed by a ridge-regression output layer. On SBU Kinect, the authors report 97.39% accuracy for DWnet versus 94.83% for HCN; on NTU RGB+D they report near-identical accuracies (84.07% vs 84.30% CS; 89.84% vs 89.90% CV). The headline claim is an approximately 100x reduction in per-sample testing time (e.g., 4.12e-2 s vs 2.17e-4 s on SBU). The paper also compares against a plain BLS and an HCNBLS variant with untrained PruHCN modules.","tokens_in":12925,"tokens_out":5977,"duration_ms":53626,"significance":"If the efficiency claim were properly supported, the idea of using a pretrained deep feature extractor with a broad-learning readout for skeleton-based action recognition would be of moderate interest, especially for latency-sensitive applications. The architecture is clearly described, and the experiments cover two standard datasets with a direct comparison to HCN. However, the paper provides no code, no repeated-run statistics, and only one strong deep-learning baseline, so the empirical support for the 'state-of-the-art' conclusion is weak. Crucially, the central speedup claim is not yet established by the reported measurements.","major_comments":[{"comment":"The roughly 100x speedup claim is not supported by the evidence because the paper does not state whether the reported times are end-to-end or only for the readout layers. DWnet retains all of HCN's convolutional layers and Fc6 (Fig. 2), so replacing only Dropout and Fc7 with a BLS readout cannot reduce total inference time by 190x on SBU if both timings include the same convolutional forward pass. Additionally, Dropout is not active at inference time in standard HCN, so pruning it has no effect on test-time cost. Please report end-to-end timings for both models with the same hardware/software stack, and state exactly which operations are included in each measurement.","section":"Section III.C, Tables II and III"},{"comment":"The number of enhancement nodes m is selected by maximizing test-set accuracy: Fig. 5 reports the best average testing accuracy at m=550, and Figs. 6-7 select 7500 and 12500 by the same criterion. This selection on the test set makes the accuracies in Tables I and III optimistic relative to what would be expected on unseen data. Please tune m on a validation split (or report accuracy as a function of m without claiming the maximum as expected performance) and provide repeated-run mean ± std to establish the stability of the reported improvements.","section":"Section IV.C, Figs. 5-7"},{"comment":"The complexity argument is incomplete and does not by itself imply the claimed speedup. On NTU RGB+D, the BLS readout with 7500 or 12500 enhancement nodes multiplies a feature vector of dimension 256 plus the enhancement dimension by the 60-class output matrix, yielding far more multiply-adds than HCN's Fc7 (256x60). On SBU, 614x8 operations for the BLS readout exceed 64x8 for Fc7. Thus the statement that 'the computational complexity of the BLS is lower than that of pruned layer of HCN' is not true in terms of multiply-add count for the configurations actually used, and the reported timing difference likely reflects differences in implementation, batching, or measurement scope rather than the architecture change.","section":"Section III.C"},{"comment":"The conclusion that DWnet 'can achieve the state-of-the-art performance' is unsupported by the comparisons presented. The only deep-learning baseline is HCN itself, and on NTU RGB+D DWnet is slightly worse than HCN (89.84% vs 89.90% CV; 84.07% vs 84.30% CS). Please either temper the claim to 'comparable to HCN' or add the relevant state-of-the-art baselines from the skeleton-based action recognition literature.","section":"Section V and Section IV.D"}],"minor_comments":[{"comment":"There are several typos: 'irregualr' in Section III.A, 'spapatial-temporal' and 'sapatial-temporal' in Section III.B, and 'pruned' used as a verb in Fig. 2 caption.","section":"Section III.A and III.B"},{"comment":"The text says 'There are 6 layers including 4 convolutional layers and 2 fully connected layers in the PruHCN,' but Fig. 2 shows PruHCN includes five convolutional layers (Conv1-Conv5) and one fully connected layer (Fc6); please correct the layer counts.","section":"Section III.B"},{"comment":"Equation (2) defines Z_i for i=1,...,n but the paper then uses Z without a subscript in Eq. (3); please clarify whether Z denotes Z_1 given that n=1.","section":"Eq. (2) and Eq. (3)"},{"comment":"The table header is confusing: the columns are labeled 'CV' and 'CS' with 'Testing Time' straddling the accuracy and time columns, making it unclear which number is accuracy and which is time. Please use explicit column headers such as 'CV Acc.' and 'CV Time'.","section":"Table III"},{"comment":"References [23] and [34] cite the same NTU RGB+D dataset paper; please merge them to avoid duplicate citations.","section":"References"},{"comment":"The paper says DWnet 'almost achieves real-time testing' but does not define real-time relative to any frame-rate requirement; please specify the target frame rate and the measurement protocol.","section":"Section IV.C"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the timing measurement: if an end-to-end measurement does not reproduce the claimed speedup, the contribution reduces to a small accuracy change from replacing the final classifier, which is likely below the novelty bar for this venue. The authors should be asked to provide precise timing methodology and code or detailed pseudocode to make the measurement reproducible. The test-set-based selection of enhancement nodes is also a significant methodological concern that should be addressed with a validation split."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is a simple, sensible assembly: take a pretrained skeleton CNN (HCN), chop off the last two layers, and replace them with a broad learning system ridge-regression readout. That specific combination is new in the literature, and the accuracy results are genuinely close to the HCN baseline: on NTU you lose less than 0.3% in both CS and CV, and on SBU you even gain a couple points. I believe the accuracy claim, as far as it goes. The paper is also honest about being non-end-to-end and about not being a full theoretical treatment.\n\nThe weak spot is the speed claim, and it is not minor. The paper reports DWnet at 2.17e-4 s/sample vs HCN at 4.12e-2 s/sample, about a 190x gap. But DWnet keeps all the convolutional layers and Fc6 from HCN (Fig. 2). Those layers account for virtually all the multiply-adds on a 16x15x3 input; the removed Dropout and Fc7 are a mask and a 256x8 dense product. If both models run the same conv stack, DWnet cannot be two orders of magnitude faster. The timing number for DWnet is suspiciously close to the BLS-alone timing (8.30e-5), which tells me the authors timed only the BLS readout, not the full PruHCN forward pass. The complexity analysis in III.C compares BLS to the pruned layers only, not to the total network, so it does not justify the headline. The paper never states what quantity was timed, and no code is released to check.\n\nSecondary issues: the enhancement-node count is chosen from test-set curves (Figs. 5-7), there are no repeated-run statistics, and the conclusion calls the result state-of-the-art after comparing with exactly one action-recognition baseline (HCN). The BLS and HCNBLS baselines are much weaker by design, so they do not make the SOTA claim meaningful. On the plus side, the related work is fine, the method is described clearly, and the accuracy comparison with HCN on two datasets is a reasonable first evidence.\n\nVerdict: the architecture is worth knowing about, but the speed evidence needs to be redone end-to-end, with a clear statement of what was measured, ideally with timing breakdowns and code. If the speedup survives that, it's a useful practical result, though not a breakthrough. This deserves a serious referee because the core question is empirical and correctable, but I would not cite it in its current form.","headline":"The BLS-plus-CNN architecture is a plausible engineering idea, but the paper's headline 100x speedup claim is not supported by the timing evidence, which appears to measure only the BLS readout, not the full pipeline.","tokens_in":13421,"tokens_out":1823,"would_cite":false,"duration_ms":19752,"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":"Replacing a deep network's final layers with a broad learning readout preserves 3D action accuracy and cuts test time by 100x.","keywords":["3D skeleton action recognition","broad learning system","hierarchical co-occurrence network","deep-wide network","test-time efficiency","ridge regression","enhancement nodes"],"falsifier":"Train DWnet while choosing enhancement-node counts on a validation split rather than the test set, then test under the same preprocessing. If accuracy falls below HCN's by a margin comparable to the reported gains (for example, more than 2.5% on SBU), the claim that random enhancement restores the pruned layers' approximation is refuted. A complementary check is to train a linear classifier directly on the same Fc6 features; if it already matches DWnet, the enhancement nodes add little.","tokens_in":12416,"feed_emoji":"⚡","tokens_out":6537,"duration_ms":58119,"temperature":0.7,"pith_summary":"The paper tries to show that a deep convolutional feature extractor and a broad learning system can be combined into one network that recognizes 3D skeleton actions almost as accurately as the deep network alone, but around 100 times faster at test time. The deep half is a pruned hierarchical co-occurrence network (PruHCN), which keeps the convolutional trunk and the first fully connected layer of the original HCN. The wide half is a broad learning readout: randomly initialized enhancement nodes expand the PruHCN feature vector, and the output weights are solved by ridge regression. On the SBU Kinect dataset the combined model reports 97.39% accuracy versus 94.83% for HCN, with average test time 2.17e-4 seconds per sample versus 4.12e-2. On NTU RGB+D, accuracy stays within 0.06 to 0.23 points of HCN while keeping a similar speed advantage.","feed_headline":"Replacing a deep network's final layers with a broad learning readout preserves 3D…","feed_subtitle":"A pruned HCN plus a broad readout matches the original model's accuracy on SBU and NTU while running about 100 times faster.","key_machinery":"The central object is the PruHCN feature mapping feeding a broad learning readout. The feature mapping writes $Z_i = \\theta(P(\\mathrm{ReLU}(X \\otimes W_{ei} + \\beta_{ei})))$; the enhancement nodes are $H_j = \\xi(Z W_{hj} + \\beta_{hj})$ with random $W_{hj}, \\beta_{hj}$ and a tansig activation; the final output is the concatenation of $Z$ and all $H_j$ multiplied by ridge-regression weights. This machine carries the argument because the random expansion is what compensates for the two pruned layers, and the closed-form solution is what makes testing fast.","core_discovery":"The central claim is that replacing the Dropout layer and the final Fc7 softmax layer of a trained HCN with a broad learning system readout preserves learning capacity and improves inference efficiency. The pruned trunk still learns local spatial-temporal and co-occurrence features; the broad readout expands the Fc6 feature vector through random nonlinear enhancement nodes and connects both to an output layer whose weights are obtained in closed form. The reported numbers are 97.39% versus 94.83% on SBU and 84.07% versus 84.30% on NTU cross-subject, with per-sample test times falling from about 4.12e-2 seconds to 2.17e-4 seconds on SBU and from about 1.1e-2 seconds to 2.26e-4 seconds on NTU. The paper interprets this as evidence that the width of the broad readout can restore the approximation ability removed by pruning.","pith_inferences":["The recipe generalizes beyond skeletons: any CNN that produces a dense feature vector could have its classifier pruned and replaced by a broad readout for faster inference, but the paper does not test this.","The reported 100x speed gain is tied to the specific enhancement-node counts; for very wide enhancement layers the ridge regression itself may dominate the runtime and weaken the advantage.","The success of random enhancement nodes suggests the Fc6 representation is already close to linearly separable; a simple probe such as a linear classifier on the same Fc6 features could reveal how much of the accuracy comes from the broad readout rather than the deep trunk."],"forward_implications":["The last fully connected layers of a trained deep CNN can be swapped for a ridge-regression readout without hurting accuracy, so the expensive softmax classifier is not a necessary part of the model.","Skeleton action recognition can run near real time on modest hardware: per-sample times near 2e-4 seconds move the bottleneck from classification to feature extraction.","The enhancement-node count is a dataset-dependent knob; the paper reports optimal values of 550 on SBU and 7500 or 12500 on NTU, so accuracy and speed trade off through a single width parameter.","Because only the readout changes, the same extracted features can be reclassified after retraining just the readout, avoiding a full end-to-end retraining pass."],"supporting_citations":[{"why":"Supplies the HCN architecture whose final Dropout and Fc7 layers are pruned to make PruHCN.","marker":"[19]"},{"why":"Introduces the broad learning system and the ridge-regression readout used in DWnet.","marker":"[20]"},{"why":"Establishes the approximation capability of BLS that the paper relies on to justify replacing pruned layers with enhancement nodes.","marker":"[27]"},{"why":"Provides the SBU Kinect Interaction dataset and the five-fold evaluation protocol used for the main accuracy and timing comparison.","marker":"[33]"},{"why":"Provides the NTU RGB+D dataset and the cross-subject and cross-view protocols used to test generality.","marker":"[23]"}],"fun_headline_variants":["Deep–wide net cuts 3D action test time 100× with stable accuracy","Pruned 3D action features plus broad readout hit real-time speed","DWnet yokes deep and wide learning for faster 3D action recognition","Broad learning readout replaces softmax, 100× faster 3D action tests","Deep trunk with wide readout preserves 3D action accuracy, runs faster"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a single trained PruHCN's Fc6 feature vector, expanded by randomly initialized enhancement nodes and solved with ridge regression, can maintain or improve the classification power of the original HCN's Dropout and Fc7 layers.","fun_headline_variants_meta":{"raw":{"variants":["Deep–wide net cuts 3D action test time 100× with stable accuracy","Pruned 3D action features plus broad readout hit real-time speed","DWnet yokes deep and wide learning for faster 3D action recognition","Broad learning readout replaces softmax, 100× faster 3D action tests","Deep trunk with wide readout preserves 3D action accuracy, runs faster"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000224,"raw_usage":{"total_tokens":1420,"prompt_tokens":863,"completion_tokens":557,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":479,"completion_tokens_details":{"reasoning_tokens":452}},"tokens_in":479,"tokens_out":557,"duration_ms":6136,"temperature":1.0,"reasoning_tokens":452,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:28:02.982073+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train DWnet while choosing enhancement-node counts on a validation split rather than the test set, then test under the same preprocessing. If accuracy falls below HCN's by a margin comparable to the reported gains (for example, more than 2.5% on SBU), the claim that random enhancement restores the pruned layers' approximation is refuted. A complementary check is to train a linear classifier directly on the same Fc6 features; if it already matches DWnet, the enhancement nodes add little.","supporting_citations":[{"cited_title":"Co-occurrence feature learning from skeleton data for action recog- nition and detection with hierarchical aggregation,","cited_arxiv_id":null,"evidence_quote":"Supplies the HCN architecture whose final Dropout and Fc7 layers are pruned to make PruHCN."},{"cited_title":"Broad learning system: An effective and efficient incremental learning sys- tem without the need for deep architecture,","cited_arxiv_id":null,"evidence_quote":"Introduces the broad learning system and the ridge-regression readout used in DWnet."},{"cited_title":"Universal approximation capability of broad learning system and its structural variations,","cited_arxiv_id":null,"evidence_quote":"Establishes the approximation capability of BLS that the paper relies on to justify replacing pruned layers with enhancement nodes."},{"cited_title":"Two-person interaction detec- tion using body-pose features and multiple instance 12 learning,","cited_arxiv_id":null,"evidence_quote":"Provides the SBU Kinect Interaction dataset and the five-fold evaluation protocol used for the main accuracy and timing comparison."},{"cited_title":"Ntu rgb+d: A large scale dataset for 3d human activity analysis,","cited_arxiv_id":null,"evidence_quote":"Provides the NTU RGB+D dataset and the cross-subject and cross-view protocols used to test generality."}],"review_version":1}