{"id":"d54f9eda-f57f-4f59-b3f3-c824e671bd8d","arxiv_id":"2412.10995","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A pure CNN mobile backbone using multi-level dilated convolutions reports faster or more accurate results than several mobile ViT, ViG, and hybrid baselines across classification, detection, and segmentation.","lead":"RapidNet is a mobile neural network backbone built from two parallel dilated convolutions at different dilation levels, plus inverted residual blocks and a large-kernel feedforward network. It reports top-1 accuracy of 76.3% on ImageNet-1K at 0.9 ms on an iPhone 13 mini NPU, claiming to beat several mobile ViT, ViG, and CNN baselines on accuracy or speed.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Speed half of the central claim is not established: Table 2's CoreML latencies lack conversion settings and variance, and the small 0.2-0.3 ms gaps could flip under an implementation-fair protocol.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern: the CoreML/ModelBench latency protocol on a single iPhone 13 Mini is not established as an implementation-fair proxy for mobile speed. My stress-test pass found no reason to reject the accuracy results: the ImageNet-1K, COCO, and ADE20K protocols follow standard practice, the ablations isolate the contribution of MLDC, and the code link provides some reproducibility support. However, the central claim is explicitly an 'accuracy and/or speed' claim, and for several baselines (EfficientFormer-L7, EfficientNet-B3, MobileViTv2-2.0) RapidNet is less accurate, so the speed disjunct is the only thing that makes the claim true for those comparisons. The reported latency gaps are small, and the paper does not report enough about the conversion and profiling process to rule out implementation unfairness. I therefore agree with the reader's conditional verdict: the paper is a plausible empirical contribution, but the speed half of the headline claim needs a controlled reproduction before it can be accepted as stated. The concrete test above would settle whether the concern actually lands: if the latency ordering is robust under a controlled CoreML pipeline, the central claim is supported; if not, the 'and/or speed' wording becomes an unsupported hedge.","tokens_in":16810,"tokens_out":5370,"duration_ms":47761,"concrete_test":"Re-run the Table 2 latency comparison under a controlled protocol: use the official released code and pretrained weights for RapidNet and every latency-listed baseline, convert each with the same coremltools version and the same compute-unit setting, preserve each model's official optimized inference mode (e.g., FastViT reparameterization, MobileViG's designated graph), and profile on the same iPhone 13 Mini across 10 separate sessions with cool-down, reporting mean and standard deviation for each model. If RapidNet-Ti is not statistically indistinguishable from MobileViG-Ti at 0.9 ms, or if RapidNet-S is not at least as fast as MobileNetV2x1.4 at 1.1 ms, then the speed half of the central claim loses support and the 'and/or speed' arguments against accuracy-superior baselines fail.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is 'outperforms state-of-the-art mobile CNN, ViT, ViG, and hybrid architectures in terms of accuracy and/or speed.' The accuracy half is supported by standard ImageNet/COCO/ADE20K protocols, but the speed half rests entirely on the latency column of Table 2. Section 4.1 reports only that 'all models are packaged as MLModels using CoreML and profiled on an iPhone 13 Mini (iOS 16) using ModelBench' with 50 inference rounds, 50 inferences per round, and a low/high trim of 10. It does not report the CoreML converter version, the compute-unit configuration (NPU/CPU/GPU), whether official converted models or authors' conversions were used, whether each baseline's optimized inference graph was preserved (e.g., FastViT structural reparameterization, MobileViG graph construction), or per-run variance and thermal state. The decisive comparisons involve tiny latency gaps: RapidNet-Ti at 0.9 ms versus MobileViG-Ti at 0.9 ms, and RapidNet-S at 1.1 ms versus MobileNetV2x1.4 at 1.1 ms. A systematic conversion choice or thermal throttling could easily change these equalities or reverse the ordering. For comparisons where RapidNet's accuracy is lower (e.g., EfficientFormer-L7, EfficientNet-B3), the 'or speed' disjunct is the only support for the central claim, so an unfair latency measurement would directly undermine the headline result. The concern is not that the measurement was deliberately manipulated, but that the paper provides no evidence that it was implementation-fair.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"RapidNet proposes Multi-Level Dilated Convolutions (MLDC), a purely CNN-based mobile backbone in which parallel 3×3 convolutions with dilation factors 2 and 3 are combined with a reparameterizable 7×7 depthwise convolution and a large-kernel FFN. The paper reports ImageNet-1K classification, COCO object detection and instance segmentation, and ADE20K semantic segmentation results, together with iPhone 13 Mini NPU latencies, and claims that RapidNet outperforms state-of-the-art mobile CNN, ViT, ViG, and hybrid architectures in terms of accuracy and/or speed at several model sizes. The central accuracy claim is supported by a broad set of benchmarks, while the speed half rests on a lightly documented CoreML/ModelBench latency measurement.","tokens_in":17158,"tokens_out":5138,"duration_ms":42685,"significance":"If the reported results hold, this is a useful contribution: it shows that a carefully designed pure CNN can match or exceed recent hybrid CNN-ViT and CNN-ViG mobile backbones on accuracy-latency tradeoffs. The paper provides broad experimental coverage across four tasks, ablations for each major design choice, and a public code repository. The main weakness is that the speed component of the headline claim depends on latency numbers whose protocol is under-specified and whose decisive gaps are very small, so the 'and/or speed' claim is not yet established at the same level as the accuracy claim.","major_comments":[{"comment":"The speed component of the central claim ('and/or speed') is not established because the latency protocol is under-specified. The paper reports only that all models were packaged as MLModels using CoreML and profiled on an iPhone 13 Mini (iOS 16) with ModelBench, 50 inference rounds, 50 inferences per round, and a low/high trim of 10. It does not report the CoreML converter version, the compute-unit configuration (NPU/CPU/GPU), whether official or author-converted models were used, whether baseline inference graphs preserved their optimized rewrites (e.g., FastViT's structural reparameterization or MobileViG's graph construction), per-run variance, or thermal state. Many decisive comparisons are ties or gaps of 0.1–0.2 ms (RapidNet-Ti 0.9 ms vs. MobileViG-Ti 0.9 ms; RapidNet-S 1.1 ms vs. MobileNetV2x1.4 1.1 ms), so these omissions leave the speed ordering unsupported. Please provide a fully reproducible protocol, per-model conversion settings, variance across multiple runs, and preferably measurements on more than one device.","section":"§4.1, Table 2"},{"comment":"The abstract states that RapidNet-Ti is 'faster and more accurate than MobileNetV2x1.4 (74.7% top-1 with 1.0 ms latency)', but Table 2 lists MobileNetV2x1.4 at 1.1 ms latency. This numerical inconsistency must be corrected, and all latency numbers should be checked for rounding consistency.","section":"Abstract vs. Table 2"},{"comment":"The justifications for core architectural choices rest on very small accuracy deltas. Table 5 reports that CPE adds 0.1%, LKFFN adds 0.1%, and MLDC adds 0.3%, with results averaged over only two runs; Table 6 reports a 0.1% gain for 5×5 kernels at +2.0M parameters and a 0.4% loss for dilation factors 3,4 or deformable convolution. These differences are within typical ImageNet training noise, yet they are used to select the final design. Please report per-run results, error bars or at least more runs, and consider whether the observed differences are statistically meaningful; as written, the component-level claims are not fully convincing.","section":"§4.4, Tables 5–6"}],"minor_comments":[{"comment":"The notation Y, Z, Out ∈ R^{N×N} omits the channel dimension and is inconsistent with the pointwise convolutions and fully connected layers in the same equations. Please use a notation such as R^{C×H×W} or describe the channel dimension explicitly.","section":"§3.2, Eqs. (4)–(7)"},{"comment":"The description of the MLDC block ordering is inconsistent: §3.2 says the two dilated convolutions are followed immediately by GeLU, while §3.4 mentions an additional pointwise convolution and BN after the dilated branches. Please align the text and equations with Figure 3f.","section":"§3.2, §3.4"},{"comment":"The hardware description 'NVIDIA RTX 6000 Ada generation GPUs' should read 'NVIDIA RTX 6000 Ada Generation GPUs' or 'Ada Lovelace GPUs', and the reference list should be checked for consistent formatting of author initials.","section":"§4, experimental setup"}],"recommendation":"major_revision","confidential_remarks":"The paper's own Section 4.1 acknowledges only minimal latency-protocol detail, yet the 'and/or speed' claim is a headline result. For a journal that values reproducibility, the authors should be asked to supply the exact conversion scripts, raw per-round latency distributions, and ideally cross-device measurements. The comparisons against MobileViG come from the same research group; this is not itself a problem, but it makes the need for an independent latency protocol more salient."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nYou should know about this paper for the accuracy-latency data point, not for any architectural breakthrough. The MLDC block is two standard 3x3 dilated convolutions (dilation 2 and 3) run in parallel, wrapped with a reparameterized 7x7 depthwise and FastViT's large-kernel FFN. That is a recombination of known components, but the specific configuration is new and the experimental coverage is genuinely broad: ImageNet-1K, COCO detection/segmentation, ADE20K, with ablations for every major choice. The accuracy story holds together. RapidNet-Ti's 76.3% top-1 at 0.6 GMACs is a real improvement over MobileViG-Ti's 75.7% at 0.7 GMACs, and the downstream numbers are consistent.\n\nThe soft spot is exactly where the stress-test put it: the speed half of the headline. All latency numbers come from CoreML/ModelBench on a single iPhone 13 Mini, but the paper does not report converter version, compute-unit settings, per-run variance, thermal state, or whether each baseline used its optimized inference graph (e.g., FastViT's reparameterization, MobileViG's graph construction). The decisive comparisons are ties or 0.1-0.3 ms gaps. That is far too fragile to support claims like \"faster than MobileNetV2x1.4\". This is not an accusation of manipulation; it is a missing evidence problem. The accuracy results do not depend on this, so the central claim can be rescued by rephrasing to emphasize accuracy/GMACs and downgrading latency to preliminary.\n\nSmaller issues: the ablation deltas are often 0.1-0.3%, and most numbers come from single runs, so the component justifications are weaker than the prose implies. There is also a minor inconsistency: the abstract quotes MobileNetV2x1.4 at 1.0 ms, the table says 1.1 ms. Fixable copyedit.\n\nOn balance this is a competent, honest empirical paper. It does not reorganize the field, but it gives a useful architecture data point with code and careful comparisons. I would send it to peer review, and I would require the authors to supply a proper latency measurement protocol and error bars (or rerun the ablations) before publication. Self-citation to MobileViG is not an issue; that is the direct baseline, and the comparison is central.","headline":"A solid empirical mobile backbone paper whose accuracy claims hold up but whose speed claims rest on thinner evidence than the headline suggests.","tokens_in":17726,"tokens_out":3328,"would_cite":true,"duration_ms":27777,"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":"A pure-CNN mobile backbone using parallel dilated convolutions claims to beat ViT, ViG, and hybrid models on accuracy or speed.","keywords":["Multi-Level Dilated Convolution","mobile vision backbone","efficient CNN","ImageNet-1K classification","COCO object detection","ADE20K semantic segmentation","receptive field expansion","mobile latency"],"falsifier":"Profile the same baseline set on a second mobile device or compiler stack with per-model optimized inference graphs; if MobileViG-Ti or MobileNetV2x1.4 runs at or below RapidNet-Ti's 0.9 ms, the speed claim is falsified. On the mechanism side, an effective-receptive-field analysis showing that MLDC's field is no wider than a plain stacked 3x3 CNN would undercut the long-range-interaction explanation.","tokens_in":16575,"feed_emoji":"⚡","tokens_out":9616,"duration_ms":69901,"temperature":0.7,"pith_summary":"RapidNet is a mobile vision backbone built entirely from convolutions, and the paper claims that this pure-CNN design beats state-of-the-art mobile transformers, vision graph networks, and hybrids on accuracy and/or speed. The key idea is Multi-Level Dilated Convolutions: two parallel 3x3 dilated convolutions with dilation rates 2 and 3 expand the theoretical receptive field to 5x5 and 7x7 at near-3x3 parameter cost, letting the network mix short-range and long-range features without attention or graph construction. On ImageNet-1K, RapidNet-Ti reaches 76.3% top-1 accuracy at 0.9 ms latency on an iPhone 13 Mini neural processing unit, compared with 75.7% at 0.9 ms for MobileViG-Ti and 74.7% at 1.0 ms for MobileNetV2x1.4; the larger RapidNet-M reaches 81.0% and leads on COCO detection and instance segmentation and on ADE20K semantic segmentation. The paper's conclusion is that a properly designed convolutional network can be the fastest and most accurate mobile backbone, without needing attention or graph mixers.","feed_headline":"Dilated convolutions beat mobile transformers on speed and accuracy","feed_subtitle":"RapidNet's pure-CNN backbone hits 76.3% top-1 at 0.9 ms, edging MobileViG and MobileNetV2.","key_machinery":"The load-bearing mechanism is the Multi-Level Dilated Convolution (MLDC) block: a reparameterizable 7x7 depthwise convolution followed by a pointwise convolution and two parallel 3x3 dilated convolutions with dilation factors 2 and 3, whose outputs are summed after batch normalization and GeLU activation. A 3x3 dilated convolution with dilation factor $d$ has theoretical receptive field $((3-1)d+1)^2$, so the two branches see 5x5 and 7x7 fields while each branch costs only a 3x3 kernel. The block is paired with a large-kernel feedforward network (a 7x7 depthwise convolution between two pointwise layers), and the whole Dilated Convolution Block occupies the last two stages, where MobileViG used graph convolution. This gives a pure-CNN path to the long-range feature mixing that attention and graph constructions provide.","core_discovery":"The paper's central discovery is that the graph-convolution block in MobileViG can be replaced by Multi-Level Dilated Convolutions with no loss of theoretical receptive field and a gain in efficiency. At the lowest-resolution stage, MobileViG's Sparse Vision Graph Attention with K=2 reaches a 7x7 theoretical receptive field over the image; a single 3x3 convolution with dilation 3 reaches exactly the same 7x7 field, and running it in parallel with a dilation-2 branch gives two scales of context at the parameter cost of two 3x3 kernels. RapidNet stacks inverted residual blocks for local processing in all four stages, then adds MLDC blocks plus reparameterizable 7x7 depthwise convolutions and a large-kernel feedforward network in the last two stages. Empirically, RapidNet-Ti outperforms MobileViG-Ti by 0.6% top-1 at the same 0.9 ms latency, RapidNet-M reaches 81.0% top-1, 42.0 AP box on COCO, and 41.5 mIoU on ADE20K, and ablations attribute the gains to the multi-level dilated branch rather than to capacity alone.","pith_inferences":["Editorial inference: if the latency ordering persists across other mobile compilers and devices, MLDC could serve as a cheap drop-in global mixer for other CNN backbones without changing their training pipeline.","Editorial inference: the equality of theoretical receptive fields between dilation-3 3x3 convolutions and MobileViG's SVGA does not guarantee equal effective receptive fields; a layer-wise effective-receptive-field analysis would test whether MLDC truly captures long-range interactions or mainly adds local multi-scale context.","Editorial inference: RapidNet-Ti has 6.6 million parameters versus MobileViG-Ti's 5.2 million, so the accuracy gain is not purely architectural; a matched-parameter comparison would isolate the MLDC contribution from the extra width.","Editorial inference: since dilation factors 2 and 3 outperformed 3 and 4, the optimal dilation appears resolution-dependent; a testable extension is per-stage dilation scheduling or input-adaptive dilation."],"forward_implications":["At the tiny scale, RapidNet-Ti reaches 76.3% top-1 on ImageNet-1K at 0.9 ms, beating MobileViG-Ti by 0.6% at the same latency and MobileNetV2x1.4 by 1.6% at lower latency.","At the medium scale, RapidNet-M reaches 81.0% top-1, 42.0 AP box on COCO object detection, 38.3 AP mask on instance segmentation, and 41.5 mIoU on ADE20K, outperforming comparable hybrids such as EfficientFormer and PoolFormer.","The same backbone transfers to dense prediction: Mask R-CNN with RapidNet-M beats FastViT-SA12 by 3.1 AP box and PoolFormer-S12 by 4.7 AP box on COCO.","Ablations show that replacing single-level dilated convolution with MLDC adds 0.3% top-1 at only 0.1 GMACs, and that dilation factors 2 and 3 beat 3 and 4, supporting the design choice."],"supporting_citations":[{"why":"Supplies the MobileViG baseline and the Sparse Vision Graph Attention block that MLDC replaces, along with the comparison numbers at each model size.","marker":"[46]"},{"why":"Provides the reparameterizable large-kernel depthwise convolution and large-kernel FFN used in RapidNet, and serves as a hybrid baseline in the downstream tables.","marker":"[58]"},{"why":"Defines the inverted residual block and provides MobileNetV2x1.0 and MobileNetV2x1.4 as the speed-accuracy comparison targets.","marker":"[52]"},{"why":"Supplies the dilated-convolution mechanism and the theoretical receptive field formula that justify multi-level dilation.","marker":"[64]"},{"why":"Defines the CoreML and ModelBench iPhone 13 Mini latency profiling protocol used for every speed comparison.","marker":"[57]"},{"why":"Provides the RegNetY-16GF teacher model used for knowledge distillation, making accuracy comparisons consistent across starred models.","marker":"[51]"},{"why":"EfficientFormer is a key hybrid CNN-ViT baseline and the source of the training recipe followed for downstream tasks.","marker":"[35]"},{"why":"PoolFormer is a CNN-Pooling baseline used across ImageNet, COCO, and ADE20K comparisons.","marker":"[65]"},{"why":"Mask R-CNN is the framework used to evaluate RapidNet on COCO object detection and instance segmentation.","marker":"[18]"},{"why":"Semantic FPN is the framework used for ADE20K semantic segmentation evaluation.","marker":"[30]"}],"fun_headline_variants":["Pure CNN beats hybrids with multi-level dilated convolutions","RapidNet: multi-level dilated convolutions outrun mobile ViTs","Dilated CNN backbone hits 76.3% top-1 at 0.9 ms on iPhone","Multi-level dilation makes pure CNN faster than mobile hybrids","RapidNet's dilated blocks beat graph attention on mobile"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The speed comparisons rest on the assumption that converting every model to CoreML and timing it on one iPhone 13 Mini gives a fair, implementation-neutral ranking of mobile latency.","fun_headline_variants_meta":{"raw":{"variants":["Pure CNN beats hybrids with multi-level dilated convolutions","RapidNet: multi-level dilated convolutions outrun mobile ViTs","Dilated CNN backbone hits 76.3% top-1 at 0.9 ms on iPhone","Multi-level dilation makes pure CNN faster than mobile hybrids","RapidNet's dilated blocks beat graph attention on mobile"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000404,"raw_usage":{"total_tokens":2169,"prompt_tokens":1078,"completion_tokens":1091,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":694,"completion_tokens_details":{"reasoning_tokens":999}},"tokens_in":694,"tokens_out":1091,"duration_ms":7476,"temperature":1.0,"reasoning_tokens":999,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:25:19.279046+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Profile the same baseline set on a second mobile device or compiler stack with per-model optimized inference graphs; if MobileViG-Ti or MobileNetV2x1.4 runs at or below RapidNet-Ti's 0.9 ms, the speed claim is falsified. On the mechanism side, an effective-receptive-field analysis showing that MLDC's field is no wider than a plain stacked 3x3 CNN would undercut the long-range-interaction explanation.","supporting_citations":[{"cited_title":"Mo- bilevig: Graph-based sparse attention for mobile vision ap- plications","cited_arxiv_id":null,"evidence_quote":"Supplies the MobileViG baseline and the Sparse Vision Graph Attention block that MLDC replaces, along with the comparison numbers at each model size."},{"cited_title":"Fastvit: A fast hybrid vision transformer using structural reparameterization","cited_arxiv_id":null,"evidence_quote":"Provides the reparameterizable large-kernel depthwise convolution and large-kernel FFN used in RapidNet, and serves as a hybrid baseline in the downstream tables."},{"cited_title":"Mobilenetv2: Inverted residuals and linear bottlenecks","cited_arxiv_id":null,"evidence_quote":"Defines the inverted residual block and provides MobileNetV2x1.0 and MobileNetV2x1.4 as the speed-accuracy comparison targets."},{"cited_title":"Designing network design spaces","cited_arxiv_id":null,"evidence_quote":"Provides the RegNetY-16GF teacher model used for knowledge distillation, making accuracy comparisons consistent across starred models."},{"cited_title":"Metaformer is actually what you need for vision","cited_arxiv_id":null,"evidence_quote":"PoolFormer is a CNN-Pooling baseline used across ImageNet, COCO, and ADE20K comparisons."},{"cited_title":"Mask r-cnn","cited_arxiv_id":null,"evidence_quote":"Mask R-CNN is the framework used to evaluate RapidNet on COCO object detection and instance segmentation."},{"cited_title":"Panoptic feature pyramid networks","cited_arxiv_id":null,"evidence_quote":"Semantic FPN is the framework used for ADE20K semantic segmentation evaluation."}],"review_version":1}