An arXiv preprint reports a compression method for large-kernel convolutional neural networks (CNNs) that was evaluated for its potential to reduce storage and parameter demands while preserving competitive performance and supporting edge deployment. The tests covered classification, scene parsing and object-level recognition on ImageNet-1K, CIFAR-100, ADE20K and COCO, along with mobile deployment measurements.
The central result is a set of reported engineering comparisons: the CGS-B variant reduced parameters by more than 77% across the evaluated RepLKNet, ConvNeXt and SLaK backbones, while accuracy drops stayed below 6.5% in each architecture. The claim is limited to those evaluated architectures and the benchmark setting.
A design aimed at the biggest parameter share
The method targets pointwise convolutions. In its Base variants, the paper reports that these operations account for over 87% of all parameters, identifying them as the dominant parameter bottleneck.
CGS uses an SVD-isomorphic low-rank structure. It shares the high-cost down- and up-projection matrices across channel groups, while keeping a separate scalable diagonal matrix for each group. In practice, the projection pieces are shared while the scaling remains group-specific.
The reported model setup applies CGS only to Stages 3 and 4, which contain approximately 97% of the model parameters. The authors describe early stages as more sensitive to compression; when all stages were compressed, the reported relative accuracy reduction was 1.6% for only 1.3 million saved parameters.
The trade-off across benchmarks
On ImageNet-1K classification, the authors report a 6.5% Top-1 improvement over PVT-Tiny with a 1.5 million-parameter increase. Compared with RepLKNet-31B, they report an 81% parameter reduction alongside a 4.2% accuracy drop. The analysis does not provide denominators for those percentage changes or a formal measure of statistical uncertainty.
The same approach was tested on two other tasks. In COCO object detection, the reported RepLKNet-31B comparison had competitive detection accuracy with more than 47% fewer parameters. On ADE20K semantic segmentation, the paper reports competitive single-scale mean intersection-over-union (mIoU), an overlap measure for segmentation, with a 57% parameter reduction against the original RepLKNet-31B.
What happened on a phone
The mobile results included RAM and loading time as well as model size. A RepLKNet-plus-CGS-B result reports 96.58 M of RAM and a 591-millisecond loading time; the accompanying comparison says loading was lower than alternatives with only slightly higher FLOPs.
On a OnePlus 13, CGS-B was listed at 198 ± 5 milliseconds of latency, 12.2 ± 0.3 milliwatt-hours of energy, 115 MB of memory and 5.0 frames per second. The corresponding RepLKNet-B figures were 242 ± 6 milliseconds, 16.3 ± 0.4 milliwatt-hours, 338 MB and 4.1 frames per second.
The paper also combined CGS-B with INT8 quantization, an eight-bit integer format. Against the INT8 baseline, CGS-B was reported to reduce latency by 25% and energy use by 43%; the combined CGS-B+INT8 result reached 135 ms and 7.8 mWh, with a reported relative accuracy drop of 0.5%.
A test beyond the device
The study also examined a device-to-cloud path. Wireless communication accounted for 54% to 62% of total latency in the test. With CGS, the reported RepLKNet-B model size fell from 306.94 MB to 58.41 MB, while transmitted data fell from 287 MB to 172 MB; end-to-end latency was reported as 27% lower on 5G and 25% lower on Wi-Fi 6.
In CIFAR-100 ablations, the reported preferred bottleneck dimension was R=200, alongside an automatically determined group count called k_auto. Performance was reported to degrade when the group count moved away from k_auto.
Where the evidence stops
These are descriptive comparisons, not universal performance estimates. The supplied analysis reports no confidence intervals or inferential tests for the accuracy comparisons. It also does not specify what the ± values mean or how many device runs produced them. The wireless results cover the tested network and device configurations, and the cross-architecture result is limited to the three evaluated backbones and benchmark setting.
Paper data and sources
Original title: Group-Shared Low-Rank Approximation for Mobile-Efficient Pointwise Convolutions in Large-Kernel CNNs
Authors: Hao Luo, Yiting Yang, Wenyi Zhao et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text