---
title: "Group-Shared Low-Rank Approximation for Mobile-Efficient Pointwise Convolutions in Large-Kernel CNNs"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.26069v1"
permalink: "https://feed.astralor.com/items/itm_8b027dc0bc37a8eab8f839c3"
published: "2026-08-26T17:36:35.000Z"
score: 72
channel: "tech"
tags: ["ConvNeXt","RepLKNet","SLaK","Low-rank Approximation","Parameter Sharing","Edge Deployment"]
---

# Group-Shared Low-Rank Approximation for Mobile-Efficient Pointwise Convolutions in Large-Kernel CNNs

## AI commentary

> This paper offers a crucial insight: while the field focuses on compressing depthwise convolutions, pointwise convolutions are the real heavyweight. For engineers working on on-device deployment, CGS group-sharing is a simple yet effective idea worth trying on their own models.

## At a glance

Large-kernel CNNs achieve excellent performance in vision tasks but suffer quadratic parameter growth, hindering edge deployment. Existing efficient architectures compress depthwise convolutions while overlooking that pointwise convolutions account for over 87% of parameters. This paper proposes a Channel Group-Shared (CGS) low-rank approximation based on SVD, sharing down/up projection matrices across channel groups and using group-specific scaling diagonals, significantly reducing parameters and enabling deployment of RepLKNet, ConvNeXt, SLaK on edge devices.

## Key points

- Identifies pointwise convolutions as the dominant parameter bottleneck (>87%) in large-kernel CNNs
- Proposes Channel Group-Shared (CGS) low-rank approximation structurally isomorphic to SVD, sharing projection matrices
- Demonstrates favorable accuracy-storage trade-offs on RepLKNet, ConvNeXt, and SLaK

## Full text

Large-kernel Convolutional Neural Networks (CNNs) deliver remarkable performance in vision tasks by significantly expanding receptive fields, yet their quadratic parameter growth critically impedes storage-efficient edge deployment. While existing efficient architectures adopt parameter-efficient depthwise separable convolution backbones that leverage techniques like low-rank approximation and weight sharing to compress depthwise convolutions, we identify a critical oversight: pointwise convolutions dominate parameter volume (>87% in models like RepLKNet-31B) and constitute the primary deployment bottleneck on resource-constrained edge devices. This results in prohibitive storage costs and severe memory-loading constraints on resource-limited devices (e.g., smartphones with 4-12 GB Random Access Memory (RAM)). To overcome this, we propose Channel Group-Shared (CGS) low-rank approximation, a novel Singular Value Decomposition (SVD)-based parameter-sharing strategy. CGS constructs a structured low-rank paradigm isomorphic to SVD decomposition, comprising shared (high-parameter-cost) down/up-projection matrices across channel groups within a layer and channel-group-specific (low-parameter-cost) scalable diagonal matrices. This group-sharing design achieves significant parameter reduction. Extensive experiments demonstrate that large-kernel CNNs (RepLKNet, ConvNeXt, SLaK) enhanced with CGS strike an empirically favorable balance between competitive performance and substantially reduced storage costs. Crucially, by alleviating storage constraints, reducing memory bandwidth pressure during loading, and minimizing model loading latency, CGS enables the feasible deployment of pre-trained large-kernel CNN models on edge devices, thereby bridging the gap between high-performance vision models and practical edge deployment.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_8b027dc0bc37a8eab8f839c3