Scholarlink

<p>With each passing year, the compelling need to bring deep learning computational models to the edge grows, as does the disparity in resource demand between these models and the Internet of Things (IoT). Understanding the structural complexities of deep models helps in realizing possible partitioning scenarios to limit the under-utilization of the edge device pool. It is also crucial to leverage the existing compression techniques to apprehend the limited processing power of edge devices. This paper proposes a well-partitioned inference model for bringing deep convolutional networks to the edge. The kernels of the entire model are compressed using Hashing Trick, and an n-ball is used to regenerate neighbourhood pixels at runtime, eliminating memory and computational bottlenecks. The proposed method uses vertical partitioning to fuse convolutional layers and horizontal partitioning for fully connected layers, modelled as a graph to optimize task scheduling across devices. This approach achieves a 40% reduction in communication overhead and <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="10586_2025_5386_Article_IEq1.gif" Format="GIF" Height="10" Rendition="HTML" Resolution="72" Type="Linedraw" Width="17" /> </InlineMediaObject> <EquationSource Format="TEX">\(\approx\)</EquationSource> <EquationSource Format="MATHML"><math> <mo>≈</mo> </math></EquationSource> </InlineEquation> 7x speedup in different partitioning scenarios, enabling scalable and efficient deep learning inference at the edge.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Adaptive partitioning of DNNs for resource-efficient inference on edge clusters

  • 期刊论文

  • 发表日期：

    2025年9月3日

查看全文
  • Azra Nazir,
  • Faisal Rasheed Lone,
  • Ashfaq Ahmad Najar

摘要

With each passing year, the compelling need to bring deep learning computational models to the edge grows, as does the disparity in resource demand between these models and the Internet of Things (IoT). Understanding the structural complexities of deep models helps in realizing possible partitioning scenarios to limit the under-utilization of the edge device pool. It is also crucial to leverage the existing compression techniques to apprehend the limited processing power of edge devices. This paper proposes a well-partitioned inference model for bringing deep convolutional networks to the edge. The kernels of the entire model are compressed using Hashing Trick, and an n-ball is used to regenerate neighbourhood pixels at runtime, eliminating memory and computational bottlenecks. The proposed method uses vertical partitioning to fuse convolutional layers and horizontal partitioning for fully connected layers, modelled as a graph to optimize task scheduling across devices. This approach achieves a 40% reduction in communication overhead and \(\approx\) ≈ 7x speedup in different partitioning scenarios, enabling scalable and efficient deep learning inference at the edge.