Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SlimEdge: Lightweight Distributed DNN Deployment on Constrained Hardware

Mahadev Sunil Kumar, Arnab Raha, Debayan Das, Gopakumar G, Amitava Mukherjee

TL;DR

SlimEdge addresses the challenge of deploying complex vision models like MVCNN on resource-constrained edge devices by introducing a view- and device-aware pruning framework. It combines a Taylor-based filter pruning strategy, view-importance estimation, device performance metrics, and a multi-objective NSGA-II optimization (with a Beta-aware initializer) to generate per-view, per-device pruned models that satisfy accuracy and memory budgets. The approach is validated on ModelNet40 with a distributed MVCNN setup across heterogeneous Raspberry Pi devices, achieving up to 3.94x latency reduction and 1.2–5.0x end-to-end speedups while maintaining target accuracy. The work demonstrates that jointly optimizing view importance and device capability yields deployable, throughput-optimized distributed inference in constrained edge environments, and suggests extensions to quantify energy and communication efficiency at larger scales.

Abstract

Deep distributed networks (DNNs) have become central to modern computer vision, yet their deployment on resource-constrained edge devices remains hindered by substantial parameter counts and computational demands. Here, we present an approach to the efficient deployment of distributed DNNs that jointly respects hardware limitations and preserves task performance. Our method integrates a structured model pruning with a multi-objective optimization to tailor network capacity to heterogeneous device constraints. We demonstrate this framework using Multi-View Convolutional Neural Network (MVCNN), a state-of-the-art architecture for 3D object recognition, by quantifying the contribution of individual views to classification accuracy and allocating pruning budgets, respectively. Experimental results show that the resulting models satisfy user-specified bounds on accuracy and memory footprint while reducing inference latency by factors ranging from 1.2x to 5.0x across diverse hardware platforms. These findings suggest that performance-aware, view-adaptive compression provides a viable pathway for deploying complex vision models in distributed edge environments.

SlimEdge: Lightweight Distributed DNN Deployment on Constrained Hardware

TL;DR

SlimEdge addresses the challenge of deploying complex vision models like MVCNN on resource-constrained edge devices by introducing a view- and device-aware pruning framework. It combines a Taylor-based filter pruning strategy, view-importance estimation, device performance metrics, and a multi-objective NSGA-II optimization (with a Beta-aware initializer) to generate per-view, per-device pruned models that satisfy accuracy and memory budgets. The approach is validated on ModelNet40 with a distributed MVCNN setup across heterogeneous Raspberry Pi devices, achieving up to 3.94x latency reduction and 1.2–5.0x end-to-end speedups while maintaining target accuracy. The work demonstrates that jointly optimizing view importance and device capability yields deployable, throughput-optimized distributed inference in constrained edge environments, and suggests extensions to quantify energy and communication efficiency at larger scales.

Abstract

Deep distributed networks (DNNs) have become central to modern computer vision, yet their deployment on resource-constrained edge devices remains hindered by substantial parameter counts and computational demands. Here, we present an approach to the efficient deployment of distributed DNNs that jointly respects hardware limitations and preserves task performance. Our method integrates a structured model pruning with a multi-objective optimization to tailor network capacity to heterogeneous device constraints. We demonstrate this framework using Multi-View Convolutional Neural Network (MVCNN), a state-of-the-art architecture for 3D object recognition, by quantifying the contribution of individual views to classification accuracy and allocating pruning budgets, respectively. Experimental results show that the resulting models satisfy user-specified bounds on accuracy and memory footprint while reducing inference latency by factors ranging from 1.2x to 5.0x across diverse hardware platforms. These findings suggest that performance-aware, view-adaptive compression provides a viable pathway for deploying complex vision models in distributed edge environments.
Paper Structure (30 sections, 28 equations, 14 figures, 6 tables)

This paper contains 30 sections, 28 equations, 14 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. ModelNet40 Dataset
  5. SlimEdge Architecture
  6. SlimEdge
  7. Components of SlimEdge
  8. Model Pruning
  9. Calculating the minimum pruning vector
  10. Computing view-importance
  11. Calculating device performance
  12. Multi-Objective Formulation
  13. Importance-aware pruning strategy
  14. Importance-aware sampling
  15. Multi-objective problem
  16. ...and 15 more sections

Figures (14)

  • Figure 1: Roadside Vehicle Monitering System using 3 Cameras
  • Figure 2: Optimization Framework using 12 Edge Devices
  • Figure 3: Views of airplane_0006 from 12 different views from the ModelNet40 dataset
  • Figure 4: Top-Level Flowchart of SlimEdge
  • Figure 5: Component-Level Flowchart of SlimEdge
  • ...and 9 more figures