Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ATM-Net: Adaptive Termination and Multi-Precision Neural Networks for Energy-Harvested Edge Intelligence

Neeraj Solanki, Sepehr Tabrizchi, Samin Sohrabi, Jason Schmidt, Arman Roohi

TL;DR

ATM-Net tackles energy-heterogeneous edge inference for energy-harvested IoT by fusing adaptive termination (multi-exit CNNs) with mixed-precision computing (32/8/4-bit) and an energy-aware scheduler (EATS). It introduces two target networks (ResNet-18 and DenseNet-121) with strategically placed exits and Quantization-Aware Training to preserve accuracy under low bit-widths. The EATS module adjusts precision based on charging rate $R_c$ and remaining energy $E_{sys}$, using thresholds $R_{th1}$, $R_{th2}$ and energy threshold $E_{th}$ to maximize energy-accuracy trade-offs at runtime. Empirical results on CIFAR-10, PlantVillage, and TissueMNIST show substantial energy reductions (PDP down to approximately $0.1$ J) and competitive accuracy, including up to 96.93% with Q4, validating the approach for energy-harvested edge intelligence.

Abstract

ATM-Net is a novel neural network architecture tailored for energy-harvested IoT devices, integrating adaptive termination points with multi-precision computing. It dynamically adjusts computational precision (32/8/4-bit) and network depth based on energy availability via early exit points. An energy-aware task scheduler optimizes the energy-accuracy trade-off. Experiments on CIFAR-10, PlantVillage, and TissueMNIST show ATM-Net achieves up to 96.93% accuracy while reducing power consumption by 87.5% with Q4 quantization compared to 32-bit operations. The power-delay product improves from 13.6J to 0.141J for DenseNet-121 and from 10.3J to 0.106J for ResNet-18, demonstrating its suitability for energy-harvesting systems.

ATM-Net: Adaptive Termination and Multi-Precision Neural Networks for Energy-Harvested Edge Intelligence

TL;DR

ATM-Net tackles energy-heterogeneous edge inference for energy-harvested IoT by fusing adaptive termination (multi-exit CNNs) with mixed-precision computing (32/8/4-bit) and an energy-aware scheduler (EATS). It introduces two target networks (ResNet-18 and DenseNet-121) with strategically placed exits and Quantization-Aware Training to preserve accuracy under low bit-widths. The EATS module adjusts precision based on charging rate and remaining energy , using thresholds , and energy threshold to maximize energy-accuracy trade-offs at runtime. Empirical results on CIFAR-10, PlantVillage, and TissueMNIST show substantial energy reductions (PDP down to approximately J) and competitive accuracy, including up to 96.93% with Q4, validating the approach for energy-harvested edge intelligence.

Abstract

ATM-Net is a novel neural network architecture tailored for energy-harvested IoT devices, integrating adaptive termination points with multi-precision computing. It dynamically adjusts computational precision (32/8/4-bit) and network depth based on energy availability via early exit points. An energy-aware task scheduler optimizes the energy-accuracy trade-off. Experiments on CIFAR-10, PlantVillage, and TissueMNIST show ATM-Net achieves up to 96.93% accuracy while reducing power consumption by 87.5% with Q4 quantization compared to 32-bit operations. The power-delay product improves from 13.6J to 0.141J for DenseNet-121 and from 10.3J to 0.106J for ResNet-18, demonstrating its suitability for energy-harvesting systems.

Paper Structure

This paper contains 10 sections, 6 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Proposed ATM-Net
  3. Quantization Synergy
  4. Multi-Exit CNN Models
  5. Target Networks
  6. Proposed Energy-Aware Task Scheduler (EATS)
  7. Results
  8. Accuracy
  9. Performance
  10. Conclusion

Figures (4)

  • Figure 1: Proposed architecture, including ATM-Net and the energy-aware task scheduler.
  • Figure 2: Design flow of the proposed ATM-Net architecture.
  • Figure 3: Accuracy trends for DenseNet-121 across different datasets.
  • Figure 4: (a) The charging rate. The (b) quantization level, (c) system's energy at runtime, and (d) behavior of the system w.r.t exit points.