Design-Space Exploration of Distributed Neural Networks in Low-Power Wearable Nodes

2025-09-17

AI Summary

Wearable AI systems suffer from high node power consumption and frequent recharging, severely limiting practical deployment. To address this, we propose DistNN, a distributed neural network framework that jointly optimizes energy efficiency and inference accuracy by collaboratively partitioning computation between ultra-low-power wearable endpoints and high-performance centralized devices. DistNN introduces a novel energy-efficiency figure-of-merit (FoM) to guide optimal computational split decisions and integrates low-precision fixed-point hardware accelerators to sustain model performance under extreme power constraints. Evaluated on image reconstruction and denoising tasks using CNNs and autoencoders, DistNN achieves SSIM scores of 0.90 and 0.89, respectively. The system delivers a 1000× energy-efficiency improvement over GPU-based implementations and reduces average power consumption by 11× compared to state-of-the-art ML-dedicated ASICs. DistNN establishes a new paradigm for real-time, scalable edge-aware wearable AI.

Technology Category

• Machine Learning: Hardware-aware ML
• Search and Optimization: Distributed Search
• Constraint Satisfaction and Optimization: Distributed CSP/Optimization

Application Category

• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs

Abstract

Wearable devices are revolutionizing personal technology, but their usability is often hindered by frequent charging due to high power consumption. This paper introduces Distributed Neural Networks (DistNN), a framework that distributes neural network computations between resource-constrained wearable nodes and resource-rich hubs to reduce energy at the node without sacrificing performance. We define a Figure of Merit (FoM) to select the optimal split point that minimizes node-side energy. A custom hardware design using low-precision fixed-point arithmetic achieves ultra-low power while maintaining accuracy. The proposed system is ~1000x more energy efficient than a GPU and averages 11x lower power than recent machine learning (ML) ASICs at 30 fps. Evaluated with CNNs and autoencoders, DistNN attains an SSIM of 0.90 for image reconstruction and 0.89 for denoising, enabling scalable, energy-efficient, real-time wearable applications.

Problem

• Optimizing energy efficiency in wearable neural networks
• Distributing computations between nodes and hubs
• Maintaining accuracy with low-power hardware design

Innovation

• Distributed Neural Networks between nodes and hubs
• Figure of Merit for optimal energy split point
• Low-precision fixed-point arithmetic hardware design