Revisiting DNN Training for Intermittently-Powered Energy-Harvesting Micro-Computers

TL;DR

NExUME addresses intermittent power in energy-harvesting edge environments by integrating energy variability directly into both training and inference. The framework combines DynFit (intermittency-aware training with dynamic dropout/quantization and QuantaTask units) and DynInfer (intermittency-aware inference scheduling with task fusion) to adapt to real-time energy availability. Empirical results across multiple datasets and a new machine-status dataset show consistent accuracy gains (6-22%) with modest compute overhead and improved energy efficiency, highlighting practical impact for battery-free IoT and Industry 4.0 deployments. The work demonstrates a holistic, hardware-conscious approach to robust, energy-aware deep learning at the edge, with broad implications for sustainable intelligent systems.

Abstract

The deployment of Deep Neural Networks in energy-constrained environments, such as Energy Harvesting Wireless Sensor Networks, presents unique challenges, primarily due to the intermittent nature of power availability. To address these challenges, this study introduces and evaluates a novel training methodology tailored for DNNs operating within such contexts. In particular, we propose a dynamic dropout technique that adapts to both the architecture of the device and the variability in energy availability inherent in energy harvesting scenarios. Our proposed approach leverages a device model that incorporates specific parameters of the network architecture and the energy harvesting profile to optimize dropout rates dynamically during the training phase. By modulating the network's training process based on predicted energy availability, our method not only conserves energy but also ensures sustained learning and inference capabilities under power constraints. Our preliminary results demonstrate that this strategy provides 6 to 22 percent accuracy improvements compared to the state of the art with less than 5 percent additional compute. This paper details the development of the device model, describes the integration of energy profiles with intermittency aware dropout and quantization algorithms, and presents a comprehensive evaluation of the proposed approach using real-world energy harvesting data.

Figures (5)

• Figure 1: An example of variable QuantaTask in a matrix multiplication scenario. Depending on the available energy, the task (vector inner product) can be divided into multiple iterations such that each QuantaTask is guaranteed to finish given the energy availability. $E$ is available energy, and $E_b$ is the energy required to finish one inner product.
• Figure 2: Software-Compiler-Hardware Driven DynInfer Flow.
• Figure 3: Sensitivity and ablation study. DN is DynNAS, DF is DynFit, and DI is DynInfer.
• Figure 4: Hardware setup of NExUME using MSP-EXP430FR5994 as the edge compute, Adafruit ItsyBitsy nRF52840 Express for communicating, Energy Harvester Breakout - LTC3588 with supercapacitors as energy rectification and storage and a Pixel-5 phone as the host.
• Figure 5: DNN computation using ReRAM xBAR.