Zygarde Task Scheduler Boosts Accuracy, Performance of TinyML Tasks Running on Battery-Free MCUs
Implemented on a TI MSP430 microcontroller, the new scheduler reduced execution times of common DNN tasks by up to 26 percent.
A team of computer scientists has unveiled a new soft-real-time task scheduling framework, which could unlock new performance from tinyML platforms running on microcontrollers β reducing execution time by as much as 26 percent while boosting inference accuracy by a claimed 21 percent.
"In this paper, we propose a time-, energy-, and accuracy-aware scheduling algorithm for intermittently powered systems that execute compressed deep learning tasks that are suitable for MCUs and are powered solely by harvested energy," the team writes in the paper's abstract. "The sporadic nature of harvested energy, resource constraints of the embedded platform, and the computational demand of deep neural networks (even though compressed) pose a unique and challenging real-time scheduling problem for which no solutions have been proposed in the literature."
"We empirically study the problem and model the energy harvesting pattern as well as the trade-off between the accuracy and execution of a deep neural network. We develop an imprecise computing-based scheduling algorithm that improves the schedulability of deep learning tasks on intermittently powered systems. We also utilize the dependency of the computational need of data samples for deep learning models and propose early termination of deep neural networks. We further propose a semi-supervised machine learning model that exploits the deep features and contributes in determining the imprecise partition of a task."
Those algorithms are implemented as a framework dubbed Zygarde, and the team claims considerable advantages from its use: Deployed on a Texas Instruments MSP430FR5994 microcontroller, the latest testing on Zygard across four data sets and six real-life audio and computer vision use-cases showed a decrease in execution time by up to 26 percent, an increase in inference accuracy of up to 21 percent, and the ability to schedule between nine and 34 percent more tasks than traditional schedulers.
The team's latest work has been presented at the ACM International Conference on Pervasive and Ubiquitous Computing, but is not yet publicly available; a paper detailing an earlier implementation of Zygarde is available under open-access terms on arXiv.org.