A Framework for Modeling Energy-Accuracy Tradeoffs in Neural Network-based Classification for Resource Constrained Embedded Systems.

摘要

Advances in the manufacturing of sensing devices and improvements in the capabilities of small, low power microcontrollers have enabled the adoption of self- contained data acquisition systems. While applications of such devices typically require that they be physically small, battery powered, and able to operate for extended periods of time, they also often require the wireless transmission of copious amounts of sensor data to a base station, where the raw data is processed into application-relevant information. The power requirements of such streaming can be prohibitive given the competing form factor and battery life requirements. Even if the data is stored in a local memory for subsequent download and offline processing, high bandwidth memory writes can be power intensive. It is therefore imperative that the bit rate of transmissions and memory writes be dramatically reduced in order to meet the requirements of many self- contained data acquisition applications.;This work explores techniques that allow for embedding intelligence, in the form of Artificial Neural Network Classifiers (ANNs), directly on the processor employed in self- contained data acquisition nodes. On-node classification of sensed data is shown to reduce the effective bit rate (i.e. only certain classes of data are deemed relevant), and therefore the energy consumption for transmission and/or storage. Among the challenges addressed is finding algorithms and techniques -- currently implemented on back-end workstations --- that will execute efficiently in the constrained computational environment typically found on self-contained data acquisition systems. Additionally, models are presented that enable systematic comparison of classifier performance based on synthetic data generation techniques. These models are shown how they may be used to predict 4 real-world classifier performance and to facilitate design-time and run-time tradeoffs between energy consumption (processing energy and unnecessary transmission/storage of false positive classifications) and application fidelity (mistaken non-transmission/storage of false negative classifications). A methodical framework for quantifying these models is also derived and is applicable for general embedded computing environments.