A Study on the Energy Sustainability of Early Exit Networks for Human Activity Recognition

Abstract

The design of IoT systems supporting deep learning capabilities is mainly based today on data transmission to the cloud back-end. Recently, edge computing solutions, which keep most computing and communication as close as possible to user devices have emerged as possible alternatives to reduce energy consumption, limit latency, and safeguard privacy. Early-exit models have been proposed as a way to combine models with different depths into a single architecture. The aim of this article is to investigate the energy expenditure of a distributed IoT system based on early exit architectures, by taking human activity recognition as a case study. We propose a simulation study based on an analytical model and hardware characterization to estimate the trade-off between the accuracy and energy of early exit-based configurations. Experimental results highlight nontrivial relationships between architectures, computing platforms, and communication link. For instance, we found that early-exit strategies do not ensure energy reductions with respect to a cloud-based solution if the same accuracy levels are kept; nonetheless, by tolerating a 1.5% decrease in accuracy, it is possible to achieve a reduction of around 40% of the total energy consumption.