Abstract:
"
Edge Intelligence based collaborative learning has been recognized as a trending
research area. It was noted that the majority of existing collaborative learning
approaches were designed in the context of servers or resources with relatively high
computational power. Therefore, it was identified that the design and implementation
of a system was required to utilize collaborative learning in resource constrained IoT
edge.
The existing collaborative learning approaches were critically reviewed and analyzed
in order to identify the most ideal collaborative learning approach for the resource
constrained IoT edge. The partitioned model training (DNN partitioning) approach was
identified as the most ideal collaborative learning approach for the IoT edge. Two
training architectures based on the partitioned model training approach were introduced
for the resource constrained IoT edge to facilitate collaborative learning in
environments with limited and adequate access to edge infrastructure. A lightweight
containerization mechanism was utilized to deploy the proposed collaborative learning
system and a hybrid deep learning model was utilized to demonstrate the system.
Furthermore, an iterative model training mechanism was integrated with the partitioned
model training approach to reduce the iterative communication overhead.
The test results and the evaluation of the research proved that the proposed Edge
Intelligence based collaborative learning system functioned efficiently with lower
CPU/memory consumption, lower power and energy consumption, lower operating
temperature and high model accuracy. Some of the future works identified were the
integration of a headless mode to perform collaborative learning, testing the
applicability of the approach in microcontrollers, testing on IoT devices with dedicated
GPU and the integration of a gradient compression mechanism.
"
