A new algorithm for efficient distributed deep learning by US army researchers
The U.S. Army Combat Capabilities Development Command's Army Research Laboratory researchers developed algorithms that facilitate distributed, decentralized and collaborative learning capabilities among devices, avoiding the need to pool all data at a central server for learning. "There has been an exponential growth in the amount of data collected and stored locally on individual smart devices," said Dr. Jemin George, an Army scientist at the U.S. Army Combat Capabilities Development Command's Army Research Laboratory. "Numerous research efforts, as well as businesses, have focused on applying machine learning to extract value from such massive data to provide data-driven insights, decisions, and predictions."
However, none of these efforts address any of the issues associated with applying machine learning to a contested, congested and constrained battlespace, George said. These battlespace constraints become more apparent when the devices are using deep learning algorithms for decision-making due to the heavy computational costs in terms of learning time and processing power.
"This research tries to address some of the challenges of applying machine learning, or deep learning, in military environments," said Dr. Prudhvi Gurram, a scientist who contributed to this research.
Earlier, they developed a new technique to significantly decrease the communication overhead, by up to 70% in certain scenarios, without sacrificing the learning rate or performance accuracy. The researchers developed a triggering mechanism, which allowed the individual agents to communicate their model with their neighbors only if it has significantly changed since it was last transmitted. Army researchers are investigating how this research can be applied to the Internet of Battlefield Things, incorporating quantized and compressed communication schemes to the current algorithm to further reduce the communication overhead.
(Content Courtesy: https://www.eurekalert.org/pub_releases/2020-02/uarl-ard021820.php)