PhD seminars – 5th, december 2022
Scientific presentation: Harikrishna Kuttivelil – Diana team
Exploring Community Structures in Decentralized Edge Intelligence
Abstract: As the number of connected devices continues to increase exponentially and intelligent applications and services continue to permeate these devices, the network edge has become a critical space for communication and intelligence infrastructures. The symbiotic relationship at the intersection of edge computing and artificial intelligence has resulted in the emergence of a new field, edge intelligence, gaining prominence in many domains, including IoT, sensor networks, autonomous vehicles, and more. Yet, intelligence networks at the edge face unique challenges compared to those at the network core, including network volatility, instability, and other communication challenges.
In this research, we consider the decentralization of intelligence networks as an enabling technology for edge intelligence while being cognizant of both the networking and learning aspects of such systems. To this end, we aim to exploit the evolving network and application affinities of the agents in a decentralized learning network to dynamically form communities of collaboration, in which agents collaborate on training and inference. In doing so, we are developing a framework to facilitate this process. As we apply this framework to real-world applications such as next-word prediction on mobile devices and environmental monitoring on heterogeneous sensor networks, the question arises: how can we evaluate such systems? As such, we also aim to answer this question by developing simulation platforms and tools that allow us to model both applications and networks to properly understand the overall performance of these systems and the interactions between the application and networking aspects.