Here is a list of how-to solutions for problems that we have encountered while doing research in the i-NRG Lab. … Read More
Decentralized Edge Learning: Paradigms and Applications of Collaborative Edge Intelligence
Introduction
As the Internet’s edge has begun to regain prominence in response to unprecedented advances in computing, storage, and wireless communication technology, wireless edge networks have grown increasingly more complex, not only due to the sheer number of connected devices but also due to the heterogeneity of those devices and the network technologies used to interconnect them. To satisfy the needs of emerging edge applications and the enormous amount of data they generate, computational resources have moved closer to the network edge, in a paradigm termed edge computing. At the same time, the pervasive growth of machine learning has not only benefitted from and encouraged this abundance of data but has also driven the development of applications on the network edge. This symbiotic relationship between networking and machine learning has enabled an emerging field of research, edge intelligence, fueling entirely new application paradigms, including autonomous vehicles, connected health services, and smart cities, while also introducing surfacing issues of security, privacy, resource utilization, and equity.
DEL, or Decentralized Edge Learning, aims to continue and further edge intelligence into the extreme peripheries of the network edge. In this research, we develop decentralized machine learning strategies to ensure proper privacy and security while also considering the heterogeneous networks these strategies will run on to ensure optimal network resource utilization and system performance. We explore various decentralized learning methods, including federated and gossip learning strategies, and understand the effect of their behavior from the networking perspective. We strive for tightly integrated machine learning and networking systems to optimize these methods. Our ultimate goal with this research is to open the door to more pervasive machine learning applications to be used in spaces they haven’t before, while still guaranteeing the safety of the end-users and the high efficiency of the computing and networking resources.
Current Efforts
Our current efforts in developing and extending edge intelligence address what we see as challenges in the space. We want to deploy full spectrum (cloud to edge) edge intelligence by enabling efficient fully decentralized learning where needed. We also want to provide a simulation solution for developing network-cognizant methods for edge intelligence. In addition to our efforts in developing systems and structures for full-spectrum edge intelligence, we are also developing extensible simulators for edge intelligence.
Frameworks for Structured Fully Decentralized Learning
As we explore decentralized learning at the edge, we consider the structures and flows of information that enable it to be more efficient and effective. To this end, we had proposed Community-Structured Decentralized Edge Learning (published at the ResilientFL workshop at ACM SOSP 2021), a paradigm of decentralized learning based on self-organizing affinity-based communities. We are continuing to build out the framework for real-world systems, while considering things like hostile network conditions, power and resource management, and security and privacy. In this process, we are also seeking out opportunities for implementation in real-world systems, including those in community health monitoring, smart city infrastrcture, and environmental and ecological monitoring.
More information coming soon.
Simulation Platforms for Edge Intelligence
An early version of our Decentralized Edge Learning simulator can be found on GitHub (GitHub – harikuts/del-step-sim). This simulator is functional and works well as a decentralized and group-based distributed learning simulator, and it has been modified to allow researchers to plug-and-play different neural network models (Tensorflow) and datasets.
While certain network conditions are considered in that first simulator, our next iteration of a more extensible simulator platform hopes to more fully consider the network; and thus we are pivoting to including the use of more powerful and standardized network simulators, such as OMNET++ and NS-3, while retaining the flexibility and modularity of the previous iteration.
More information coming soon.
People
Publications