The future of AI is distributed: Peer-to-peer learning and multi-agent AI at the edge

Description

The field of artificial intelligence has made great strides over the past 10 years, benefiting greatly from the adoption of GPU technology and a rising flood of training data. Old algorithms can finally show their full power, and practitioners are able to be more creative and innovative when developing new architectures to solve AI problems. However, these gains do not transfer well to smaller systems that lack the computational resources of large, monolithic models, such as typical IoT devices, smartphones, and other highly portable technologies. Bringing AI to these devices necessitates compromises in terms of responsiveness, connectivity, and model fidelity. New approaches such as federated learning and other distributed frameworks show promise, but building these systems requires us to make very real trade-offs in where and how data is collected, stored, transferred, and ultimately incorporated into the model and in how the model is deployed and adapted to fit within the constraints of edge computing devices.

Noah Schwartz explores the most recent advances in distributed and cooperative learning systems, including federated learning systems for real-world, edge-based AI. Along the way, Noah discusses topics and methods such as asynchronous learning of distributed representations, multi-agent federation in both centralized and peer-to-peer topologies, and online learning algorithms that enable true learning in edge-based systems—demonstrating concepts and solutions live using the Quorum AI Framework.