Tue, Nov 18 · 5:00 PM CET
Machine learning (ML) services have become deeply integrated into daily life and industrial systems, most of which are delivered over networks. The rapid surge in ML demand has led to soaring computational workloads and escalating data volumes, creating substantial challenges for traditional architectures that rely on ever more powerful servers and expanded network capacity. In-network computing emerges as a promising and scalable alternative, enabling packet-level processing directly within the data path and allowing early traffic termination. Yet, applying this paradigm to ML inference represents a new and largely unexplored frontier. Network devices are optimized for high-performance packet processing and forwarding, not for machine learning computation.
This talk will introduce the concept of in-network machine learning, its core technologies, and implementation strategies, including three general model mapping methodologies. As in-network machine learning is inherently a resource-constrained machine learning problem, two complementary deployment strategies, distributed deployment and hybrid deployment, will be presented. The talk will also showcase a set of in-network machine learning applications, such as anomaly detection, IoT traffic classification, and bot detection, among others. Finally, it will introduce a rapid prototyping framework for deploying in-network machine learning across diverse hardware targets, along with several open-source projects available for community use.
Speaker: Changgang Zheng received his PhD in Engineering Science from the University of Oxford. His research interests lie at the intersection of networking, in-network computing, and machine learning, with a focus on in-network machine learning and its applications across domains from cybersecurity to financial systems. His research has been published in venues such as ACM SIGCOMM Computer Communication Review (CCR), IEEE/ACM Transactions on Networking, and ACM CoNEXT, and has received multiple awards, including Best of CCR (SIGCOMM 2024), Best Paper Award at IEEE HPSR 2025, and Best Project in Data Science and AI-Enabled Solutions at Prototypes for Humanity 2024.
