Networks That Ask for Help: Uncertainty, guidance, and trust on the road to AI-native 6G

#networking #engineering #wireless #security

Networks That Ask for Help: Uncertainty, guidance, and trust on the road to AI-native 6G

Zero-touch, self-configuring networks are a defining ambition of AI-native 6G. This talk explores a pragmatic path toward that goal, built on two simple design principles, namely (1) AI-algorithms complementarity: learned and classical components achieve more when composed together, each supplying what the other lacks (as when a classical solver draws global guidance from a language model while lending it algorithmic guarantees in return); and (2) calibrated escalation: systems to quantify their own uncertainty and ask (an oracle for labels, or a human operator for clarification) when that uncertainty warrants it, and less and less as confidence is earned. We develop these principles through three recent works from our NetAI group. First, an active machine learning framework for 6G, in which the network queries an oracle for the labels and samples it is most uncertain about, cutting data acquisition and annotation costs while improving mmWave throughput prediction. Second, an LLM-aided A* algorithm for non-geometric network graphs, where a classical solver asks a language model for guidance, thereby reducing search exploration while keeping imperfect guidance harmless. Third, an uncertainty-driven intent translation system, in which a fine-tuned LLM turns operator intents into configurations and uses two tiers of uncertainty, namely predictive uncertainty to decide whether to deploy or escalate, and token-level entropy to pinpoint what clarifying questions to ask. These threads rely on our open-source NWDAF-enabled 5G testbed, which extends Free5GC with event-driven analytics and an LLM/RAG agentic intent interface, grounding the vision in a system that runs today. Together, they suggest a design philosophy for trustworthy network autonomy: (1) compose generative models with classical guarantees, (2) quantify what the system does not know.

• Add Event to Calendar
  iCal
  Google Calendar

• Western Engineering
• London, Ontario
• Canada
• Building: SEB
• Room Number: 3102

---

• Contact Event Host
• Co-sponsored by Western Optimized Computing and Communications

• Starts 07 July 2026 04:00 AM UTC
• Ends 09 July 2026 04:00 AM UTC
• No Admission Charge

•