MobiCom'26

Thyaga Nandagopal

Head of Emerging Technologies Group, Samsung Research America

Title: The Undocumented Interface: Split Inference on a Billion Phones
Abstract: Before 2028, the installed base of generative-AI-capable smartphones will likely exceed a billion units. Each of these phones splits its AI features between the device and the manufacturer's servers, reached over the wireless interfaces. The phone decides, per request, how this split is handled. It is less known what that decision is based on, or what the phone does when the radio conditions turn against it. This talk poses that as an open question: does either side of this split even know the other exists? Drawing on recent research across on-device inference, digital health, communications, and robotics at Samsung Electronics, this talk argues mobile AI needs what real-time robotics already has: measured behavior under sustained load, a defined response to a missed deadline, and a way for the two halves of a split computation to see each other.


Danijela Cabric

Professor, IEEE Fellow, Electrical Engineering Department, University of California, Los Angeles

Title: Learning for Spectrum Awareness, Spatial Intelligence, and Robustness to Hardware Impairments Abstract: Next-generation wireless systems are evolving beyond the pursuit of higher data rates toward intelligent radio networks capable of supporting dense user connectivity, improved spectral efficiency, and integrated sensing and communication. Under this vision, wireless infrastructure is expected not only to exchange information but also to sense, localize, and adapt to the surrounding environment.  In this talk, I will present recent research on integrating intelligence into wireless systems through three representative applications. First, I will discuss AI-driven approaches for wideband spectrum sensing that combine signal detection, classification, and RF fingerprinting using complex-valued neural networks. Second, I will present learning-based techniques for user localization, tracking, and spatial prediction based on deep learning, hypernetworks, and in-context learning. Third, I will discuss in-context learning approaches for sample-efficient adaptive self-interference cancellation, enabling robust full-duplex operation under realistic hardware impairments and time-varying conditions. Finally, I will introduce ongoing research on learning-enabled channel estimation and super resolution angle-of-arrival estimation in large-scale MIMO systems.

Kailash Narayanan

Senior Vice President - Keysight Technologies, President - Keysight’s Communications Solutions Group