How Can Integrated Sensing and Communication Enable Generalizable Learning in 6G Networks?

Terahertz (THz) communications offer ultra-high data rates and large bandwidth, making them a strong candidate for 6G wireless networks. However, accurate channel prediction using deep learning (DL) techniques is crucial for maintaining reliable connectivity, especially given the susceptibility of THz signals to blockages. Existing DL-based channel prediction models are typically developed under the assumption of stationary channel statistics. In practice, however, channel statistics are non-stationary due to environmental dynamics, particularly changes in user density. These variations introduce a phenomenon known as concept drift, where the statistical properties of the channel change over time, resulting in a severe degradation in model performance ranging from 4% to 51%, as we demonstrate. To overcome this limitation and develop robust channel prediction models, we explore integrated sensing and communications (ISAC) as an effective solution. Specifically, we propose a system that leverages the impact of room occupancy on THz channel gain to estimate user density. Based on this estimation, we dynamically select a DL model customized for the identified density level. Our approach offers two key advantages: 1) it eliminates the need for separate sensing hardware by utilizing the inherent characteristics of the wireless channel and 2) it effectively mitigates concept drift and achieves generalization regardless of user density, ensuring stable prediction performance across varying environmental conditions.