张钧凯 助理教授(西安交通大学)

马俊杰 副研究员

Implicit Layer Empowered Deep Learning Networks for 6G Adaptive Channel Estimation

12月27日 (周六) 14:00-15:00，思源楼615

With the evolution toward sixth-generation (6G) wireless systems, artificial intelligence is set to become a core driver of intelligent and adaptive mobile communications. A critical challenge lies in achieving accurate channel estimation—especially in dynamic environments—under the tight computational constraints of user equipment. While existing AI-based methods often rely on stacking multiple deep learning layers, they struggle to adapt flexibly to varying channel conditions and resource limits.

In this talk, we introduce ICENet—an Implicit Channel Estimation Network designed for adaptive, efficient, and accurate channel state information recovery in 6G scenarios. By moving from traditional explicit deep learning architectures to a lightweight implicit network design, ICENet dynamically balances estimation accuracy and computational complexity based on real-time channel quality.

We demonstrate through numerical experiments that ICENet not only matches or exceeds the accuracy of conventional stacked networks but does so with significantly lower memory overhead. Importantly, we also delve into the training dynamics of implicit networks, analyzing key factors in forward and backward propagation, and propose a Jacobian regularization strategy that ensures stable convergence.

This talk will explore how implicit learning architectures can pave the way for scalable, adaptive, and hardware-friendly AI solutions in future wireless systems—enabling intelligence that is both powerful and practical for 6G.