Abstract

As deep neural networks (DNNs) continue to revolutionize computing, a critical challenge emerges: how can we efficiently deliver these increasingly large models over bandwidth-constrained mobile networks? This challenge is particularly pressing as state-of-the-art DNNs become more specialized for specific times, locations, and tasks, requiring on-demand delivery to users' devices. Similarly, distributed learning scenarios require frequent exchange of model parameters across wireless networks. I will highlight that these problems are examples of `semantic communications’, where the goal is not to recover network parameters reliably, but to make sure they serve for the desired task at the receiver.

In this talk, I will first present our recent breakthroughs in efficient DNN parameter compression and transmission over wireless channels, demonstrating how these techniques can reduce bandwidth requirements while preserving model performance. Building on these foundations, I will then explore how our techniques extend naturally to implicit neural representations (INRs), which treat signals as continuous functions represented by neural networks. This innovative approach bridges traditional signal processing with modern deep learning, yielding remarkable results for practical applications. Specifically, I'll demonstrate how our INR-based methods achieve state-of-the-art performance in MIMO channel state feedback and image compression tasks, reducing transmission overhead while maintaining signal fidelity.