Analog Coding: Theory and Applications

Zoom link

Passcode: 842382

Error-correcting codes are an essential building block of today’s communication systems that enable reliable end-to-end communication in the presence of inevitable noise in the physical layer. Moreover, coding theoretic tools have been recently utilized to enable resiliency, robustness, and data privacy in large-scale distributed computing and learning algorithms, among other applications in distributed storage systems, cryptography, etc. Most of the existing tools in the coding theory literature have been developed over discrete spaces with finite alphabet size. However, constructing codes in the analog domain offers significant performance improvements in various applications such as communication over massive wireless networks, distributed learning on the cloud, and prediction serving systems.

This thesis investigates the advantages of analog codes over the conventional codes constructed over discrete spaces in specific applications. In the first part, we provide a novel framework to study analog subspace codes for non-coherent communication in wireless networks and a corresponding family of algebraic codes that outperform existing constructions. The second part proposes privacy-preserving distributed learning algorithms that directly apply to real-valued data available in floating-point numbers. Furthermore, we propose a model-agnostic approach to prediction serving systems that offer inference results over neural networks as a service by leveraging analog coding techniques.  Our method is robust against malicious adversaries and significantly improves the end-to-end tail latency.

Chair: Professor Hessam Mahdavifar