Efficient Transmission and Recovery of Salient Information

The thesis concerns fundamental problems that emerge from the proliferation of devices in increasingly distributed systems. As resources migrate from centralized clouds to a heterogeneous edge-cloud continuum, payloads from inference requests and telemetry required for monitoring overwhelm networks. Existing solutions cannot adequately meet the demanding requirements of edge-cloud systems as they are too complex for adoption, report results that are virtually impossible to reproduce, or do not address root causes. We follow a strict bottom-up approach, focusing on simple methods that yield significant improvements with little effort to integrate into existing systems. A minimally opinionated reference architecture emphasizes legacy compatibility and the importance of observability for automated decision mechanisms, such as scheduling. A rigorous empirical methodology for Machine Learning research in edge-cloud systems demonstrates the importance of simplicity when designing methods that are expected to run in complex systems. A training algorithm and novel knowledge distillation approach for task-agnostic compression achieves significant rate reductions, outperforming even the most competitive baseline. The base model consists of an encoder with just 140,000 parameters and is efficient enough for the latency penalty from the encoding and decoding overhead to be more than offset by the reduced transmission costs. An extension handles networks where connectivity is only intermittently available and throughput is favored, increasing downlinkable data volume by over two orders of magnitude. Finally, a method that augments algorithms for constructing statistical summaries for monitoring improves their coding efficiency. The augmentation provably maintains the base algorithm's underlying mathematical guarantees and compatibility with legacy systems. Together, the contributions establish a principled foundation for scalable, transparent, and coding-efficient systems that recover and transmit only the most salient information under stringent constraints for emerging paradigms in distributed computing.