Poster
in
Workshop: Information-Theoretic Methods for Rigorous, Responsible, and Reliable Machine Learning (ITR3)

Single-Shot Compression for Hypothesis Testing

Fabrizio Carpi · Siddharth Garg · Elza Erkip


Abstract:

Enhanced processing power in the cloud allows constrained devices to offload costly computations: these remote executions call for a new communication paradigm that optimizes performance on the computational task within a rate constraint. We consider a simple binary hypothesis testing scenario where the transmitter performs fixed-length single-shot compression on data sampled from one of two distributions; the receiver performs a hypothesis test on multiple received samples to determine the correct source distribution. We formulate the task-aware compression problem as finding the optimal source coder that maximizes the asymptotic performance of the hypothesis test on the receiver side under a rate constraint. We propose a new source coding strategy that outperforms universal fixed-length single-shot coding scheme for a range of rate constraints.

Chat is not available.