Single-Shot Rate Distortion Classification Trade-off
Oguzhan Kubilay Ulger
Abstract:
Motivated by the increasing interest in semantic communications, we investigate a goal oriented, single-shot lossy compression problem where the source has an underlying class. This intrinsic class can not be observed directly and has to be inferred from the source. Under our model, the encoder has to convey its observation in a single-shot and rate limited manner. The decoder has two tasks: to reconstruct the original signal under some distortion criterion and to infer the class of the source. We study different encoding strategies and their impacts on the classification error rate and distortion.