The problem of compressing a source and reconstructing it at the decoder with respect to some distortion metric is well-studied in information theory. However, in some settings, we may be interested in more than just reconstructing the compressed source. For example, in an image compression problem we might wish to classify the image on top of reconstructing it. To address this, we model the problem as a single-shot joint direct-indirect source coding one. We measure the quality of both the inference and reconstruction tasks using separate distortion criteria and study attainable rate distortion region.

follow this research