Task-aligned training objective for a sustaining codec

Determine whether optimizing a differentiable surrogate of the two-sided detection-agreement criterion, rather than a pixel-fidelity loss, is sufficient for a radar compression codec to reach a sustaining operating point.

Background

The learned codec evaluated in the paper was trained with a conventional rate-distortion objective based on squared reconstruction error and did not reach a rate satisfying the two-sided detection criterion. The paper therefore proposes replacing pixel fidelity with a differentiable surrogate of detection probability and false-alarm control.

The authors explicitly leave unresolved whether aligning the training objective with the operational evaluation criterion would actually produce a codec that sustains utility.