Evaluate model-generated feedback and cycle-consistency training for closed-loop reconstruction

Evaluate whether incorporating model-generated feedback or cycle-consistency training into the SBNO framework improves recursive forward/inverse reconstruction while retaining randomly sampled observation masks.

Background

The paper evaluates SBNO in a closed-loop setting in which a prediction generated in one direction is reused as a partial observation for inference in the opposite direction. The results show directional compatibility but also recursive error accumulation, particularly in more difficult settings.

The authors suggest augmenting training with model-generated feedback or a cycle-consistency objective, but explicitly state that the effect of such an extension has not yet been determined. This constitutes an unresolved methodological question concerning whether closed-loop stability and reconstruction accuracy can be improved.