Isolate boundary learning from prediction collapse in contested-pair intervention

Determine whether the reduction in mutual information for the A220–A350 aircraft pair after targeted oversampling reflects genuine learning of the contested class boundary rather than collapse toward A350, using a fully specificity-isolated intervention that separates boundary-instance targeting from class-level data inflation.

Background

The framework diagnoses A220–A350 as a contested epistemic pair because the classes have negligible physical-size overlap but high ensemble disagreement. Targeted oversampling reduces the pair’s mutual information, but A220’s validation and test AP move in opposite directions, while predictions shift toward A350. This leaves unresolved whether the intervention genuinely sharpened the boundary or merely induced a systematic prediction collapse toward A350.

The paper identifies the needed resolution: an experiment that isolates boundary-instance targeting from class-level inflation and distinguishes genuine boundary learning from prediction collapse. This is a concrete unresolved validation problem for the contested-boundary remedy.

References

Because oversampling is at the image level, it also inflates the A220/A350 populations as a whole, so exposure and targeting are entangled; a fully specific test, isolating boundary instances without class-level inflation and separating genuine boundary learning from prediction collapse, is left to future work.

— Not All Confusion Is Equal: A Source-Aware Uncertainty Diagnosis for Fine-Grained Aircraft Detection  (2609.29959 - Huang et al., 24 Sep 2026) in Section 4.5, subsection “A caveat: a val/test divergence on A220”