Determine whether longer retraining increases the performance gap between selection methods

Determine whether extending retraining increases the performance gap between the proposed error-linked image-selection method and external-feature $k$-center selection.

Background

The study compared the proposed selection method with external-feature kk-center selection under the main retraining schedule and under a longer 40-epoch schedule. Although the proposed method performed better under both schedules, the authors explicitly note that the observed difference does not establish whether longer training itself increases the gap between the methods.

References

The difference between these estimates does not establish whether longer training increases the gap between the methods.

— From internal representations to model improvement through prediction errors  (2609.35449 - Nakaya et al., 28 Sep 2026) in Discussion, paragraph beginning “In the BDD100K experiment with 40 epochs at every stage”