Combine error-linked selection with full-collection diversity

Investigate whether combining prediction-error information from the target model with diversity-based selection across the entire image collection improves active image selection.

Background

The proposed method links internal representations to predicted error types and their expected effects on the evaluation metric, whereas diversity-based methods seek broad coverage of an image or feature space. The results varied across datasets and seed conditions, and the authors identify as unresolved whether these complementary signals should be combined across the full candidate collection.

References

The combination of error-based selection with diversity-based selection across the entire image collection remains to be tested.

— From internal representations to model improvement through prediction errors  (2609.35449 - Nakaya et al., 28 Sep 2026) in Discussion, paragraph beginning “The performance of the proposed method relative to the external-feature”