Kenichi Higuchi

Collecting more errors doesn't guarantee better models: targeting only the high-impact failure modes signaled by a network's own internal representations outperforms standard rarity-based active learning while saving annotation budget.