A prototype-based ensemble head with explicit diversity regularization improves worst-group accuracy on nine subpopulation shift benchmarks, often matching or beating prior state of the art.
We observe that inference time increases only slightly as the number of prototypes increases, from 0.0031s with 15 prototypes to 0.0045s with 100 prototypes
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Diverse Prototypical Ensembles Improve Robustness to Subpopulation Shift
A prototype-based ensemble head with explicit diversity regularization improves worst-group accuracy on nine subpopulation shift benchmarks, often matching or beating prior state of the art.