VICAL Framework Reduces Prediction Variance for Long-Tailed Recognition
September 7, 2026
VICAL shifts the focus of multi-expert models from maximizing expert diversity to reducing prediction variance via Self-Consistency Learning and Deep Ensemble Distillation. This approach mitigates overfitting on tail classes by smoothing the local loss landscape.
HOW THIS AFFECTS YOU
●
researcherYou may achieve better ensemble accuracy in long-tailed datasets by prioritizing variance reduction over diversity.