Dual-Stage Reweighted MoE for Long-Tailed Egocentric Mistake Detection

DR‑MoE merges frozen and LoRA‑tuned ViViT features and uses classifiers—reweighted cross‑entropy, AUC loss, and label‑aware loss—to improve rare egocentric mistake detection. getnews.me/dual-stage-reweighted-mo... #egocentric #mistakedetection