概念
distillation optimization
创建2026-04-15
更新2026-04-15
阅读量级1 分钟
概念导读

KL (Kullback-Leibler) divergence plays a central role in knowledge distillation as the primary loss function measuring the discrepancy between teacher and student output distributions. The on-policy-distillation-survey frames distillation methods through the lens of f-divergence measures, with KL divergence being the most commonly used instance. The choice b...

  1. Related
  2. References

KL Divergence in Distillation

KL (Kullback-Leibler) divergence plays a central role in knowledge distillation as the primary loss function measuring the discrepancy between teacher and student output distributions. The on-policy-distillation-survey frames distillation methods through the lens of f-divergence measures, with KL divergence being the most commonly used instance. The choice between forward KL, reverse-kl-distillation, and other f-divergences fundamentally shapes the bias-variance tradeoff in distilled models.

References