Knowledge Distillation

KL-Based Divergences 给定两个离散分布$P(\mathcal C)$和$Q(\mathcal C)$,它们的KL散度定义为: $$ \mathcal D_{KL}(P\Vert Q)=\sum_{c\in\mathcal C}P(c)\log\frac{P(c)}{Q(c)} $$ 由于KL散度的不对称性:$\mathcal D_{KL}(P\Vert Q)\neq \mathcal D_{KL}(Q\Vert P)$,定义前向KL散度(forward KL)为$\mathcal D_{KL}(P\Vert Q)$,定义反向KL散度(reverse KL)为$\mathcal D_{KL}(Q\Vert P)$。 ...

 2025-11-01        4 min        Rs