Generalized Doubly-Reparameterized Gradient Estimators
Matthias Bauer 1 Andriy Mnih 1
Abstract usually prefer unbiased estimators as they tend to be better-
behaved and are better understood. Lower variance is also
Efficient low-variance gradient estimation en-
preferable because it enables faster training by allowing
abled by the reparameterization trick (RT) has
...


雷达卡




京公网安备 11010802022788号







