hominot / metric_learning

0 stars 0 forks source link

Importance sampling이 Gradient variance reduction 효과가 있음을 보이는 실험 디자인 하기. #57

Open hominot opened 5 years ago

hominot commented 5 years ago

distance metric learning의 단점으로, 이미지 pair 전체를 사용하지 못해 learning process가 stable하지 않다고들 함. (3개 이상의 reference가 있음)

실제로 random pair sampling으로 학습해보면 제대로 학습이 안됨.

이 "unstable"한 정도를 mini-batch gradient 벡터의 variance로 측정할 수 있음 (이것도 reference 있음)

importance sampling을 하면 gradient 벡터의 variance가 reduce가 될텐데, 이것을 이쁜 그래프로 보여줄 수 있는 실험을 디자인해보자.