left: setting; right: tricks
on each gpu, there's only one half of all the neurons
training classification early -> more gradient in such a deep NN
[图片上传失败...(image-c8f98-1572805673386)]
from test error -> deeper is worse -> overfitting ???
from training error -> deeper is worse -> not overfitting !!! -> optimization problem ??? -> deeper means harder to train ???
cifar -> smaller than imagenet
stack MLP on top of the conv
local network -> precursor of googlenet
网友评论