ݺߣ

ݺߣShare a Scribd company logo
Understanding deep
learning requires rethinking
generalization (2017)
- 1/2
Chiyuan Zhang, Samy Bengio, Moritz
Hardt, Benjamin Recht, Oriol Vinyals
170322 @mikibear_ ?? ??
https://arxiv.org/abs/1611.03530
?? 1,
"??? ????? ????
???? ?? ????"
? ???, ???
'???? ? ?? ??'?
??? ??? ? ?????
??? ???? ????
????.
??? ??? ?????
???? ?? ??.
?? Generalization Error
??? ??? ?????
???? ?? ??.
?? Generalization Error
= |training error - test error|
??? Generalization error? ?? ??,
??? ???? ??? ??? ?? ? ?? ????.
??? ?? ?????, ??? ?? ???
??????.
?? ???? ???? ??? ??
??? ?? ?????, ??? ?? ???
??????.
?? ???? ???? ??? ??
??? ???
??? ?? ?????, ??? ?? ???
??????.
?? ???? ???? ??? ??
??? ???
?? ?? ????? ??? ?? ??? ??
??
??!
?? ????!
?? ? ????
"?? ???"
?!
"?? ???"
?!
?? ??? VGG, Inception, AlexNet?? ?? ???
????? ?? ??
????.
?? ?? ??? ???.
?, ??? ??? ?? ???
?? ??,
??? ?? ?????.
??? ?????.
??
??!
?? ????!
?? ????!
?? ????!
??
??!
???
??? ???????.
?? ??? ????.
"?? ???"
??...
?? ?? ???? ?? ???.
"?? ???"
??...
????? ?? ??
????.
??? ???? ???? ??
??? ? ?????.
??? ??? ????,
??? ???? ???? ? ? ?????
"?? ???"
'???'?? ???? ??
??? "??..."???.
??? ?? ?????
???? ??? ???
???? ??? ???? ??
??? ????? ?? ??
??? ?????.
?? ???? ?? ??????
"?? ???"
'???'?? ???? ??
???
!
?? ??? '???'?? ????,
??? ????? ???
???? ?? ???,
???? ?? ??? ???
????
???? ???? ??
???? ?????.
??? ??? ?? ??
???? ??? ???.
??? ????? ???.
"?? ???"
'???'?? ???? ??
???
!
?? ? ??
?? ??
??? ???? CIFAR10 ?????
??? (??? 10?)
('?? ???' - '??? ?? ???') ??
??? ?? ??? ????????.
? ??? ??? ?? ????, ??? ??? ???????.
??
??
Training error? 0?? ??, Testing error? 0.9? ???.
Training error? 0???,
Generalization error = Testing error
?, ?? ??? ??? ? ? ???? 0.9?
Generalization error? ??? ???.
Training error? 0???,
Generalization error = Testing error
?, ?? ??? ??? ? ? ???? 0.9?
Generalization error? ??? ???.
? ?? ???...
????? ? ???
??? ?? ???? ????.
????? ? ???
??? ?? ???? ????.
?? ??? ???? ??? ???
?????
'? ??? ??? ??? ???? ??? ??
????.
Data augmentation??
Dropout??
Weight decay?
BatchNorm??
Early stopping??
?? ?? ?? ? ?? ?????'
?? 2,
"?? ??? ? ???? ? ?
???? ??? ???"
Dropout?? weight decay? ??? ????
?????? ? '??'?? (...)
?, ?? weight decay? ? ????
??????. ?? ???? train data? ??
???? ????
'?? ?? ??? ???? ????
??? data augmentation?? weight
decay? batch norm ?? ? ??
???...?'
'?? ?? ??? ???? ????
??? data augmentation?? weight
decay? batch norm ?? ? ??
???...?
??? ??? ?? ?? ?????
???? ?? ??? ? ? ????
????? ?? ???? ????
???'
Understanding deep learning requires rethinking generalization (2017) 1/2
Understanding deep learning requires rethinking generalization (2017) 1/2
?, ??? ???? ??? ???
???? ??? ? ?????. ??,
batch norm? ?? ??? inception??
??? ??? ????.
????? ????? ? ?? ????
????, ????? ?????.
'??, ??? ???? ??
regularizer?? ???? ??? ???,
???? ????'
Understanding deep learning requires rethinking generalization (2017) 1/2
'? ???? ???, ?? regularizer?
???? ???? ??? ????
???? ??? ? ???? ??? ??
?????'
'? ???? ???, ?? regularizer?
???? ???? ??? ????
???? ??? ? ???? ??? ??
?????'
??? ??
????...
?? 3,
"??? ??? ??? ??? ???, ?? ?
Capacity? ??? ???"
?? 4,
"SGD? ??? ???? ? ???? ????
??? ??? ??? ???, ?? ????"
???
(?? ?? ???? ? ????)
?? ??? ??? ???? ???? ?? ?????!
@mikibear

More Related Content

Understanding deep learning requires rethinking generalization (2017) 1/2