1. Adversarial Training
to avoid overfitting
NBME top#2 Solution and Discussion
https://www.kaggle.com/competitions/nbme-score-
clinical-patient-notes/discussion/323085
Feedback top#1で活用されたが,NBMEでは
“Although its CV score was quite higher than the one I selected above, both its public LB score and private LB score were lower.
It seems that my way of doing pseudo labeling was better. It may be that being quite new to these techniques I didn't tune
them correctly. I will try them in future competitions for sure.”
とあるので汎用性についてはさらなる実装と議論が必要か。
2. Adversarial Training
Inputs Perturbation
into “Local” worst-case
Weights Perturbation
into “Global” worst-case
Gradient-based Adversary Not gradient-based
Need Labels Not need Labels
FGM, SiFT
VAT, TRADES,
SMART
MART
AWP