During the development of the project we used accuracy and cross-validation to compare different methods and models. Will we be penalized (in terms of grading) for using accuracy instead of loss?
Performing model selection (e.g., using cross validation) based on accuracy (or sometimes based on F1-score when classes are imbalanced) for a classification task is in general a better idea than based on a loss. What we usually care about in the first place is accuracy, while we minimize a loss during training rather as a proxy since in general we can't efficiently minimize directly accuracy.
So doing cross validation based on accuracy is the right thing to do.
Comparing different models
Good afternoon,
During the development of the project we used accuracy and cross-validation to compare different methods and models. Will we be penalized (in terms of grading) for using accuracy instead of loss?
Thank you
Hi,
Performing model selection (e.g., using cross validation) based on accuracy (or sometimes based on F1-score when classes are imbalanced) for a classification task is in general a better idea than based on a loss. What we usually care about in the first place is accuracy, while we minimize a loss during training rather as a proxy since in general we can't efficiently minimize directly accuracy.
So doing cross validation based on accuracy is the right thing to do.
1
Add comment