I still do not get the difference between validation and test data... validation data is also a new set of data for the model right. which it has never seen before. so is test data. the result of the validation and test data will be more of less the same. As I've understood test data is just a 2nd layer of testing after validation data. is that correct ?
i have same thoughts with you. validation data is used for determining the best model to use after we train the data with different models. then we checked again that model with the test data to make sure if it is actually good or not. like double check. anyway, still a little confusing to me.
So it can be confusing but the real difference lies when you are not training and testing the model once. This might happen in a case where you are doing hyperparameter tuning where your training and testing data might get exposed to the model multiple times. But you hold out validation set till the end once the model is finalise and you figured out the hyperparameters you test it on validation to be sure that it is consistent with the test set. This is done because there is a chance in the process of tuning the model it actually starts remembering the training and testing sets and the performance improvement you are seeing will not be generalisable to new unseen datapoints.
Best video ever in regard to this content! Love it!
Very well explained! Thanks!
Yay, glad it was helpful :)
Really Effective
well explained. saved my life
What is the best percentage for all the 3 data to accept the model
讲得好！good！
I still do not get the difference between validation and test data... validation data is also a new set of data for the model right. which it has never seen before. so is test data. the result of the validation and test data will be more of less the same. As I've understood test data is just a 2nd layer of testing after validation data. is that correct ?
i have same thoughts with you. validation data is used for determining the best model to use after we train the data with different models. then we checked again that model with the test data to make sure if it is actually good or not. like double check. anyway, still a little confusing to me.
So it can be confusing but the real difference lies when you are not training and testing the model once. This might happen in a case where you are doing hyperparameter tuning where your training and testing data might get exposed to the model multiple times. But you hold out validation set till the end once the model is finalise and you figured out the hyperparameters you test it on validation to be sure that it is consistent with the test set. This is done because there is a chance in the process of tuning the model it actually starts remembering the training and testing sets and the performance improvement you are seeing will not be generalisable to new unseen datapoints.
Super clear
teacher can you share this presenantion?
thanks, very useful
thanks maam