UPENN | ML Questions for GATE DA

Question

In fitting some data using radial basis functions with kernel width $σ$, we compute training error of $345$ and a testing error of $390$.

(a) increasing $σ$ will most likely reduce test set error

(b) decreasing $σ$ will most likely reduce test set error

(C) not enough information is provided to determine how $σ$ should be changed

Answer 1

I think Option (A) is correct.

training data has less error compared to test set error and test set has high error compared to training that means it’s leading to overfitting.

As increasing sigma will make the RBF function to be more broader and less prone to overfitting. Lower sigma indicates that the curve is capturing all the noise and peculiarities of the training data and more chances to get overfit. That’s why increasing sigma will lead to reduce test set errror.

Comments

the correct option is C as per the answer key I have...may be we can think it over how and why