In MLS course 2’s Practice lab, the error in the cross-validation step in Neural Network was much lower when the input data was a polynomial (i.e., features were x,x^2…) rather than a linear data (only x). I expected the errors to be same since the NN can learn and develop non-linear models per se right?
Can someone please throw light on this?
Well, maybe you have trained the neural network enough, or with the polynomial maybe during training, it sees a different part of the data than when using the neural network!
Check whether both solutions used a normalized data set.
Normalizing the data will greatly impact the cost value.
Cost isn’t an absolute measurement, it’s relative to the specific dataset.