I am working on the lab on overfitting. For the regression example, it seems like the quadratic model works best. However, when I add more data points, it seems like a cubic model works better.

The same happens for the classification case. The ‘best fit degree’ seems to depend on the data points.

How do I decide which model works best?

Yes, that is exactly the point of the assignment.

You modify your model by using results from the validation set.

Do you mean we keep adding data points to the validation set until we see no change in the ‘best fit degree’ of the model? And then use that model for prediction?

No, that’s not what I mean.

You use the cost computed on the validation set for guidance about what degree gives the best results.