So the lecture explains how to choose polynomial features for regression and how to a choose neural network architecture based on MSE. But I have a more fundamental question. For any given supervised ML problem, how do you know whether to use a simple regression model or a neural network? If this is already addressed in lecture I apologize; I wasn’t able to find a clear answer to the question.
Start with the simplest possible model. If you don’t get “good enough” results, then start adding complexity.
Is there any way to know if a model won’t be good enough without actually having to test it?
What I’ve learned, is regression and NN are different types of models, regression is good but is not good enough, it has some disadvantages:
you need to do manual feature engineering in contrast NNs do that automatically
also one major point that if you collect more data a large data and train your regression model on it you will find that the accuracy doesn’t getting better, it’s like that regression models have a limit level even if you have more data, in contrast in NNs if you have large data you can simply use a complex NN and you will find that the accuracy of the model getting better.
and finally in regression models they don’t support transfer learning.
simply always use NNs regression models are a basics models.
and good luck.
1 Like
No.