Chapter 9 Gradient Descent For Machine Learning

Question 1

HOW DOES PLOTTING COST VS TIME HELP US EVALUATE THE ALGORITHM? P41

Answer 1

A well performing gradient descent, decreases the error (cost) in each iteration, if it does not decrease, we should try reducing the learning rate.

Question 2

HOW MANY PASSES ARE NEEDED FOR STOCHASTIC GRADIENT DESCENT? P41

Answer 2

SGD often doesn’t need more than 1-10 passes through the training set to converge on good or good enough coefficients

Question 3

IS LINEAR REGRESSION SENSITIVE TO HAVING GAUSSIAN DISTRIBUTION? P47

Answer 3

Yes, Linear regression will make more reliable predictions if your input and output variables have a Gaussian distribution. You may get some benefit using transforms (e.g. log or BoxCox) on your variables to make their distribution more Gaussian looking