Appendix : Machine Learning Bias Versus Statistical Bias

is if and 0 if. This high variance may help to explain why there is selection pressure for weak (machine learning) bias when the (machine learning) bias correctness is low. The reason that statisticians are interested in (statistical) bias and variance is that squared error is equal to the sum of squared (statistical) bias and variance. Therefore minimal (statistical) bias and minimal variance implies minimal squared error. (Note that the error is the error relative to the target , not the error relative to the testing data. The difference between these two errors is random noise, known as irreducible prediction error.) Typically there is a trade-off between bias and variance — decrease in one usually causes increase in the other — but we have seen here that it is sometimes possible to minimize both simultaneously.