We have a parameterised function f (x; w) which we will use as a predictor—our esti-mated model of φ. When we do not wish to refer to any specific input x or parameter vector
w, we will use simply f , or f
i
to refer to a specific predictor. For any input x ∈ X, the
predictor will map it to an output in Y . The accuracy of its mapping (i.e. how closely it
models φ) is determined by two factors: firstly the exact functional form of f , and secondly
by the contents of the parameter vector w ∈ R
k
. We will consider the form of the function
in the next section; for the moment imagine it to be any arbitrary form, with the parameter
vector w also set arbitrarily; f (·; w) is now our predictor. For a particular (x, d) pair chosen
from t, we can quantify how well our predictor models φ with an error measure:
(f (x; w) − d)
2