The role of noise
I'm having trouble understanding the role of noise. The generalization bound depends on N, the VC dimension of H, and delta.
I notice that in later lecture slides, noise forms an explicit term in the bias-variance decomposition, i.e. more noise increases the expected E_out (apologies for referring to slides that haven't been discussed yet).
Why doesn't it feature in the generalization bound? Is it because it is captured in the E_in term, i.e. more noise will increase our training error? In earlier lectures, N was written in terms of the growth function, to see how much data we need; and a rule of thumb was given that says N >= 10*VCdim. I'd like understand quantitatively how our need for data grows with noise, but I don't see how to do this using the generalization bound or bias-variance.
|