0
$\begingroup$

Lets consider a simple derivation of GPs I've seen.

Given a dataset $ ( X, \vec{y} ) = (\vec{x}_i, y_i)_{i=0}^N $, we wish to find a function $ f $ such that $ y_i = \vec{w} \cdot \vec{\phi}(\vec{x_i}) + \epsilon_i $ where $ \vec{\epsilon} \sim \mathcal{N}\big( 0, \sigma^2 \mathbb{1} \big) $ is the gaussian noise. From various sources, it seems typical to assume $ w \sim \mathcal{N}(0, \mathbb{1}) $ (sometimes with a non-unit variance).

Then, $ \vec{y} \sim \mathcal{N}(0, \Phi \Phi^T + \sigma^2) $ with $ \Phi_{ij} = \phi_i(\vec{x}_j) $. This looks great because we don't have any parameters anymore, but this simple computes $ Pr(\vec{y} | X) $ and ignores the output of $ \vec{w} \Phi $!

I would imagine we would want to add priors for $ \phi $ and use $ Pr(\vec{y} | \Phi, X) = \mathcal{N}(w \Phi, \sigma^2 \mathbb{1}) $ (by inverting the noise distribution) for prediction.

How are we justified in ignoring $ \vec{w} \Phi $ and only using $ Pr(\vec{y}) $? What am I misunderstanding?

$\endgroup$

1 Answer 1

0
$\begingroup$

I think the papers I was reading were just confusing me. I was right, I just wasn't fully going through the derivation. Chapter 2 of C. E. Rasmussen & C. K. I. Williams, Gaussian Processes for Machine Learning resolved my issues by providing a simple derivation in Section 2.1.1

In case anyone actually reads this and is similarly confused. Fear not, as it's actually fairly simple. Maybe code will help you understand better. A little 16 line prediction algorithm I wrote is here that produces the below.

enter image description here

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.