This should be fairly easy, but for some reason i'm having hard time getting it to work and I've spent a long time trying to figure it out myself.

In the last paragraph of page 4 of the original Sum-Product Networks paper the authors described how to compute the posterior marginals of sum nodes, i.e

$$P(Y_k = i | e) propto w_{k,i} frac{partial S}{partial S_K}$$

, where Y_k is a sum node and i is an edge to one of its child nodes.

Let's assume that I've a very simple network like this one:

Simple SPN

and I want to compute the marginal of the sum node given an evidence (which can be either: e=x or e=not_x). The partial derivitev of the sum node should be 1 because it's the root:

$$frac{partial S}{partial S_{sum1}} = 1$$

then the marginals should be:

$$P(sum = 1 | e) propto w_{sum,w_1} $$

$$P(sum = 2 | e) propto w_{sum,w_2} $$

(where 1 and 2 referees to the edges)

My problem is that these two equations don't depend on the evidence and they give the same results regardless of what the evidence is. I tried with different structures and I keep getting the same results.

How can I correctly compute the posterior marginal of the sum nodes?

asked Mar 11 '14 at 22:46

TinyProton's gravatar image

TinyProton
51347

Be the first one to answer this question!
toggle preview

powered by OSQA

User submitted content is under Creative Commons: Attribution - Share Alike; Other things copyright (C) 2010, MetaOptimize LLC.