Custom message operator that uses other divergences RRS feed

  • Question

  • Hi all,

    As I understand, in the standard EP, computing an EP message involves a projection to ExpFam (exponential family). The projection is in the KL sense i.e., proj[p] = arg min_q KL[ p || q].

    - If I change the divergence for the projection (much like in power EP) and assume that I can compute this new EP outgoing message easily, will the InferenceEngine still work if I do this in a custom operator, say, MyFactorOp ?
    - Also, can I mix divergences ? That is, I use some other divergence only for computing outgoing messages from my custom factor. However, other built-in factors in Infer.NET still use the usual KL. Will this give a valid result ?

    Monday, October 6, 2014 8:34 AM


  • A custom operator can do anything it wants.  It doesn't even have to correspond to divergence minimization. Whether the result is 'valid' is up to you.
    • Marked as answer by wittawatj Monday, October 6, 2014 10:13 AM
    Monday, October 6, 2014 9:30 AM