This paper introduces variational message passing. Formula (8) is based on Fig 1.

Formula (a) is $\ln Q^*_j(H_j)=\langle\ln P(H_j\mid\vec{pa_j})\rangle_{\sim Q(H_j)}+\sum_{k\in ch_j}\langle\ln P(X_k\mid pa_k)\rangle_{\sim Q(H_j)}+\text{const}$

$\langle\cdot\rangle$ is said to be an expectation. But I don't know what exactly the first term on RHS look like.

Is the first term $=\sum_{\vec{pa_j}}P(\vec{pa_j})\ln P(H_j\mid\vec{pa_j})$ or $\sum_{\vec{pa_j}}\ln P(H_j\mid\vec{pa_j})$? I think the first guess is more reasonable. However, if that is the case, what if the parents nodes have their parents? Do we still use $\sum_{\vec{pa_j}}P(\vec{pa_j})\ln P(H_j\mid\vec{pa_j})$ or $\sum_{\vec{pa_j}}P(\vec{pa_j}\mid\vec{ppa_j})\ln P(H_j\mid\vec{pa_j})$? where $ppa_j$ denotes the parents of parents.

And, my understanding for the second term on RHS is $\sum_{k\in ch_j}\sum_{\vec{cp_k^{(j)}}}P(\vec{cp_k^{(j)}})\ln P(X_k\mid\vec{pa_k})$.

Are my understandings correct? Thanks!


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.