Assinar:
Postar comentários (Atom)
skip to main |
skip to sidebar
teste
The algorithm proceeds by cyclical updates of the individual components
$g_{\mu}(\mathbf{w}|\boldsymbol{\xi})$ refining the parameters of
$\boldsymbol{\xi}$ until converging sufficiently close to a
fixed point. To update the $\nu$ th
term, one defines the so - called {\em tilted} distribution which is
built by removing the $\nu$-th term of the approximate distribution and putting in its place the original likelihood, i.e.
\begin{equation}\label{definition-Q}
Q_{\nu}( \mathbf{w}) = \frac{1}{Z_\nu} \frac{ q( \mathbf{w}| \boldsymbol{\xi} )}{g_{\nu} (\mathbf{w}| \boldsymbol{\xi} )}
F\left( \frac{\mathbf{w} \cdot \mathbf{s}^{\nu} }{\sqrt{N}}\right),
\end{equation}
where $Z_\nu$ is a normalization constant. The new updated parameter set $\boldsymbol{\xi}^{new}$
is obtained by minimising the KL divergence
\begin{equation}\label{KL-Q-q-new}
KL \left( Q_{\nu}(\mathbf{w}) \Big|\Big| q (\mathbf{w}| \boldsymbol{\xi}^{new}) \right). %= KL \left( \frac{1}{Z_{Q_{\nu}}} \frac{q (\mathbf{w}| \boldsymbol{\xi}) }{ g_{\nu} (\mathbf{w}| \boldsymbol{\xi}) } F\left( \frac{\mathbf{w} \cdot \mathbf{s}^{\nu} }{\sqrt{N}}\right)
%\Big|\Big| q^{new} (\mathbf{w}| \boldsymbol{\xi}) \right).
\end{equation}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\subsection*{EP Algorithm}
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
Seguidores
Quem sou eu
- Marcelo
- Mineiro. Professor de Sociologia na Rede Estadual de S.P. Mestrando em Educação pela Unicamp.
Nenhum comentário:
Postar um comentário