next up previous
Next: Hierarchical Nonlinear Variance Model Up: Example structures Previous: Variance Neurons

Linear Independent Factor Analysis

The addition and multiplication nodes can be used for building an affine transformation from Gaussian source nodes $ \mathbf{s}(t)$ to Gaussian observation nodes $ \mathbf{x}(t)$. This corresponds to linear factor analysis. With an independent mixture-of-Gaussians prior for each of the sources, the model corresponds to linear independent factor analysis [2]. Figure 3 shows how switches can be used to build such a prior for a source $ s(t)$. A variance neuron is used in order to prevent multiple paths from the discrete node to the source.

Figure 3: A mixture-of-Gaussians prior for $ s(t)$ is achieved using switches.
\begin{figure}\begin{center}
\epsfig{file=mog.eps,width=0.18\textwidth} \vspace{-6mm}
\end{center}
\end{figure}



Harri Valpola 2001-10-01