Expressions

The simplest expression is the Parameter, which holds a value that is optimized by the Optimizer (typically, the Adam optimization algorithm is used). It is however possible to do a deterministic transformation of these parameters before using them as the defining parameters of a distribution.

An example of such a transformation is the Variational Auto Encoder. It does not define variational parameters for each data point (say, an image). Instead, a neural network is used to calculate parameters for the guides from the data point. In this way, the number of variational parameters is greatly reduced. Furthermore, the trained neural network can be used to obtain good approximations to the posterior distributions of hidden variables for new data points.