Next: Why simple methods fail
Up: BAYESIAN NONLINEAR FACTOR ANALYSIS
Previous: BAYESIAN NONLINEAR FACTOR ANALYSIS
The nonlinear factor analysis model described in publication V has the following structure:
- The mapping from factors to observations is modelled by an MLP
network with one hidden layer as described by (34).
- The factors can have Gaussian or mixtures-of-Gaussians models.
This corresponds to a nonlinear extension of linear factor analysis
or linear independent factor analysis, respectively. The factors
are assumed to be independent.
- Hierarchical models are used for describing the prior
information about the parameters. Gaussian distributions are used
- The variance parameters of the Gaussians are parameterised by the
logarithm of the standard deviation. This yields a roughly Gaussian
posterior probability for these parameters.