Observations and likelihoods#
The likelihood entry of a modality gives the distribution of a reading given the state factors
it observes. It accepts:
a table over discrete readings:
ConditionalCategorical, orConditionalDirichletfor a table that is learned;a linear reading with Gaussian noise:
LinearGaussianObservation, whose noise may vary over the observed state through a noise map;a neural conditional:
MLPConditional, or aPredictorHeadwith a named loss;a reference to a shared table:
Parameter(see Shared parameters, emissions and recognition models).
Linear-Gaussian readings#
|
|
A Gaussian emission over categorical parents alone, |
Noise that depends on state#
These are parametrisations of an observation’s noise, not factors of the model.
An observation noise that varies over the observed state: a module that gives |
|
An observation noise that varies over the observed state: |
|
A rectangular grid over some coordinates of a continuous state factor's state. |
|
The probability vector over the cells of |