A latent variable, is a variable in a Bayesian network that is neither an input or output variable, and hence has no data associated with it during parameter learning.
Latent variables can either be discrete, or continuous.
Tip |
---|
Latent variables are also known as hidden variables. |
Discrete latent variables
An example of a discrete latent variable is the Cluster node in the mixture model network shown below.
Instead of a single multivariate Gaussian, we have made the model more flexible by allowing multiple multivariate Gaussians, as shown in the image below.
Continuous latent variables
An example of a continuous latent variable is the x node in the Kalman filter network shown below.
The continuous latent variable is used in this case to represent an unobservable process.
Tip |
---|
Models such as this are often used to track missiles. |
Tip |
---|
Continuous latent variables can also be used for dimensionality reduction (e.g. probabilistic PCA). |