| | |
Summary: Investigations of Gaussian
Products-of-Experts Models
Felix Agakov
MSc in Arti cial Intelligence
Division of Informatics
University of Edinburgh
2000
Abstract
In 1999 Hinton introduced the Products of Experts (PoE) architecture for
density estimation, where individual expert probabilities are multiplied and
renormalized. In this work we consider speci c types of PoE models where
each expert is a Gaussian. Even though the resulting distributions of the
products in these cases are also Gaussians, it is shown that they may lead
to interesting models, such as probabilistic analogues of non-probabilistic
dimensionality reduction techniques. This makes products of Gaussians po-
tentially useful in a wide range of data analysis applications.
In rst two principal parts of the investigation we consider products
of constrained Gaussians (PoCG), elongated in all directions except one,
and products of probabilistic principal component analyzers (PoPPCA). We
|
