Undirected Graphical Model
An undirected graphical model, also known as a Markov random field (MRF), is a type of probabilistic graphical model that represents a set of random variables having a Markov property described by an undirected graph. These models are powerful tools for modeling the joint distribution of a set of variables encoded through graphs.
Structure and Components
In an undirected graphical model, the structure of the graph is constructed from nodes and edges:
- Nodes: Represent random variables that may be continuous or discrete.
- Edges: Indicate a dependency or probabilistic relationship between the connected nodes. Unlike directed graphical models, the edges in an undirected graphical model do not imply causality or direction.
This structure allows the model to capture arbitrary dependencies between variables, making it suitable for various domains such as image analysis, protein structure prediction, and social network analysis.
Mathematical Representation
The probability distribution of an undirected graphical model is represented as a product of potential functions (also known as clique potentials) over cliques in the graph. A clique is a fully connected subgraph within the larger graph. The probability of a particular configuration ( x ) of the graph's nodes is given by:
[ P(x) = \frac{1}{Z} \prod_{C \in \mathcal{C}} \psi_C(x_C) ]
Where:
- ( \psi_C(x_C) ) is the potential function of clique ( C ).
- ( Z ) is the partition function that normalizes the distribution.
Applications
Undirected graphical models are used in numerous applications across various fields:
- Computer Vision: They are utilized for tasks such as image segmentation and object recognition.
- Natural Language Processing: Used for language modeling and sequence predictions.
- Bioinformatics: Employed in modeling biological processes and protein interactions.
Connection to Other Models
Undirected graphical models are part of a broader family of probabilistic graphical models. They differ from Bayesian networks, which use directed edges and encode conditional dependencies. However, both types of models are unified under the framework of factor graphs, which can represent both directed and undirected models.
Prominent researchers like Geoffrey Hinton and Yee Whye Teh have contributed significantly to the development and application of undirected graphical models. Their work includes methodologies such as the Bethe free energy approximation and contrastive divergence for learning and inference in these models.