ABCDE
Analysis · Concept hub

Probabilistic Graphical Models

1988 CE20th-century United States (Pearl)

Concept

Variables and their dependencies as a graph. Bayesian networks (directed) and Markov random fields (undirected) — the framework of AI uncertainty.

Understand it in one breath

"Compress a high-dimensional joint distribution as a graph." With 100 variables a joint distribution has 2¹⁰⁰ entries, but conditional independencies drawn as edges shrink that to a tractable number. Medical diagnosis (symptoms → disease), robotic SLAM, and HMMs for speech recognition all live here.

At a glance

Model

Structure

Application

Naive Bayes

Class → all features

Spam filtering; text classification

Bayesian Network

Directed acyclic graph

Medical diagnosis; causal inference

HMM (Hidden Markov)

Chain: zₜ₋₁ → zₜ → xₜ

Speech, NLP, and DNA

MRF (Markov Random Field)

Undirected graph

Image segmentation

Conditional Random Field

Conditional MRF

NER; semantic segmentation

Variational Autoencoder

Deep Bayesian model

Image generation

Represent conditional independence with graph edges: a 2¹⁰⁰-state joint distribution over 100 variables can collapse to a few hundred terms.

Key formula

P(X1,,Xn)=iP(XiPa(Xi))P(X_1, \dots, X_n) = \prod_i P(X_i \mid \mathrm{Pa}(X_i))

Modern applications

Medical AI diagnosis, SLAM for autonomous systems, hidden Markov models in speech recognition, and causal inference.

Beyond MathVoyage

Loading…