# Markov Networks ### Key ideas * Our goals is almost always to learn a joint distribution, $P(X)$. Once we have a joint distribution, inference tends to be straightforward. * In order to learn the joint distribution, we introduce *factors*. Factors allows us to capture the *affinity* between connected random variables. * In order to learn the factors, we must introduce *parameters*. See more [here](https://towardsdatascience.com/markov-networks-undirected-graphical-models-dfb19effd8cb) and [here](https://www.youtube.com/watch?v=m-W0gLpOT94&list=PL3pGy4HtqwD2kwldm81pszxZDJANK3uGV&index=132). * We learn these parameters in a way such that some *objective function* is satisfied, see [here](https://youtu.be/H3yWjhthGw0?list=PL3pGy4HtqwD2kwldm81pszxZDJANK3uGV&t=598) * The selection of our parameters are a *modeling choice*. ### Restricted Boltzman Machine * Goal: learn an abstract representation of our observed random variable space * Gibbs distribution: see [here](https://youtu.be/m-W0gLpOT94?list=PL3pGy4HtqwD2kwldm81pszxZDJANK3uGV&t=1265) ### Random notes * When dealing with latent variables, we never know what they actually represent (see [here](https://youtu.be/lXrFX3vjtjQ?list=PL3pGy4HtqwD2kwldm81pszxZDJANK3uGV&t=1296)) --- Date: 20210622 Links to: [003-Data-Science-MOC](003-Data-Science-MOC.md) References: * [Markov Network, NPTEL course](https://www.youtube.com/watch?v=xlqn-hfrqeU&list=PL3pGy4HtqwD2kwldm81pszxZDJANK3uGV&index=131&t=10s) * [Graphical Models: A Combinatorial and Geometric Perspective - Lecture 1](http://www.fields.utoronto.ca/talks/Graphical-Models-Combinatorial-and-Geometric-Perspective-1) * [Graphical Models: A Combinatorial and Geometric Perspective - Lecture 2](http://www.fields.utoronto.ca/talks/Graphical-Models-Combinatorial-and-Geometric-Perspective) * [Graphical Models: A Combinatorial and Geometric Perspective - Lecture 3](http://www.fields.utoronto.ca/talks/Graphical-Models-Combinatorial-and-Geometric-Perspective-0)