继续啃PRML
第八章:
Basic notation:
- node --> random variable or group of random variables
- link --> probabilistic relation ship
- notation of random var and non-random var, observed and unobserved var
Conditional Independence:
- three example: block means independent conditionally
1. tail-tail, observed(block), unobserved(unblock)
2. head-tail, observed(block), unobserved(unblock)
3. head-head, observed(unblock), unobserved(block)
- D-separation theorm: regrard graph as filter for distribution p(x)
- Markov blanket/Markov boundary
Directed graphical model --> Bayesian Networks:
- Discrete variables: three ways to control number of parameters
1. chain nodes
2. sharing parameters
3. model with latent parameter
- Continues variables: Linear-Gaussian model
Undirected graphical model --> Markov random field:
- conditional independence property in undirected graph
- factorization property in directed graph
- potential function and energy function
- how to convert directed graph to undirected and vice versa
- I map, D map, perfect map
Inference:
- chain: using potential function, local messages pass to get an efficient algorithm
- trees: undirected tree, directed tree, polytree, use efficient algorithm in a broader situation
factor graph:
- translate directed and undirected graph to factor graph to become tree
- sum-product algorithm: read it later
- max-sum algorithm: read it later
important view:
1. Basic notation
2. Conditional property and factorization property
3. Directed graphical model --> Bayesian Networks, Undirected graphical model --> Markov random field
4. Inference in graphical model