http://info.usherbrooke.ca/hlarochelle/neural_networks/content.html

these characteristics may come from a word. (hand writting data) 

sequence of observation => model the joint distribution over the whole sequence

linear chain CRF

usually => iid assumption

but for the adjacent positions in a sequence => linear chain CRF

first term: from x_k

seconde term: from V matrix

context window

 three neural network, weighted by a(0) a(-1) a(+1)

 alternative: only one NN

computing the partition function

 y‘ ≠ y

y_k is the resultant sequence

y‘_k is all the probable sequence

the goal here is to calculate Z(X) in polynomial time (dynamic programming)

if someone gives me y2‘ then we can calculate \alpha_1(y2‘)

https://www.spaces.ac.cn/archives/5542/comment-page-1 

advantage function????

a = max x_n

V(s) = max_a Q(a|s)

A(a|s) = Q(a|s) - V(s)

computing marginals

performing classification

factors, sufficient statistics and linear CRF

Markov network

factor graph

another visualization to get rid of the ambiguity.

belief propagation

[Neural Network] {Université de Sherbrooke} C3: Conditional Random Field

原文：https://www.cnblogs.com/ecoflex/p/10884617.html