机器学习与概率图模型中科院自动化所系列报告(王立威).pptVIP

Mean field method: Using first order approximation of the target distribution: The fixed point equation and updating rule have simple forms. * Sampling Based Methods * Sampling-based (Monte Carlo) algorithms: Idea: Using sample based frequency to approximate the probability. Probability is Expectation: Expectation (of a function) can be approximated by sample average: * Monte Carlo is useful in many problems: High dimensional numerical integration. How to generate the sample when the target probability distribution is difficult to compute? Markov Chain Monte Carlo (MCMC) Key idea :Generat