18-bayesianin想ference.pptxVIP

Homework 2 Mean: 120.5 Median: 129 St.Dev: 31 Max: 169 Min: 37 Bayesian Inference Explaining Away Independence Is there a principled way to determine all these dependencies? Yes! It’s called D-Separation – 3 rules. (not in the book) d-separation Clarification: A “blocked” path makes X and Z independent. A path is “blocked” if… A path is “blocked” if… Bayesian Inference Compute posterior probability distribution for a set of query variables given some observed event Event: an assignment of values to evidence variables Bayesian Inference Exact Inference Compute the exact value of the posterior distribution Inference by Enumeration This is what we did with the Conditional Probability Tables... Exact Inference Inference by enumeration is computationally intractable (NP-hard) in the general case Optimizations: Variable elimination Dynamic programming Variable clustering … All of these help, but the problem is still NP-hard. Approximate Inference Since exact inference is so hard, approximate inference methods are often used in real world applications. Most common approaches use randomized sampling Approximate Inference: Sampling We will talk about Direct sampling Markov chain sampling All these techniques are called Monte Carlo methods: Approximate algorithms whose accuracy depends on the number of samples generated. Sampling the Joint Distribution Example of sampling from the joint distribution Example of sampling from the joint distribution Example of sampling from the joint distribution Example of sampling from the joint distribution Example of sampling from the joint distribution Example of sampling from the joint distribution Example of sampling from the joint distribution [true, false, true, true] Inference via sampling Total number of samples. Inference via sampling How do we compute P(X = x | e)? Rejection Sampling Inference via Sampling Inference via Sampling Markov chain sampling Markov Chain Monte Carlo (MCMC) Instead of generating each sample from sc