190 likes | 212 Views
This session covers essential distributions and concepts in probability theory, their application in machine learning, and key techniques such as biased and unbiased estimators. Learn about the Exponential Family, Conjugate Priors, and how to calculate Expectation and Variance in various distributions.
E N D
Distributions and Concepts in Probability Theory 10701 Recitation PengtaoXie
Outline • Important Distributions • Exponential Family • Conjugate Prior • Biased and Unbiased Estimators
Outline • Important Distributions • Exponential Family • Conjugate Prior • Biased and Unbiased Estimators
Distributions • Usage in machine learning • Expectation and variance • Bernoulli, Beta, multinomial, Dirichlet, Gaussian
Usage of distributions in ML • Gaussian: Least Square Regression, Mixture of Gaussians, Kalman Filtering, Gaussian Markov Random Field, Gaussian Process • Multinomial: Hidden Markov Model, Mixture of Gaussians, Latent Dirichlet Allocation, Naive Bayes classifier • Dirichlet: Latent Dirichlet Allocation, Dirichlet Process • Bernoulli: Logistic Regression, switching variables in Graphical Models • Beta: Beta Process
Expectation and variance • Expectation: the average value of a random variable under its probability distribution • Variance: a measure of how much variability there is in x around its mean value
Distributions Random Variable Discrete Random Variable Continuous Random Variable Two Outcomes Multiple Outcomes Bernoulli Distribution Multinomial Distribution Gaussian Distribution Conjugate Conjugate Beta Distribution Dirichlet Distribution Conjugate
Bernoulli • Model binary variable {0,1} • Probability mass function • Expectation
Multinomial distribution • Model variables taking K possible states • 1-of-K coding • Probability mass function • Expectation
Beta • Prior of the parameter in Bernoulli distribution • Probability density function • and are pesudo counts Refer to note1.pdf
Dirichlet distribution • Prior of parameters in multinomial distribution • Probability density function • are pesudo counts
Univariate Gaussian distribution • Model continuous variables • Probability density function • Expectation and variance
Multivariate Gaussian distribution • Defined on a continuous random vector • Probability density function • Expectation and covariance
Outline • Important Distributions • Exponential Family • Conjugate Prior • Biased and Unbiased Estimators
Exponential Family Distribution • A class of distributions sharing a certain form • Natural parameters and sufficient statistics • Special cases: Bernoulli, Beta, multinomial, Dirichlet, Gaussian • Moment generating property Refer to note2.pdf
Outline • Important Distributions • Exponential Family • Conjugate Prior • Biased and Unbiased Estimators
Conjugate Prior both Beta both Dirichlet both Gaussian From the same distribution Refer to note3.pdf
Outline • Important Distributions • Exponential Family • Conjugate Prior • Biased and Unbiased Estimators
Estimator Bias • Bias of an estimator • Unbiased estimator and biased estimator • Example: MLE for Gaussian mean and variance Refer to note4.pdf