17. Conditional Probability Models

In this lecture we consider prediction functions that produce distributions from a parametric family of distributions. We restrict to the case of linear models, though later in the course we will show how to make nonlinear versions using gradient boosting and neural networks. We develop the technique through four examples: Bernoulli regression (logistic regression being a special case), Poisson regression, Gaussian regression, and multinomial logistic regression (our first multiclass method). We conclude by connecting this maximum likelihood framework back to our empirical risk minimization framework. Access the full course at https://bloom.bg/2ui2T4q