Summary
This chapter was about non-parametric Bayesian methods. Non-parametric methods are extremely useful because we don’t have to make parametric assumptions about the form of the relationship between our features and the target variable. That has required us to learn a new approach to probabilistic modeling and new concepts. Those new concepts include the following:
- Non-parametric Bayesian methods focus on specifying a modeling function as coming from a prior distribution over functions
- The priors in non-parametric Bayesian methods are stochastic processes such as GPs or DPs
- A GP is specified by a kernel (covariance) function and a mean function
- In GPR, our model is the mean function of the response (target) variable and we put a GP prior on that mean function
- In GPR, we can fit the kernel function parameters to the training data by maximizing the marginal likelihood
- A DP is defined in terms of its base distribution,
, and its concentration parameter...
