Marginal Likelihood / Evidence
Definition
The marginal likelihood, also called the evidence, is
$$p(x) = \int p(x, z)\,dz.$$It normalizes the posterior and is often intractable in latent variable models.
Why it matters
It is the reason direct posterior computation is difficult, and why VI introduces a surrogate objective.
Related
- Posterior Distribution
- ELBO
- STA414 W6+A3