Singular Covariance Collapse
If a component's mean lands on a single data point, its covariance can shrink toward zero and the likelihood diverges to infinity — a pathological spike, not a good fit; regularization adds a small floor to Sigma's diagonal (or a Wishart prior) to keep it positive definite.
This Concept is waiting for its first lesson!
If a component's mean lands on a single data point, its covariance can shrink toward zero and the likelihood diverges to infinity — a pathological spike, not a good fit; regularization adds a small floor to Sigma's diagonal (or a Wishart prior) to keep it positive definite.
Are you a teacher? Sign in to start contributing.
Sign In