E-step
Compute posterior q(z) = P(z|x, θ) over hidden variables given current parameters.
Advertisement
M-step
Update θ maximizing E_q[log P(x, z | θ)]. Often closed form.
Advertisement
Convergence
Monotonic increase in likelihood. Converges to local maximum. Random restart or careful init needed.
Applications
Gaussian Mixture Model. Hidden Markov Model. Latent Dirichlet Allocation. Missing data imputation.