In Blei's original paper about LDA(Latent Dirichlet Allocation), he used VBEM method to train the model. In E step, update the posterior distribution, in M step, update the hyperparameter. But in other tutorials about VBEM, I didn't see them use such method. VBEM in other tutorials just compute $q(z)$ in E step and compute $q(\theta)$ in M step. I'm confused what is a standrad variational bayes EM.
- Is standard variational bayes EM need to update hyperparameters?
- Or we can say Blei's method is a combination of variational bayes EM and empirical bayes?