12.2.3 Bayesian PCA - Pattern Recognition and Machine Learning

Описание: An important problem that arises when fitting data with PCA is choosing the dimension of the principal subspace. The Bayesian approach to this model selection problem would be to compare marginal likelihoods of models with different dimensionality. In this video, we discuss an alternative approach: automatic relevant determination. The idea is to place a Gaussian prior on the columns of the matrix determining the principal subspace with zero-mean and learnable precisions. The precision of superfluous columns will tend to infinity, forcing the corresponding columns to zero, removing them from the model. We discuss how to modify the EM updates for probabilistic PCA to accommodate the precisions and two approaches to updating the precisions.

My notes on this section: https://sinatootoonian.com/index.php/...