ONLINE LEARNING OF GAUSSIAN MIXTURE MODELS - A Two-Level Approach
Arnaud Declercq, Justus Piater · 2008
We present a method for incrementally learning mixture models that avoids the necessity to keep all data points around. It contains a single user-settable parameter that controls via a novel statistical criterion the trade-off between the number of mixture components and the accuracy of representing the data. A key idea is that each component of the (non-overfitting) mixture is in turn represented by an underlying mixture that represents the data very precisely (without regards to overfitting); this allows the model to be refined without sacrificing accuracy.