In this talk, we present a novel approach to online change detection problems,
based on estimating the expected information content of a new data point.
In the case of the Gaussian as the distribution generating the data, an
expression can be derived which does not depend on the statistics estimated
from the training data, but only on the size of the training set.
This result leads naturally to a novelty detection method that, in the Gaussian
case, is closely related to statistical testing.
Such test allows to control the false positive rate even when a small training
set is available.
Although this represents a remarkable result, its applications is limited to a
very small class of real problems, given the strong assumption on the
We discuss the connections between the statistical and information theoretic
approaches, as well as the extension of the proposed method to other
We will discuss in more detail the extension to the case of the mixture of