Adaptive on-line learning of probability distributions from fieldtheories

摘要

An adaptive algorithm is considered in on-line learning ofprobability functions, which infers a distribution underlying observeddata x₁, x₂, …, x_N. Thealgorithm is based on how we can detect the change of a source functionin an unsupervised learning scheme. This is an extension of an optimalon-line learning algorithm of probability distributions, which isderived from the field theoretical point of view. Since we learn notparameters of a model but probability functions themselves, thealgorithm has the advantage that it requires no a priori knowledge of amodel