Slow subspace learning from stationary processes

The talk presents a method of unsupervised learning from stationary, vector-valued processes. The method selects a subspace on the basis of an objective which can be used to bound the expected classification error for a family of tasks posessing a temporal continuity property. We prove bounds on the objective’s estimation error in terms of mixing coefficients and consistency for absolutely regular processes. Experiments with image processing demonstrate the algorithms ability to learn geometrically invariant feature maps.