"Incremental Learning"

5/22/1998: [time not recorded]
[location not recorded]

Abstract: The ability to learn internal models about the environment is often assumed to be one of the key ingredients for intelligent information processing, be it in robotics, process control, or biological organisms. In autonomous learning systems that have a high input rate of sensory data, internal models need to be acquired by incremental learning, i.e., data points are only used once for updating the model and are then discarded. This talk will discuss the problems that need to be faced in incremental learning, in particular the problems of catastrophic interference and how to allocate sufficient resources for the learning network. By using techniques from nonparametric statistics, it will be shown how principled and fast solutions to incremental learning can be found in the framework of locally weighted regression. We will also discuss how the presented methods scale to high dimensional learning problems, a domain that has usually been assumed to be not suitable for local learning systems. The usefulness of our algorithms and the validity of their assumptions are illustrated with charts and videos for synthetic data and in applications using an actual 7 degree-of-freedom anthropomorphic robot arm that learns various dynamic manipulations tasks.

Last updated: Mon Jun 19 17:44:06 2006