The Role of Transfer in Learning (extended abstract)
Sebastian Thrun · 1996
Introduction Virtually all of today's approaches to artificial neural network learning generalize considerably well if sufficiently many training examples are available. However, they often work poorly when training data is scarce. Various psychological studies have illustrated that humans are able to generalize accurately even when training data is extremely scarce. Often, we generalize correctly from just a single training instance. In order to do so, we appear to massively re-use knowledge acquired in our previous lifetime. Lifelong learning is a framework that addresses the issue of knowledge re-use and inductive transfer in learning. In lifelong learning, it is assumed that the learner faces an entire family of learning tasks, not just a single one. When facing a new learning task, the learner may transfer knowledge acquired in previous learning tasks to boost generalization. Three questions are of fundamental importance for any approach to lifelong learni