Learning Inductively and Analytically
Sebastian Thrun
University of Bonn and Carnegie Mellon University
4:00 pm Thurs. March 23 in room 1325
Research on Machine Learning and AI has led to the identification of
two major learning paradigms: inductive and analytical. Inductive
techniques learn purely by observing statistical regularities in the
data. Analytical approaches generalize more rationally from less
training data, relying instead on prior knowledge about the learning
problem ("domain knowledge"). While many researchers have noted the
importance of combining inductive and analytical learning, we still
lack combined learning methods that are sufficiently effective in
practice.
In this talk, I will present the explanation-based neural network
learning algorithm (EBNN). EBNN integrates inductive neural network
learning and analytical explanation-based learning, smoothly blending
both learning principles. In a variety of application domains (mobile
robot control, robot perception, game playing) EBNN has shown to yield
superior generalization accuracies.
One of the key features of EBNN is its ability to transfer knowledge
from previously encountered learning tasks to other, new learning
tasks. This makes it particularly applicable to scenarios in which a
learner faces a whole collection of learning tasks, e.g., over its
entire lifetime. In robotics domains, which will be of particular
interest in this talk, the transfer of knowledge is crucial due to the
costs involved with operating robot hardware. I will argue that
approaches like EBNN are necessary to overcome some of the scaling
problems faced by current machine learning technology, and will
outline research strategies for the design of a lifelong-learning
robot.