ICML-98 Submission #145
A Supra-Classifier Architecture for Scalable Knowledge Reuse
Kurt Dewitt Bollacker and Joydeep Ghosh
Department of Electrical and Computer Engineering,
University of Texas at Austin, 78712
USA
Abstract
When faced with inadequate information, humans often use knowledge
gained from previous experience to help them in making decisions.
Even when this knowledge is spread thinly among many previous
experiences, humans are able to effectively accumulate and apply it to
a current classification task of interest. Inspired by human
knowledge reuse, we have previously introduced a general framework for
the use of knowledge embodied in existing classifiers to aid in a new
classification task. In this framework, a supra-classifier is built to
make decisions based on the outputs of large numbers of previously
trained classifiers designed for different, but possibly relevant
tasks. In this article, we discuss the Hamming Nearest Neighbor (HNN)
supra-classifier architecture and mathematically show its usefulness.
Experiments on public domain data sets demonstrate the practicality of
the framework and HNN supra-classifier when faced with very few
training examples.
Keywords:
Knowledge Transfer, Nearest Neighbor, Curse of Dimensionality
Email address of contact author: kdb@pine.ece.utexas.edu
Phone number of contact author: 512-471-2358