Neural Computation, Vol 10, 1815-1830, Copyright © 1998 by The MIT Press
Neural Feature Abstraction from Judgements of Similarity
Michael D. Lee
The common neural network modeling practice of representing the elements of
a task domain in terms of a set of features lacks justification if the
features are derived through some form of ad hoc preabstraction. By
examining a featural similarity model related to established
multidimensional scaling techniques, a neural network is developed that
generates features from similarity data and attaches weights to these
features. The network performs a constrained search of a continuous
solution space to determine the features and uses a previously developed
regularization technique to minimize the number of features it derives. The
network is demonstrated on artificial data, from which it recovers known
features and weights, and on two real data sets involving the similarity of
a set of geometric shapes and the abstract conceptual similarities of the
10 Arabic numerals. On the basis of these results, the relationship between
the multidimensional scaling approach adopted by the network and an
alternative additive clustering approach to feature extraction is
discussed.
