@incollection{cybenko:effectiveness,
    author = {George Cybenko and S. Saarinen and Robert Gray and Yunxin Wu and
Alexy Khrabrov},
    title = {On the effectiveness of memory-based methods in machine learning},
    booktitle = {Dealing with Complexity},
    editor = {K. Warwick},
    year = {1997},
    publisher = {Springer-Verlag},
    copyright = {Springer-Verlag},
    group = {agents},
    url = {http://agent.cs.dartmouth.edu/papers/cybenko:effectiveness.ps.gz},
    urlpdf = {http://agent.cs.dartmouth.edu/papers/cybenko:effectiveness.pdf},
    keyword = {machine learning},
    abstract = {Many memory-based methods for learning use some form of nearest
neighbor inference. By memory-based, we mean methods that localize data in
the training sample to make inferences about novel feature values. The
conventional wisdom about nearest neighbor methods is that they are subject
to various curses of dimensionality and so become infeasible in high
dimensional feature spaces. However, recent results such as those by Barron
and Jones suggest that these dimensionality problems can be overcome in the
case of parametric models such as sigmoidal neural networks which are
patently nonlocal. This creates a paradox because memory-based methods have
been shown to perform well in a number of applications. They are often
competative with parametric methods in terms of prediction error and actually
superior in terms of training time. In this paper, we study the unreasonable
effectiveness of memory-based methods. We analyze their performance in terms
of new metrics that take into consideration ...}
}