Similarity-Based Neural Networks for Applications in Computational Molecular Biology

Abstract

This paper presents an alternative to distance-based neural networks. A distance measure is the underlying property on which many neural models rely, for example self-organizing maps or neural gas. However, a distance measure implies some requirements on the data which are not always easy to satisfy in practice. This paper shows that a weaker measure, the similarity measure, is sufficient in many cases. As an example, similarity-based networks for strings are presented. Although a metric can also be defined on strings, similarity is the established measure in string-intensive research, like computational molecular biology. Similarity-based neural networks process data based on the same criteria as other tools for analyzing DNA or amino-acid sequences.