This paper presents a neural network approach to modelling similarity computations. Generally, in the many applications of similarity, geometric and contrast models have been the primary tools used to perform similarity computations. We discuss both of these models and show that they have specific theoretical and empherical problems. We, then, use neural network approaches (e.g., simple recurrent networks) in two implementations which overcome these problems. We demonstrate these findings by presenting a set of experiments which illustrate the failings of the general models and which highlight the utility of these neural network techniques.