up in suboptimal architectures. Genetic algorithms applied to selection of architectures do
not guarantee good solutions and are computationally very demanding [7]. Missing inputs
are especially difficult to handle since filling the unknown features with the most frequently
appearing values may lead to poor results.
MLPs are widely used for classification and approximation problems, while many interesting
problems involve pattern completion and association. Associative memory models are
usually based on recurrent networks. It would be very interesting to accomplish similar task
using feedforward MLP networks. MLPs, SVMs [2] and other methods based on discriminant
analysis, perform mappings that are rather difficult to interpret. Proponents of the logical rulebased
machine learning methods consider it to be the biggest drawback of neural networks,
limiting their applications in safety-critical fields such as medicine. Similarity-Based
Methods (SBMs), for example the k-nearest neighbor (k-NN) method, retrieve the relevant
context for each query presented to the classification system, providing some interpretation
and estimating probability of different class assignment. Such interpretation is also possible
for the Radial Basis Function (RBF) networks using Gaussian or other localized functions, or
the Learning Vector Quantization (LVQ) method based on optimization of reference vectors.
It may seem that such an interpretation is not possible for MLPs since they belong to the
discriminant rather than to memory-based techniques. One way to obtain an interpretation
of MLP decisions is to study the transition from MLPs to networks performing logical
operations [8]. Although discriminant methods and prototype methods seem to be quite
different in fact the two approaches are deeply connected. A single hyperplane discriminating
vectors belonging to two classes may be replaced by two prototypes, one for each class. For
N prototypes one can generate N(N −1)/2 pair-wise discriminating hyperplanes providing
piece-wise linear approximation to the decision borders.
All these shortcomings of the MLP networks are overcome here. Recently a general
framework for Similarity-Based Methods (SBMs) used for classification has been presented
[9]. It is briefly presented in the next section, and several examples of well-known and new
neural methods derived using this framework are presented. In particular the Distance-Based
Multilayer Perceptrons (D-MLPs) are introduced, improving upon the traditional approach
by providing more flexible decision borders, using information about the structure of the data
derived from clusterization procedures and enabling a prototype-based interpretation of the
results. Symbolic values used with probabilistic distance functions allow to avoid ad hoc
procedure to replace them with numerical values. SBM perspective allows to initialize all
D-MLP network parameters starting from some one of standard clusterization procedures
and thus using information that may be easily obtained from the data. A simple procedure
to change D-MLP models into associative memories and to use them in pattern completion
problems is described in the fourth section. As a result missing values are handled in an
efficient way. Finally to avoid writing computer programs for the backpropagation method
for each type of distance function a simple transformation of the input data is proposed,
allowing for distance-based interpretation. An illustration of this method on the Iris data is
presented in the sixth section. The paper is finished with a short discussion.