On improvements of competitive networks principles

Abstract

Artificial neural networks are very powerful computational systems and are capable of solving many various problems. They provide a nonlinear input to output mapper, and therefore, applications in classification, feature extraction or pattern recognition may be expected. This dissertation thesis focuses on competitive networks which belong to the unsupervised learning systems family. The aim of this thesis is to present possible improvements to this area and provide necessary experiments which support suggested ideas. First, an introduction to artificial neural networks is given and several common models are stated including basic feed-forward network, Self-organized Feature Map, Growing Neural Gas etc.. Next, several innovative approaches and methods are suggested. In particular, the main contributions of the thesis are as follows:

1. A classifier based on the competitive network is presented. An innovative method for learning is suggested where the competition is not supervised in the standard way but the input signals to the network are altered in order to achieve the supervision.

2. An application for the presented supervised competitive network is presented. The RBF network initialization process is addressed and possible improvements brought by the supervised competition are examined.

3. An application of competitive network in combination with the force-based graph plotting algorithm is presented. The high-dimensional data visualization technique is addressed, advantages of such approach are explored, and several datasets are visualized.

4. A new insight into competitive network principles is given. An innovative physically based model for the competition is stated and various attributes of this model are examined. The model is built above the particle system which is driven by mechanical laws of motion and all competitive principles are transformed into forces affecting particles. The model is tested against several well-known datasets to prove functionality and usability.