Optimizing Neural Networks Based on Concept Lattices and Clusterization

The current work discusses new approaches to optimize a neural network architecture based on concept lattices. During the study the abilities of this architecture were further explored and its performance was evaluated depending on its architecture variations using different sets of 'best' concepts, chosen by their purity and object coverage. In addition, another architecture of the neural network built on neurons-clusters obtained from pre-clustered data was considered. The new approach of optimizing this architecture by changing the number of clusters-neurons was tested, as well as different clustering methods were observed. Based on the results of experiments in this work, it was found that the considered neural network architectures can be optimized by the proposed methods without loss of metric quality, and achieve results comparable to classical machine learning methods.