This paper includes a presentation of experiments performed on Contextual Neural Networks with a dynamic field of view. It is checked how their properties can be affected by the usage of not-uniform numbers of groups in different layers of contextual neurons. Basic classification properties and activity of connections are reported based on simulations with H2O machine learning server and Generalized Backpropagation algorithm. Results are obtained for data sets with a high number of attributes (gene expression of bone marrow cancer and myeloid leukemia) as well as for standard problems from UCI Machine Learning Repository. Results indicate that layer-wise selection of numbers of connection groups can have a positive influence on the behavior of Contextual Neural Networks.