Using predefined vector systems as latent space configuration for neural network supervised training on data with arbitrarily large number of classes