Monte Carlo Subsampling for Training Polynomial Neural Networks on Large Data Volumes

摘要

Self-organizing polynomial neural networks observe principles from the theory of evolution of complex systems in dynamically building their structure. This article gives an overview of the process of self-organization of neural networks. Next, training of a neural networks on a large data volume is considered; the latter can be a problem for ordinary self-organizing algorithms. The fact that each neuron is a simple problemsolver allows one to limit the number of data samples required for its training. The proposed method of neural network self-organization is based on selecting random training data samples when educating each neuron. The number of samples required for neuron training is estimated by using Monte Carlo methods.