Non-Linear Data for Neural Networks Training and Testing

摘要

Highly nonlinear data sets are important in the field of artificial neural networks. It is not feasible to design a neural network and try to classify some real world data directly with that network. N-bit parity is one of the oldest data used to train and test neural networks. The simplest is the 2-bit parity also known as the XOR classification problem. Some researchers say that N-bit parity s set though highly nonlinear it is a simple task to learn by neural networks, others were drifted to tailor special purpose neural networks to solve only the N-bit parity problem without explaining why there is such a need. Is it possible to judge the N-bit parity is a simple data due to the fact that it can be modeled by a deterministic finite accepter? Moreover, should patterns that are in the form of context free which require a pushdown automaton, or context-sensitive and recursively enumerable that require a Turing machine be harder to learn by neural networks? The aim of this paper is to focus on and propose some complex nonlinear data to be used in training and testing of neural networks. The most important in these parity data is that the developer can tune the complexity of nonlinearity through various amounts of degrees; the user can select various numbers of categories, huge number of pattern samples, and many hybrid symbols. Testing for various neural networks and their generalization and ability to classify unseen patterns can be more effective. Experimental results on the classification of prime numbers showed that neural networks can learn the classification of prime numbers.