Neural network construction using multi-threshold quadratic sigmoidal neurons

A new type of neuron called the multi-threshold quadratic sigmoidal neuron is proposed. In cooperation with single-threshold quadratic sigmoidal neurons, the multi-threshold quadratic sigmoidal neuron can be used to construct multilayer neural networks in order to dichotomize arbitrary dichotomy defined on any given training set. For such constructed neural networks, it is proved that the number of required hidden neurons is only one-fourth of those networks with the standard architecture that is often assumed in theoretical studies