A Quadratic Neural Net for Classification
H. Qarten · 2005
Neural Nets, used for cassification purposes, are implemented by analog, VLSI techniques. If the dimension of the feature apace is high, the resulting chip size is large and the power consumption is increased as well. This work discusses the use of a Quadratic Neural Net (QNN) for classsification. In this scheme, the decision boundaries are approximated by quadratic function and not by lines or hyperplanes, as is the case in the regular Neural Net. It is shown that the QNN scheme offers a reduced number of connections and the total number of active elements is reduced as well. As a result, there is an improvement of 50%, in the chip size and at least 50% in the power consumption. Due to the reduction in the number of weights the convergence rate during the learning phase is improved as well.