A New Improved Maxnet Based on a Hybrid Neural Network That Does Not Need to be Trained

The maximum of a set is the element that is greater or equal to all the remaining ones. This seams obvious but it is this idea that is behind our Maxnet based on an hybrid neural network with multiplication units. Although this approach does not need training it implies N hard limit perceptrons, N analog switches units and one linear neuron for a set of N elements. In a first approach we consider the simpler case where we only want to get the order of the input variable(s) that is/are the maximum(s), in a second approach we consider the case where we want to get the value(s) of the maximum(s), in a third approach we solve the same problem but with only one output introducing mutual inhibitio and finally we solve the same problem without mutual inhibition and introducing a division unit to divide the sum of all maximums by the number of maximums. Finally we compare our Maxnet with the recent published proposals and we show the great advantages of our approach either for software implementation or hardware realization.