A Division Based Neuron for Neural Networks

We propose an alternative neuron design for artificial neural networks that replaces wx+b in a typical neuron with w1x/w2x. The design allows more complex calculations, such as division, to be performed efficiently within a neural network, and thus more efficient and specialized prediction models to be made. Along with the new design, we developed an algorithm to dynamically adjust the learning rate for model training. The algorithms were tested in the training of single-layer and single-neuron models to predict the outcome of various division-based operations. The initial test results showed that the average prediction errors are between 0.0241% and 0.898%. That is significantly more accurate than traditional neural networks with more layers.