APPROXIMATE-LINEAR NEURON AND NON-LINEAR NEURON

摘要

The neuron model normally widely applied in artificial neural network (ANN) has the property of monotone and bounded I/O. This paper names it as approximate-linear neuron (ALN). It simplifies the complexity of the neural network: at the same time, it becomes a bottleneck that limits the network's abilities. In contrast, by comparing the stability of Hopfield neural network with the complex behavior of Cconway's life game, this paper prefers a new concept, non-linear neuron (NLN), of which I/O property is non-linear completely. NLN can further enhance the non-linear properties of ANN, so an ANN with NLN will be more powerful. Remainder ANN is an example.