A CORDIC implementation of a digital artificial neuron

Digital implementations of neural networks are either very complex or very simple, often complicated by the difficulty in building the sigmoidal activation function; in fact, many implementations use hard limiters or saturated linear activation functions to avoid the issue. This paper presents the CORDIC implementation of a digital neuron, achieving a data rate of 0.988 million synaptic connections/second, suitable for multilayer perceptrons. The CORDIC hardware algorithm is well known for its ability to compute difficult transcendental functions. Furthermore, the same CORDIC hardware can be used to calculate the net value to reduce hardware complexity.