Tunable Neuron With PWL Approximation Based on the Minimum Operator

In this brief, hardware implementation of a novel tunable neuron is proposed based on choosing the minimum operator to generate the piecewise linear approximation of the neuron's transfer function. The proposed structure is compatible with analog and mixed-signal neural networks. An accurate model of the sigmoid function is developed by using minimum operator. Moreover, if required, the activation function of the neuron can be adjusted during the training to provide various maximal slope sigmoid, hard-limit, and linear functions. The shape and the slope of the transfer function can be controlled by a 2-bit digital to analog converter. This allows the neuron to have different transfer functions based on the network requirements and during the training.