Ring-VCO-based ReLU activation function with linearity improvement for pulsed neuron circuits

Motivated by the linearity of the ring VCO's V-to-F characteristics in this paper we present a low power ReLU activation function for pulsed neural network. The proposed rectifier is based-on our linearity improved ring-VCO. Furthermore, our studies shows that nonlinearity is affected by process variation and we discuss a compensation technique. These improvements are illustrated using simulation results of the proposed neuron circuit which is designed in
$0.18-\mu\mathrm{m}$
CMOS technology. The designed pulsed neuron circuit which can be used in DNN is benchmarked by MNIST performance.