Validation of a CMOS SNN network based on a time-domain threshold neuron circuit achieving 114.90 pJ/inference on MNIST

This paper proves the computing capability of a recently proposed spiking neuron circuit. The novelty of the neuron resides in being based in a subthreshold-operated ring oscillator that is indirectly powered by the input spikes. This allows a very efficient power usage. In the paper, we derive an analytical model of the neuron. The neuron is then co-simulated using the analytical model along with a transistor-level circuit to check the model accuracy. Afterward, the analytical model is used to construct an spiking neural network that can be trained to 98% of accuracy on the MNIST data set, proving equivalent performance than other contemporary circuits. As an advantage, the estimated power for a LeNet-5 implementation is 114.90pJ/inference, which is competitive with the state-of-the-art.