Revisiting the Ultra-Low Power Electronic Neuron Towards a Faithful Biomimetic Behavior

The spiking neural networks (SNN) have been considered as the third generation of neural networks, since it better mimics the biological behavior of the brain cortex which processes spike trains. This paper proposes to revisit electronic neuron implementations in a novel way able to deal with SNN limitations in deep learning. Post-layout simulation results prove that nonlinear action function is able to shift from right to left considering a reference sizing to create excitation and inhibition synaptic current. Proposed model is extended to handle up to three neurons in a parallel association, i.e. SNN with a fan-in/fan out of 3 synaptic branches. Revisited electronic neuron achieves an area of
$10.8\times 9.7\ \mu m^{2}$
for a energy efficiency below 10 fJ/spike, while including synapse circuitry. Relative standard deviation of synapse current is below 1.8% and synapses weight mismatch leads to less than 4 pA current error.