Stochasticity In Neuromorphic Computing: Evaluating Randomness For Improved Performance

Several researchers have proposed that random noise in highly non-linear biological systems helps with learning and information processing. Neuromorphic systems could potentially harness the computational power of such biological systems by utilizing stochasticity to imitate random biological noise. Systems implementing this kind of probabilistic behavior would open up an entirely new realm of potential algorithms and applications to neuromorphic developers. This paper proposes an efficient and reliable stochastic spiking neuromorphic topology that utilizes variations of the membrane capacitance to realize stochastic behavior. We also present analyses of network-level effects of this stochastic spiking neuron behavior, showing the performance impact on classification and other neuromorphic applications. This network-level analysis shows that stochastic noise does in fact provide powerful generalization properties that improve performance in emerging neuromorphic systems.