Stochastic resonance in noisy spiking retinal and sensory neuron models

Two new theorems show that small amounts of additive white noise can improve the bit count or mutual information of several popular models of spiking retinal neurons and spiking sensory neurons. The first theorem gives necessary and sufficient conditions for this noise benefit or stochastic resonance (SR) effect for subthreshold signals in a standard family of Poisson spiking models of retinal neurons. The result holds for all types of finite-variance noise and for all types of infinite-variance stable noise: SR occurs if and only if a sum of noise means or location parameters falls outside a 'forbidden interval' of values. The second theorem gives a similar forbidden-interval sufficient condition for the SR effect for several types of spiking sensory neurons that include the Fitzhugh-Nagumo neuron, the leaky integrate-and-fire neuron, and the reduced Type I neuron model if the additive noise is Gaussian white noise. Simulations show that neither the forbidden-interval condition nor Gaussianity is necessary for the SR effect.