Stochastic Resonance in Retinomorphic Neural Networks with Nonidentical Photoreceptors and Noisy McCulloch-Pitts Neurons

Abstract—We propose a simple retinomorphic neural network that consists of photoreceptors generating nonuniform outputs for common optical inputs with random offsets, an ensemble of noisy McCulloch-Pitts neurons each of which has random threshold values, local synaptic connections between the photoreceptors and the neurons with variable receptive fields (RFs), output cells, and local synaptic connections between the neurons and output cells. Through numerical simulations, we observed stochastic resonance among the proposed pixels. We calculated correlation values between the optical inputs and the outputs as a function of the RF size and intensities of the random components in photoreceptors and the McCulloch-Pitts neurons, and then found nonzero optimal RF sizes as well as optimal noise intensities of the neurons under the nonidentical photoreceptors. This implies that SR-based night-scope image sensors with an array of nonidentical photosensors would be developed with less efforts to implement uniform pixel devices.