Noise-induced alternations and data-driven parameter estimation of a stochastic perceptual model

Neural systems are inherently noisy and our perceptual system can be then influenced from time to time. In this paper, we considered a perceptual model perturbed by Lévy colored noise, which is much easier to be satisfied in real-world environments than the general Gaussian noise. To elucidate the mechanism underlying the alternation behaviors induced by noise, we characterized the perceptual dynamics in terms of three statistical measures: the mean dominance duration, the number of alternations and the predominance of each interpretation. Numerical simulations showed that the stability index as well as the scale factor and the correlation time of the noise can lead to distinct changes in these measures. Then, attention was paid to data-driven parameter estimation which has typically received less attention than the exploration of stochastic behaviors. A distinctive neural network was proposed to give rise to joint estimates of system parameters and noise parameters, which can also give the measurement to describe the accuracy of estimation. The good performances of our method are shown by simulation tests.