| Binocular rivalry is a paradigmatic phenomenon of multi-stable perception in which, given two visual stimuli presented to the two eyes, perception alternates between the two percepts, with dominance intervals irregularly distributed in time. The associated rich and intriguing phenomenology has prompted the development of several models proposing dynamic neural mechanisms underlying the temporal properties of the perceptual alternation. Most models are based on two main ingredients: a fast competition mechanisms between the neural representations of the two percepts (through various forms of mutual inhibition) for the exclusive alternative percepts dominance; slow adaptation-like effects which allow a percept to gain a chance to come to conscious perception after a dominance period of the other; noise, to account for the wide observed distribution of dominance times. Such models are essentially memoryless, and have difficulties in explaining the observed modulation of dominance times statistics on long time scales, in particular in experimental protocols in which the presentation of stimuli is repeatedly interrupted blanks. The model we propose is able to account for such long memory effects, while retaining agreement with experimental evidence compatible with previously proposed models, and generates some testable predictions. The proposed architecture includes two families of neural populations, defined as 'evidences populations' (EP), which integrate on short time scales visual information on the stimuli, and 'memory populations' (MP), which slowly integrate information of the active percept. Each population in EP or MP is assumed to randomly switch, forward and backward, from a 'spontaneous' to an 'active' state with assigned probabilities per unit time, which depend on the stimulation and perceptual conditions; neural populations are not explicitly modeled as sets of spiking neurons, but effectively described through their switching dynamics. Accordingly, the system dynamics is governed by a pair of Master Equations for the time evolution of the probabilities of having a given number of active EP and MP. A readout stage defines the active percept according to a threshold mechanism: when the number of EP + MP crosses a threshold the corresponding percept becomes active; upon switching percept, EP of the loser percept are temporarily inhibited. The model reproduces the well known and robust feature of binocular rivalry, by which the histogram of dominance times can usually be well fitted by a Gamma distribution. Dominance intervals of the two percepts have also been reported to show non-negligible serial correlation for uninterrupted presentations; the model is consistent with such evidence, and predicts a modulation of the correlation profile for interrupted presentations. Interrupted presentations have been observed to stabilize the percepts; this somewhat counterintuitive effect is correctly accounted for by the model in terms of the 'survival probability', i.e. the probability of having the same percept before and after the interruption; the model also predicts a specific dependence of the survival probability on the persistence time of the stimuli between two interruptions. The model is consistent with another piece of evidence, concerning the increase of the survival probability with increasing dominance time of the percept before the interruption, and predicts a highly nonlinear dependence of the average dominance intervals on the persistence time of the stimuli. |
|