Abstract
Binocular rivalry occurs when two eyes are presented with incompatible images; observers experience temporal alternations of the two eyes’ images. Such descriptions often mislead readers to assume regular rhythmic rivalry dynamics.
However, rivalry dynamics are noisy: percept durations show high variability. (1) The percept durations follow a gamma distribution. (2) Autocorrelations in the percept durations, although widely known, receive little attention in modelling. (3) Similarly, there is a substantial chance of and individual differences on returning to the same exclusive percept after a mixture percept (van Ee, 2009). These observations are unlikely to be explained by injecting Gaussian noise in the adaptation term of oscillator models.
Instead, we treat the rivalry process as a Bayesian inference process that is constantly inferring the true state of the world. As full inference is too difficult, we use a sampling approach to approximate the optimal solution at each time point, treating perceptual states as samples from an internal distribution of hypotheses (e.g., Gershman et al. 2009). Here, samples are autocorrelated, meaning new samples are related to preceding samples, and have momentum, meaning new samples continue in the same direction as previous samples.
This Bayesian sampling model qualitatively captures the shape of the percept distributions, autocorrelations in percepts, and the return rate, matching the dynamics which previous models have not captured. Using this model, binocular rivalry can be viewed as a decision-making process performed by the early visual cortex, helping to draw connections between early visual systems and higher-level decision-making.
Keywords
Binocular Rivalry, Bayesian Sampler Model, Noise
Start Date
15-5-2024 9:30 AM
End Date
15-5-2024 10:30 AM
Recommended Citation
Qian, Stella; Spicer, Jake; and Sanborn, Adam N., "Accounting for the structure of noise in binocular rivalry with a Bayesian Sampling Approach" (2024). MODVIS Workshop. 4.
https://docs.lib.purdue.edu/modvis/2024/pav/4
Accounting for the structure of noise in binocular rivalry with a Bayesian Sampling Approach
Binocular rivalry occurs when two eyes are presented with incompatible images; observers experience temporal alternations of the two eyes’ images. Such descriptions often mislead readers to assume regular rhythmic rivalry dynamics.
However, rivalry dynamics are noisy: percept durations show high variability. (1) The percept durations follow a gamma distribution. (2) Autocorrelations in the percept durations, although widely known, receive little attention in modelling. (3) Similarly, there is a substantial chance of and individual differences on returning to the same exclusive percept after a mixture percept (van Ee, 2009). These observations are unlikely to be explained by injecting Gaussian noise in the adaptation term of oscillator models.
Instead, we treat the rivalry process as a Bayesian inference process that is constantly inferring the true state of the world. As full inference is too difficult, we use a sampling approach to approximate the optimal solution at each time point, treating perceptual states as samples from an internal distribution of hypotheses (e.g., Gershman et al. 2009). Here, samples are autocorrelated, meaning new samples are related to preceding samples, and have momentum, meaning new samples continue in the same direction as previous samples.
This Bayesian sampling model qualitatively captures the shape of the percept distributions, autocorrelations in percepts, and the return rate, matching the dynamics which previous models have not captured. Using this model, binocular rivalry can be viewed as a decision-making process performed by the early visual cortex, helping to draw connections between early visual systems and higher-level decision-making.