Nonparametric Bayesian Inference Of Contrast Sensitivity Function

The contrast sensitivity function(CSF)is crucial in predicting functional vision both in research and clinical area.It has been established that the CSF could be interpreted in terms of a two-dimensional(2-D)psychometric function,and be estimated by 2-D adaptive Bayesian methods more efficiently compared with traditional methods.In this study,we first investigated the use of 2-D Bayesian inference to extract the information contained in experimental data that were sampled by different strategies,such as staircase,?-methods,or 2-D Bayesian adaptive methods,to estimate the parameters of CSFs.We extensively simulated its performance as well as validated the results in a psychophysical experiment.Our results first show that 2-D Bayesian inference is much more efficient and obtains the same precision in one fourth of the trial numbers of the standard convergence points at each spatial frequency.In addition,we explore the efficiency of different sampling methods and show that when applying Bayesian inference on the various methods,the sampling efficiencies of traditional staircase methods,?-methods and modern 2-D adaptive methods are comparable.This similarity suggests that there are small differences between the sampling efficiency of traditional adaptive methods and modern 2-D adaptive methods,and the superiority of modern 2-D adaptive methods upon traditional ones is mainly due to their 2-D Bayesian inference.The psychometric function(PF),describing the subject's response to the strength of a stimulus,is fundamental to psychophysics.Traditionally,this function is estimated by fitting a parametric model,such as Weibull or logistic function,to the experimental data.This approach works well if the model is correct,but it will mislead if not.This thesis describes a model free Bayesian fitting approach(Gaussian Process Classification-GPC)for the estimation of PF from psychometric measuring procedures.This approach does not make any assumption about the underlying mathematical model,but instead simply assumes smoothness and continuity of the PF.We investigated the statistical properties of the estimates on slope and threshold of the psychometric function obtained from this method.They were derived from Monte-Carlo simulations and compared with those of model-based maximum likelihood(ML)estimates and another model-free strategy,the local linear fitting(LLF).We found that to estimate PF parameters from binomial response data,the GPC approach frequently performed better than and merely worse than ML estimates or LLF estimates.At last,we investigated the use of 2-D Gaussian processes classification to extract the information contained in experimental data for contrast sensitivity measuring.With the help of Monte-Carlo simulation,we extensively invetigated the statistical properties of the estimates on peak,area under log contrast sensitivity function(AULCSF),as well as the local deficits such as notches of CSF.The ML method performed a bit better than GPC approach if the CSF was normal.However,if the CSF had some local deficits,the ML method failed to estimate these local deficits,while GPC had both better global and local estimates of the CSF.We suggest that the GPC strategy can be applied to directly estimate contrast sensitivity function for its reliability and suitability.