A statistical perspective to visual masking

Abstract

A stimulus (mask) reduces the visibility of another stimulus (target) when they are presented in close spatio-temporal vicinity of each other, a phenomenon called visual masking. Visual masking has been extensively studied to understand dynamics of information processing in the visual system. In this study, we adopted a statistical point of view, rather than a mechanistic one, to investigate how mask-related activities might influence target-related ones within the context of visual masking. We modeled the distribution of response errors of human observers in three different visual masking experiments, namely para-/meta-contrast masking, pattern masking by noise, and pattern masking by structure. We adopted statistical models, which have been used previously in studies of visual short-term memory, to capture response characteristics of observers under masking conditions. We tested the following scenarios: (i) mask activity may reduce a target's signal-to-noise ratio (SNR) without interfering with its encoding precision. (ii) Mask activity may "interfere" with the encoding of a target and cause decreased precision in observer's reports. (iii) Decreased performance due to masking may result from the confusion or "misbinding" of a mask's features with those of the target, when they are similar as in the case of pattern masking by structure. Our results show that in all three types of masking, the reduction of a target's SNR was the primary process whereby masking occurred. A significant decrease, correlated with masking strength, in the precision of the target's encoding was observed in para-/meta-contrast and pattern masking by structure, but not in pattern masking by noise. We interpret these findings as the mask reducing the target's SNR (i) by suppressing or interrupting the signal of the target in para-/meta-contrast, (ii) by increasing noise in pattern masking by noise, and (iii) a combination of the two in pattern masking by structure.

Metacontrast masking and attention do not interact

“…Therefore, we determined the number of trials required to reject the null hypothesis (i.e., the transformed performances with and without a mask are equal) by a two-sample t-test with a power level larger than 0.7. In a previous study, we found that the average standard deviation of transformed-performance across observers is roughly 0.15 (Agaoglu et al, 2015). We also found small-to-moderate effect sizes (defined as the difference between two conditions divided by the standard deviation of one group (Cohen, 1988)) between baseline and weak masking conditions.…”

“…In sum, these findings support the aforementioned informal observations that there is no clear or consistent trend across observers in the dependence of the standard deviation parameter on SOA and set-size. In our previous work (Agaoglu et al, 2015), we found that both the standard deviation of the Gaussian and the weight of the uniform distribution in the GU model correlated with the metacontrast function. The correlation of the weight parameter was higher than the correlation of the standard deviation.…”

“…We adopted the statistical models that have been previously used in modeling VSTM (Bays, Catalao, & Husain, 2009;Zhang & Luck, 2008) and several visual phenomena such as crowding (Ester, Zilber, & Serences, 2015) and masking (Agaoglu, Agaoglu, Breitmeyer, & Ogmen, 2015;Harrison, Rajsic, & Wilson, 2014). The simplest model is a single Gaussian (referred to as the G model from now on) whose mean and standard deviation may be modulated by attention and/or masking.…”

“…The mean of the target Gaussian was set to zero, i.e., centered on target orientation. This was motivated by our recent study on masking where we found that the mean of the Gaussian is not significantly different from zero (Agaoglu et al, 2015). Therefore, in the following analyses, the target Gaussians were centered on target orientations (i.e., zero mean in error space).…”

Can (should) theories of crowding be unified?

“…We used Bayesian model comparison (BMC) technique (MacKay, 2003; Wasserman, 2000) to select the best statistical model to capture the data (for a step-by-step derivation of this technique, see Agaoglu, Agaoglu, Breitmeyer, & Ogmen, 2015, or Ester et al, 2015). In short, this technique computes the likelihood of data given a certain model, penalizes models with more free parameters, and assigns proportionately more weight to the penalizing factor for larger data sets in estimating the model performance.…”