Reply to Pachai et al.

Abstract

Peripheral vision is fundamentally limited by the spacing between objects. When asked to report a target's identity, observers make erroneous reports that sometimes match the identity of a nearby distractor and sometimes match a combination of target and distractor features. The classification of these errors has previously been used to support competing 'substitution' [1] or 'averaging' [2] models of the phenomenon known as 'visual crowding'. We recently proposed a single model in which both classes of error occur because observers make their reports by sampling from a biologically-plausible population of weighted responses within a region of space around the target [3]. It is critical to note that there is no probabilistic substitution or averaging process in our model; instead, we argue that neither substitution nor averaging occur, but that these are misclassifications of the distribution of reports that emerge when a population response distribution is sampled. This is a fundamentally different way of thinking about crowding, and on this basis we claim to have provided a mechanism unifying categorically distinct perceptual errors. Our goal was not to model all crowding phenomena, such as the release from crowding when target and flanks differ in color or depth [4]. Pachai et al.[5] have suggested that our model is not unifying because it inaccurately predicts perceptual performance for a particular stimulus. Although we agree that our model does not predict their data, this specific demonstration overlooks the critical aspect of the model: perceptual reports are drawn from a weighted population code. We show that Pachai et al.'s [5] own data actually provide evidence for the population code we have described [3], and we suggest a biologically-plausible analysis of their stimuli that provides a computational basis for their 'grouping' account of crowding.

Can (should) theories of crowding be unified?

“…In a recent critique, Pachai et al (2016) demonstrated that crowding is reduced when the target C is surrounded by multiple flankers with the same orientation rather than just one, and the HB model fails to account for this reduction in crowding. In their reply, Harrison and Bex (2016) claimed that Pachai et al (2016) overlooked the critical aspect of their model, the weighting field. Moreover, the authors proposed that the apparent reduction of crowding with multiple surrounding flankers (having the same orientation) can be accounted for by a simple change in the front end of their model.…”

“…Moreover, the authors proposed that the apparent reduction of crowding with multiple surrounding flankers (having the same orientation) can be accounted for by a simple change in the front end of their model. Instead of using a bank of orientation-tuned filters in the first stage, Harrison and Bex (2016) used a filter-rectify-filter model of early visual texture processing (Bergen & Landy, 1991) and claimed that this updated version of their model can account for the data put forth by Pachai and colleagues (2016). Although Harrison and Bex (2016) did not perform any quantitative comparisons, here our aim was not to determine what type of front end would be best in a crowding model.…”

“…Instead of using a bank of orientation-tuned filters in the first stage, Harrison and Bex (2016) used a filter-rectify-filter model of early visual texture processing (Bergen & Landy, 1991) and claimed that this updated version of their model can account for the data put forth by Pachai and colleagues (2016). Although Harrison and Bex (2016) did not perform any quantitative comparisons, here our aim was not to determine what type of front end would be best in a crowding model. Instead, we focused on the backbone of the HB model, the weighting field, and showed that regardless of the type of front end used, the weighting field approach fails to predict perceptual errors made for inner and outer Cs.…”

“…In short, their model consists of three stages: (a) a filtering stage, (b) a weighted-averaging stage, and (c) a decision-making stage. The essential property of the HB model, as the authors emphasized more recently (Harrison & Bex, 2016; Pachai, Doerig, & Herzog, 2016), is the weighting field centered on the target object. The contrast energy obtained from the first stage is weighted such that the contribution of a feature (regardless of its ownership) to the final percept decreases with its distance from the center of the target object.…”

Visual crowding is a combination of an increase of positional uncertainty, source confusion, and featural averaging

“…Indeed, even the term crowding refers to the result of some visual process and not a mechanism. The underlying cause of crowding has previously been explained by various computational models1123404142474849 and higher-level mechanistic hypotheses2450. Population code models, in which all visual features probabilistically contribute to perceptual reports, can produce a wide variety of data51, including so-called averaging and substitution errors41.…”