Stable individual signatures in object localization

In Brief

Kosovicheva et al. report large, stable individual differences when human observers judge the perceived locations of brief, static visual targets. These idiosyncratic localization signatures are consistent across different measures and stable over several months, suggesting that even basic visual judgments differ substantially between individuals.

Perceptual processes in human observers vary considerably across a number of domains, producing idiosyncratic biases in the appearance of ambiguous figures [1], faces [2], and a number of visual illusions [3–6]. This work has largely emphasized object and pattern recognition, which suggests that these are more likely to produce individual differences. However, the presence of substantial variation in the anatomy and physiology of the visual system [4,7,8] suggests that individual variations may be found in even more basic visual tasks. To support this idea, we demonstrate observer-specific biases in a fundamental visual task – object localization throughout the visual field. We show that localization judgments of briefly presented targets produce idiosyncratic signatures of perceptual distortions in each observer and suggest that even the most basic visual judgments, such as object location, can differ substantially from one individual to another.

To reveal this bias, observers (N=5; 2304 trials each) reported the location of a brief (50 ms), stationary random dot noise patch shown at one of 48 random angular stimulus positions along an invisible isoeccentric ring with a radius of 7 degrees of visual angle (d.v.a.; see Figure 1A and Supplemental Information). Subjects fixated the center of the display during the stimulus window, and then indicated perceived patch location by moving a cursor from the display center to the previously seen target location (“outward adjustment”), or by adjusting the position of a cursor constrained at an eccentricity of 7 d.v.a., starting from a random angular location (“angular adjustment”). In a separate session, to determine whether errors could be reproduced when the retinal location of the stimulus was dissociated from the retinal location of the cursor, subjects completed the outward adjustment method while moving their eyes freely during the response window. Finally, subjects completed a separate session in which they made a saccade as quickly as possible to the center of the target. For each of the four methods (see Fig. 1C legend), we calculated the mean angular difference between the subject’s response (or saccade landing location) on each trial and the angular location of the target center.

Figure 1. Stimulus arrangement and localization errors.

(A) On each trial, subjects were shown a brief (50 ms) noise patch at one of 48 randomly selected locations (the possible locations of the patch center, indicated by ‘x’s, and the markings shown in white were not visible to subjects). Subjects reported the patch location using one of four possible response methods (see Methods and panel C legend). The mean angular difference between the subject’s response and the true center of the noise patch was calculated for each location. (B) Mean response errors at each location for every subject (outward adjustment method, red: clockwise errors, blue: counterclockwise errors) reveal substantial individual variation in response error between subjects. (C) Response errors from a representative subject show a high degree of consistency across the four different response methods, shown in the legend (white dotted circles represent gaze location at time of response). Mean error is plotted as a function of angular location for each method, with positive values corresponding to clockwise errors, and negative values corresponding to counterclockwise errors. (D) To determine the degree of within-observer similarity, we correlated the errors for each response method (see legend in panel C) with the other three response methods within the same observer (solid bars). Between-observer similarity for each method was calculated by averaging all pairwise comparisons between subjects (hatched bars). Horizontal bars represent the upper bound of the central 95% of the permuted null distribution (see Supplemental Information). (E) To determine stability in subjects’ response errors over time, correlations between pairs of sessions within a subject were sorted into three bins based on their temporal separation (in weeks). Open circles represent individual correlations between pairs of sessions, and filled circles represent binned averages. Average correlations were calculated from individual Fisher z values and then transformed to Pearson’s r. Error bars represent bootstrapped 95% confidence intervals.

Figure 1B shows the errors from each observer in the outward adjustment response method at each location. Subjects’ errors revealed large, idiosyncratic mislocalizations, up to 9.15° (1.11 d.v.a) or 3.43 times the just-noticeable difference (JND) at a single location. Across the four response methods, the average absolute angular deviation was 4.94° (0.60 d.v.a.). To evaluate between-observer mislocalization similarity for each response method, we first calculated pairwise comparisons of errors at each of the 48 locations between observers (e.g., Subject 1’s error at the 90° location compared to Subject 2’s error at 90°, and so on for each location) and then computed the average of all pairs of subjects. This analysis produced weak inter-observer correlations across the four response methods (Figure 1D), significant only in the angular adjustment condition (p = 0.004, all other p-values > 0.10, permutation test using a Bonferroni-corrected alpha, α_B = 0.006; see Supplemental Procedures).

In contrast, response errors within any individual observer were highly consistent across the four response methods (see Figure 1C for errors from a single subject). To quantify this degree of similarity for each response method, we correlated the errors from one response method with the other three within an observer, and then averaged the resulting values. Average within-observer correlations for each response method (see Figure 1D) were significantly greater than those expected by chance (all p-values < 0.001 based on permutation tests; α_B = 0.006, see Supplemental Procedures). In addition, we assessed the stability of each observer’s localization signature over time by carrying out these measurements over the course of several months. Figure 1E shows the correlations between all pairs of sessions within an observer as a function of the length of time separating them (mean: 11.0 weeks, range: 0–24 weeks), sorted into three time bins. Mean correlations indicated a high degree of stability over time, with significant correlations within each time bin (all p-values < 0.001; α_B = 0.017).

The stability of subjects’ errors over time, and across different types of adjustments (e.g., circular, outward) suggests that they are unlikely to be a product of motor response biases. We further excluded the possibility of response bias in a second experiment, in which subjects reported patch position relative to a stable reference dot in a two-alternative forced choice (2AFC) task. Subjects’ responses in this task indicated that the patch appeared aligned with the reference dot only when they were physically misaligned, in a pattern consistent with their individual errors in the main experiment (Figure S1).

If there are systematic localization errors, why is the perceived cursor position unaffected? The presence of identical perceptual shifts in the perceived location of the noise patch and cursor should cancel out any measurable error. One possibility is that these localization errors emerge under spatial or temporal uncertainty—for instance, when the noise patch is briefly presented or spatially diffuse. We tested this by measuring subjects’ errors, varying both stimulus duration and size. When either spatial or temporal noise was reduced, such that the noise patch more closely resembled the cursor, the magnitude of the errors also decreased. Variations in patch size also shifted the pattern of errors, as indicated by reduced within-observer correlations across different patch sizes (Figure S2).

Our finding of stable, idiosyncratic localization signatures overturns long-standing assumptions about perceptual judgments of basic visual attributes – that they are homogenous across individual observers, and invariant to retinal location within an observer. While it is often assumed that different observers generally agree about the locations of objects, our results demonstrate that this judgment can result in wildly different responses across individuals. Moreover, these errors were reproduced with saccadic responses, similar to correlations between perception and action observed in some illusions [6,9]. As saccadic responses and cursor adjustment responses occur on very different timescales, the observed errors are unlikely to be memory-driven, and we observe no correlation between reaction time and the magnitude of response errors (Figure S2). The stability of the errors observed suggests that these biases may be anatomically driven [10], similar to previously reported relationships between V1 anatomy and perceived size [4]. Further work will be needed to determine the anatomical locus of these errors, and to establish any relationship between these low-level biases and more cognitive, high-level effects [1].

Supplementary Material

s1

Document S1. Experimental Procedures and Two Figures