Isolating objective and subjective filling-in using the drift diffusion model.

Spatial context is known to influence the behavioral sensitivity (d') and the decision criterion (c) when detecting low-contrast targets. Of interest here is the effect on the decision criterion. Polat and Sagi (2007) demonstrated that, for a Gabor target positioned between two similar co-aligned high-contrast flankers, the observers' reports of seeing the target (Hit and False Alarm) decreased with increasing target-flanker distance. This effect was more pronounced when the distance was randomized within testing blocks compared to when it was fixed. According to signal detection theory (SDT), the latter result suggests that the decision criterion is adjusted to a specific distance-dependent combination of signal (S) and noise (N) when the S and N statistics are fixed, but not when they vary across trials. However, SDT cannot differentiate between changes in the decision bias (the criterion shift) and changes introduced by variations in S and N (the signal and noise shift). To circumvent this limitation of SDT, we analyzed the reaction time (RT) data within the framework of the drift diffusion model (DDM). We performed an RT analysis of the target-flanker interactions using data from Polat and Sagi (2007) and Zomet et al. (2008; 2016). The analysis revealed a stronger dependence on flankers for faster RTs and a weaker dependence for slower RTs. The results can be explained by DDM, where an evidence accumulation process depends on the flankers via a change in the rate of the evidence (signal and noise shift) and on observers' prior knowledge via a change in the starting point (criterion shift), leading to RT-independent and RT-dependent effects, respectively. The RT-independent distance-dependent response bias is attributed to the observers' inability to learn multiple internal distributions required to accommodate the distance-dependent effects of the flankers on both the signal and noise.

ImageNet-trained deep neural networks exhibit illusion-like response to the Scintillating grid.

Deep neural network (DNN) models for computer vision are capable of human-level object recognition. Consequently, similarities between DNN and human vision are of interest. Here, we characterize DNN representations of Scintillating grid visual illusion images in which white disks are perceived to be partially black. Specifically, we use VGG-19 and ResNet-101 DNN models that were trained for image classification and consider the representational dissimilarity (\(L^1\) distance in the penultimate layer) between pairs of images: one with white Scintillating grid disks and the other with disks of decreasing luminance levels. Results showed a nonmonotonic relation, such that decreasing disk luminance led to an increase and subsequently a decrease in representational dissimilarity. That is, the Scintillating grid image with white disks was closer, in terms of the representation, to images with black disks than images with gray disks. In control nonillusion images, such nonmonotonicity was rare. These results suggest that nonmonotonicity in a deep computational representation is a potential test for illusion-like response geometry in DNN models.

A decision-time account of individual variability in context-dependent orientation estimation.

Following exposure to an oriented stimulus, the perceived orientation is slightly shifted, a phenomenon termed the tilt aftereffect (TAE). This estimation bias, as well as other context-dependent biases, is speculated to reflect statistical mechanisms of inference that optimize visual processing. Importantly, although measured biases are extremely robust in the population, the magnitude of individual bias can be extremely variable. For example, measuring different individuals may result in TAE magnitudes that differ by a factor of 5. Such findings appear to challenge the accounts of bias in terms of learned statistics: is inference so different across individuals? Here, we found that a strong correlation exists between reaction time and TAE, with slower individuals having much less TAE. In the tilt illusion, the spatial analogue of the TAE, we found a similar, though weaker, correlation. These findings can be explained by a theory predicting that bias, caused by a change in the initial conditions of evidence accumulation (e.g., priors), decreases with decision time (*Communications Biology 3 (2020) 1-12). We contend that the context-dependence of visual processing is more homogeneous in the population than was previously thought, with the measured variability of perceptual bias explained, at least in part, by the flexibility of decision-making. Homogeneity in processing might reflect the similarity of the learned statistics.

Interaction of contexts in context-dependent orientation estimation.

The processing of a visual stimulus is known to be influenced by the statistics in recent visual history and by the stimulus' visual surround. Such contextual influences lead to perceptually salient phenomena, such as the tilt aftereffect and the tilt illusion. Despite much research on the influence of an isolated context, it is not clear how multiple, possibly competing sources of contextual influence interact. Here, using psychophysical methods, we compared the combined influence of multiple contexts to the sum of the isolated context influences. The results showed large deviations from linear additivity for adjacent or overlapping contexts, and remarkably, clear additivity when the contexts were sufficiently separated. Specifically, for adjacent or overlapping contexts, the combined effect was often lower than the sum of the isolated component effects (sub-additivity), or was more influenced by one component than another (selection). For contexts that were separated in time (600 ms), the combined effect measured the exact sum of the isolated component effects (in degrees of bias). Overall, the results imply an initial compressive transformation during visual processing, followed by selection between the processed parts.

Perceptual bias is reduced with longer reaction times during visual discrimination.

Fast and slow decisions exhibit distinct behavioral properties, such as the presence of decision bias in faster but not slower responses. This dichotomy is currently explained by assuming that distinct cognitive processes map to separate brain mechanisms. Here, we suggest an alternative single-process account based on the stochastic properties of decision processes. Our experimental results show perceptual biases in a variety of tasks (specifically: learned priors, tilt aftereffect, and tilt illusion) that are much reduced with increasing reaction time. To account for this, we consider a simple yet general explanation: prior and noisy decision-related evidence are integrated serially, with evidence and noise accumulating over time (as in the standard drift diffusion model). With time, owing to noise accumulation, the prior effect is predicted to diminish. This illustrates that a clear behavioral separation-presence vs. absence of bias-may reflect a simple stochastic mechanism.

Expectations and visual aftereffects.

The tilt aftereffect (TAE) is traditionally regarded as a consequence of orientation-selective sensory adaptation, a low-level stimulus-driven process. Adaptation has been recently suggested to be the outcome of predictive coding. Here, we tested whether the TAE is modulated by predictability, and specifically, whether TAE depends on the congruency of adapted and expected orientations. Observers were presented with successive pairs of oriented Gabor patches. Pairs were arranged in blocks, forming two conditions with the orientation of the second pair member either predictable or not. For all pairs, the orientation of the first Gabor was tilted clockwise (CW) or counterclockwise (CCW) (±20° relative to vertical, randomized). In the "Expected" conditions, the orientation of the second Gabor was fixed relative to the first Gabor (the same or a mirror orientation, blocked). In the "no-expectation" condition, the orientation of the second Gabor was independent of the first Gabor (randomized ±20°). Intermixed test pairs were used to measure observers' perceived vertical, with the second pair member serving as a target, oriented around the vertical, permitting an estimate of the TAE produced by the presentation of the first Gabor. Results show an increase in TAE with the expected orientation matching the inducing orientation, but a decrease with the expected mirror orientation, consistent with additivity of the adaptation and the expectation effects. A second experiment, with the first oriented Gabor replaced by a colored circular blob, showed that expectation alone does not modulate the perceived orientation. These findings indicate a role for expectation in generating the perceptual TAE and are in line with predictive coding models of perception. We suggest that orientation dependent adaptation is affected by both the mean orientation (first order statistics) and by temporal contingencies (second order statistics).

Tilt aftereffect due to adaptation to natural stimuli.

The human visual system continuously adjusts to the current environment. To investigate these adjustments, biases in observers' perceptions owing to changes in the visual environment are measured (visual aftereffects). Typically, the stimuli used are synthetic and are composed of oriented patterns such as lines or gratings. These patterns are known to activate individual neurons in the visual cortex, but cover only a small subset of actual visual stimulations. To overcome this drawback, recent research has focused on synthetic patterns that mimic several aspects of natural stimulation. However, the aftereffects of natural stimulation per-se remain largely unexplored. Here, we interleaved presentations of unmodified natural image adaptors, selected according to criteria favoring content at a particular orientation, with presentations of targets that test a perceived orientation. This allowed us to measure the change in the perceived orientation, namely the tilt aftereffect (TAE), which resulted from repeated image presentations. Results show a close to standard TAE with adaptor durations around 500ms, which is reduced with longer presentations. Importantly, our method can be generalized to investigate other aftereffects by selecting images differently.