Center-surround patterns emerge as optimal predictors for human saccade targets

Abstract

The human visual system is foveated, that is, outside the central visual field resolution and acuity drop rapidly. Nonetheless much of a visual scene is perceived after only a few saccadic eye movements, suggesting an effective strategy for selecting saccade targets. It has been known for some time that local image structure at saccade targets influences the selection process. However, the question of what the most relevant visual features are is still under debate. Here we show that center-surround patterns emerge as the optimal solution for predicting saccade targets from their local image structure. The resulting model, a one-layer feed-forward network, is surprisingly simple compared to previously suggested models which assume much more complex computations such as multi-scale processing and multiple feature channels. Nevertheless, our model is equally predictive. Furthermore, our findings are consistent with neurophysiological hardware in the superior colliculus. Bottom-up visual saliency may thus not be computed cortically as has been thought previously.

Visual orienting in dynamic broadband (1/f) noise sequences

“…To ensure that the values were approximately comparable with those from other methods, we used the reverse-correlation method that was developed for the analysis of visual neurons (Simoncelli, Paninski, Pillow, & Schwartz, 2004), as well as the supportvector machine classification method by Kienzle, Franz, Schölkopf, and Wichmann (2009). With all of these methods, similar ROC area values were obtained that hinted that the values were independent of the method.…”

Does Cognitive Science Need Kernels?

“…Over the last years, we and several colleagues used kernel methods to identify those features that best predict a subject's response in psychophysical tasks with natural stimuli (Wichmann, Graf, Simoncelli, Bülthoff, & Schölkopf, 2005;Kienzle, Wichmann, Schölkopf, & Franz, 2007;Kienzle, Franz, Schölkopf, & Wichmann, 2009;Yovel, Franz, Stilz, & Schnitzler, 2008). Like other black-box methods, these methods substitute a very hard to analyze complex systemthe complete human observer-with a less complex system that is sufficiently sophisticated to re-create human decisions during a psychophysical task but is still amenable to mathematical analysis.…”

“…The features that were recovered in this example are neurophysiologically plausible: The more similar an image patch is to a center-surround structure (z 3 and z 4 ) and the less similar it is to a ramp (z 1 and z 2 ) the more likely it will be a saccade target. This figure is adapted from Kienzle et al (2009). rized according to Marr's levels of explanation: the computational, the algorithmic, and the implementational level (Marr, 1982). Kernel methods are related to RBF networks on the implementational level and exemplar models on the algorithmic level.…”

“…Recently, Kienzle et al (2007Kienzle et al ( , 2009 showed one natural extension of our approach to identifying non-linear features of a certain form. This work also illustrates that kernel methods are, of course, not limited to predicting subject's behavior in standard categorization tasks.…”

Computational Models of Human Visual Attention and Their Implementations: A Survey

“…Kienzle et al [77] proposed a framework to discover relevant visual features for saliency calculation using human gaze data. They assume that there are local image patterns (perceptive fields) that guide human gaze.…”