Abstract
Many statistical methods exist for decomposing visual stimuli into collections of informative features. Which features contain information that human observers are sensitive to? We examine this question in the context of a speed discrimination task with naturalistic stimuli. Overall stimulus contrast impacts speed sensitivity, but it is less well known how the presence (or absence) of features defined by different contrast patterns modulates sensitivity. Here, we directly examine how different features impact human speed perception. To do so, we composed stimuli from two feature libraries learned from natural stimuli. The first library, determined using principal components analysis (PCA), emphasizes stimulus reconstruction; it consists of rank-ordered features that best account for contrast in natural image movies. The second library, determined using accuracy maximization analysis (AMA), emphasizes task-relevance; it consists of rank-ordered features that are most useful for estimating speed in natural image movies. We measured human foveal speed sensitivity in a two-alternative forced choice discrimination experiment. Experimental stimuli (~3°/s; 1×1°; 250ms) were composed by first filtering and then reconstructing natural stimuli with features from each library. Multiple results emerge. First, when stimulus contrast is held constant, stimuli with more task-relevant features yield better discrimination performance. Second, when contrast is mismatched between stimuli, lower contrast stimuli with more task-relevant features yield better performance than higher contrast stimuli that more closely approximate the original images. Third, when uninformative features are added to stimuli that already contain task-relevant features, discrimination performance decreases dramatically. Thus, task-relevance is a critical variable for predicting the outcome of additional information on performance. The current study draws an explicit link between human visual performance and principled statistical methods for assessing the task relevance of different stimulus features. Our data suggest that for speed discrimination, human observers are indeed sensitive to speed-relevant features identified by these principled methods.
Acknowledgement: R01-EY028571