Purchase this article with an account.
Benjamin Chin, Johannes Burge; A model grounded in natural scene statistics predicts human performance with both natural and artificial stimuli. Journal of Vision 2018;18(10):340. doi: 10.1167/18.10.340.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Models developed with artificial stimuli tend to generalize poorly to natural stimuli. Do models developed with natural stimuli generalize well to artificial stimuli? We examine this question in the context of motion estimation. Accurate estimation of self-motion and of the motion of objects in the environment is critical for survival and reproduction, but the visual system must first accurately estimate the motion of images on the retina. Previously, we developed an ideal observer for retinal speed estimation with natural image movies and used it to tightly predict human responses in a speed discrimination experiment with a single efficiency parameter (R2>0.95; Chin & Burge, 2017). We further showed that the value of the efficiency parameter nicely predicts human response agreement with natural stimuli in a double-pass experiment, a result predicted by the hypothesis that inefficiency is due only to internal noise. How well does this ideal observer predict human performance with artificial stimuli? Here, with zero additional free parameters, we challenged the ideal observer to predict human speed discrimination and human response agreement in a double-pass experiment. Each human observer performed a speed discrimination experiment with drifting Gabors in a 2IFC design (1deg, 250ms) using the method of constant stimuli, and each observer performed the experiment twice (2800 trials = 2 standard speeds x 7 levels/standard x 100 trials/level x 2 passes). The ideal observer predicts i) a 30% improvement in thresholds compared to natural stimuli, and ii) significantly less response agreement between passes. Both predictions are confirmed by the data. These results show that careful task-specific analysis of natural signals can provide powerful (and interpretable) models that predict human performance with both natural and artificial stimuli.
Meeting abstract presented at VSS 2018
This PDF is available to Subscribers Only