Abstract
When detecting targets under natural conditions, the visual system almost always faces multiple, simultaneous, dimensions of extrinsic uncertainty. Each dimension of uncertainty (i.e., each dimension of random variation in background and target properties) has individual effects on behavioral performance. Measuring how these individual effects combine is essential for understanding the normal functioning of the visual system. Furthermore, an important real-world factor that interacts with uncertainty is target prior probability. Our aim is to measure and model the combined effects of the various dimensions of uncertainty, together with the effect of target prior probability. Here, we consider a simple task: detection of a horizontal windowed sinewave target with random amplitude in Gaussian white noise with random contrast, for target prior probabilities of 0.5 and 0.2. We examined these two dimensions of uncertainty first, because exact expressions for model performance can be derived. On each trial, the amplitude of the target and the contrast of the noise were randomly and independently sampled from uniform distributions having eight discrete levels that covered a broad range (i.e., each trial was a random sample from 64 conditions). Participants were asked to report whether the target is present or absent. We derived the performance of the ideal observer (IO), which has a dynamic decision criterion, a template-matching (TM) observer with a single criterion, and a contrast normalized template-matching (NTM) observer with a single criterion. Interestingly maximum-likelihood fits showed that human accuracy as a function of target amplitude and background contrast were well predicted by all three models, when the target prior was 0.5, but that only the IO and NTM observers were able to predict accuracy when the prior was 0.2. The results reveal the value contrast normalization under real-world conditions where target priors are low, and the value of manipulating target priors for discriminating models.
Acknowledgement: NIH grant EY024662