Abstract
Numerous studies have reported that human behavior on perceptual inference tasks – such as cue combination and visual search – is well accounted for by optimal models. However, others have argued that optimal models are often overly flexible and, therefore, lack explanatory power. In addition, it has been suggested that inference performed by neural systems is inherently noisy, which would preclude optimality in many perception tasks. Here, we reconsider human performance on visual search by devising an approach that strongly reduces model flexibility and tests for suboptimalities due to imprecisions in neural inference. Subjects performed a target detection task in which targets and distractors were ellipses with orientations drawn from Gaussian distributions with different means. We controlled the level of sensory uncertainty through stimulus presentation time (short vs. unlimited) and the elongation of the ellipses (low vs. high). Moreover, we created four levels of external uncertainty by varying the amount of overlap between the target and distractor distributions. Since sensory noise was negligible in the conditions with unlimited display time, we were able to estimate deviations from optimality without having to fit free parameters. In conditions with short display time, we limited the flexibility of the optimal model by using a separate task to estimate sensory noise levels. We found clear evidence for suboptimalities in all tested conditions. Moreover, we estimate that the performance loss due to computational imperfections was of comparable magnitude to the loss due to sensory noise. Our results provide support for the proposal that neural inference is inherently imprecise and challenge previous claims of optimality in perception.
Author summary The main task of perceptual systems is to create truthful representations of the world. They do so by using sensory information that is often astonishingly imprecise due to measurement errors in our senses. Consequently, it is often impossible to be 100% correct all the time on tasks that involve perception, such as judging whether a visual target is present in a cluttered scene. Observers are typically defined as optimal if they perform as well as theoretically possible given the sensory imprecisions. Numerous studies have reported that humans are optimal observers in perception-based tasks, but the validity of these findings has recently been questioned for two different reasons. First, it has been argued that a lot of the evidence is based on studies that used overly flexible models. Second, there are indications that inference performed by brains is inherently imprecise, due to limitations in the neural systems performing the inference. In this study, we reconsider optimality in perception by devising a research method that makes several improvements over previous studies. We apply this method to a visual search task and find clear indications of suboptimalities. Our findings imply that the perceptual systems may indeed not be as perfect as previously thought.