Learning features in a complex and changing environment: A distribution-based framework for visual attention and vision in general

What are the building blocks of our visual representations? Whatever we look at, the things we see will have some feature variability: even snow is not purely white but has a range of shades of white. However, in most studies investigating visual perception, homogeneous displays with all stimuli having a very limited range of features have been used. In contrast, recent studies using heterogeneous displays have shown that our perceptual system encodes surprisingly detailed information about stimuli, representing parameters such as the mean, variance, and most importantly the probability density functions of feature distributions. Learning the parameters of the distributions takes time as distribution representations are continuously updated with incoming information. However, the mechanisms guiding this process are not yet known. We will review current knowledge about the sampling and updating of representations of feature distributions in heterogeneous displays and will present new findings providing further insights into this process. Overall, the results show that representations of distributions can be remarkably detailed and shed light on how the information provided affects the learning of feature distributions. Observers’ ability to quickly encode the probability density function of distributions in the environment may potentially provide novel interpretations of a number of well-known phenomena in visual perception.