Abstract
Why do some visual stimuli look "random" and others more like a pattern? Conventional wisdom suggests that repeating, alternating, or symmetrical arrangements of features play a role in the perception of a pattern, but this evidence is often subjective and anecdotal. We aimed to more precisely characterize the factors that lead to the perception of randomness and pattern-ness starting with a simple set of visual stimuli: five-element sequences of horizontal and vertical Gabor patches arranged in a vertical column.
In our main experiment, we administered a two-alternative forced-choice task to subjects via Amazon’s Mechanical Turk platform. On each trial, subjects chose which of two presented sequences appeared most random. Subjects made over 10,000 behavioral responses across all 496 possible pairings of the 32 possible sequences. We fit the results with a simple quantitative model that projects each sequence onto a single dimension that ranges from perceived randomness to perceived pattern-ness. We found that by using only a few objective, calculable descriptions of the stimuli, we could predict the subjects’ choice data with extremely high accuracy and reliability. Specifically, we found that (1) low entropy, (2) a low probability of alternation, and (3) a lack of symmetry all contribute systematically to the likelihood that one stimulus will appear more like a pattern than another.
In two further behavioral experiments, we then show that the model generalizes very well to both longer and more abstractly represented stimulus sequences. Since our model can predict the perceived randomness or pattern-ness of any binary sequence, we also demonstrate an interesting application: given a binary sequence with a subset of pre-specified elements, we can fill in the remaining elements to produce a sequence that will appear either maximally random or maximally patterned to a typical human observer.
Meeting abstract presented at VSS 2013