Purchase this article with an account.
Phillip Isola, Nicholas Turk-Browne, Brian Scholl; Multidimensional visual statistical learning. Journal of Vision 2007;7(9):43. doi: https://doi.org/10.1167/7.9.43.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Statistical relationships between objects in space and time are automatically extracted via visual statistical learning (VSL). Such processing has traditionally been thought to operate over visual objects, but visual input is also highly structured at the level of individual surface features. Here we studied VSL in temporal sequences of colored shapes, exploring how features are combined into objects. Observers were familiarized to sequences of colored shapes that appeared one at a time, with statistical regularities present in the order of repeated shape subsequences. In Experiment 1, half of these were bound-color subsequences, in which each shape was always presented in its own unique color; the other half were random-color subsequences, with colors randomly drawn upon each presentation from a different set of possible values. During a later test phase, observers repeatedly judged which of two shape subsequences — now presented all in black — was more familiar: one previously encountered during familiarization vs. a misordered foil subsequence constructed from the same shapes. Observers reliably chose the previously encountered subsequences for both bound-color and random-color conditions. Since each shape had been encountered equally often, this performance must reflect learning of the shapes'; statistical ordering. Moreover, performance in these conditions did not differ, suggesting that the covariance between individual feature values did not affect the expression of VSL for black shapes. In Experiment 2, however, familiarization consisted of only bound-color subsequences, and performance at test with black shapes was significantly (and surprisingly) lower. Thus, color appears to have been more integral for the learned representations of the bound-color subsequences in Experiment 2 — compared to those same subsequences in Experiment 1, which were encountered in the context of additional random-color subsequences. In sum, what determines the input to VSL is the diagnosticity of feature dimensions, not only of individual feature values.
This PDF is available to Subscribers Only