Attention is drawn spontaneously to regularities during statistical learning

摘要

The visual environment contains widespread regularities, but this structure represents only a subset of the complex and noisy input available at any given moment. The challenge for statistical learning is thus to identify what aspects of the environment to learn about. Here we propose that regularities themselves capture attention, prioritizing their own locations and features for further processing. In Experiment 1, we examined whether regularities cue spatial attention. Observers viewed four simultaneous streams of shapes. Unbeknownst to them, the stream in one a??Structureda?? location was generated from triplets, while the streams in three a??Randoma?? locations were randomized. To probe spatial attention, we presented occasional search arrays where the target appeared randomly at one of the shape locations. Target discrimination was reliably faster for targets at Structured vs. Random locations, suggesting prioritization of locations containing regularities. To generalize this finding, in Experiment 2 we examined whether regularities cue feature-based attention. Observers viewed a single stream at fixation containing red and green shapes. Shapes in the a??Structureda?? color appeared in triplets, while those in the a??Randoma?? color appeared in a randomized order. We probed feature-based attention with search arrays that now contained a color singleton: either a distractor or target appeared in either the Structured or Random color. Target discrimination was faster overall for target vs. distractor singletons as expected, but critically, this capture was significantly stronger for Structured color singletons, suggesting prioritization for features of objects embedded in regularities. These findings reveal a new type of automatic orienting to regularities, driven neither by inherent stimulus salience nor by intentional goals, which may in turn encourage further statistical learning about matching locations and features. Such orienting provides both a novel implicit and online measure of statistical learning, and a compelling demonstration of the influence of statistical learning over other parts of cognition.