József Fiser¹, Zoltán Rácskay; ¹Central European University

A staple feature of object representation and perception is invariance: stable representation of new objects is learned from perpetually varying visual inputs and objects are identified as same despite of large differences in the appearance of their local features. Yet, visual statistical learning (VSL) paradigms investigating the formation of such object representations typically use static images of standardized visual patterns as local features. In three experiments using a Virtual Reality environment, we explored how human VSL is affected when the stable visual structures to be learned appear in a more natural dynamic manner by where the classical spatial VSL patterns appeared in a plane that perpetually changed its 3D position, rolled, pitched and yawed in a random fashion. First, we replicated the classical static results showing that adults have a significant familiarity preference (59%) for shape-pair structures that were used to generate composite scenes in the preceding task-free exposure phase of the experiment. Next, we rerun the experiment with moderate dynamic movement of the plane during the exposure phase and found that preference became indistinguishable from chance performance indicating the disappearance of implicit learning under these conditions. In the third experiment, we increased further the magnitude of the dynamic movement of the plane holding the composite patterns. Surprisingly, instead of staying in the ”no-learning” regime, participants developed a significant familiarity preference for the alternative lure shape pairs during the test trials (40%). This preference was significantly different not only from chance performance but also from the preference in the moderate condition. These results suggest that instead of being simply eliminated, the effects of dynamic variations in the input during exposure are associated with the features and can take part in active interference during the familiarity judgement.

Acknowledgements: This work has been supported by Grant ANR-FWF I 6793-B.