Element predictability not high occurrence frequency determines feature learning from multi-element scenes

Previous studies have demonstrated that humans, by mere exposure, can become sensitive to clusters of elements (objects) from multi-element visual scenes. However, a key question in object recognition - which of several statistics of the scene (e.g., element relative frequency or conditional probability) humans prefer for breaking complex unfamiliar objects into subparts - has not been investigated before. We addressed this question by presenting naïve observers with 184 unfamiliar artificial multi-element scenes, where the occurrence frequency and conditional probability (predictability) statistics were different across the participating elements. Subjects passively viewed each scene for 2 second during a familiarization session. Each scene consisted of 7 geometric shapes positioned apparently randomly on a 5×5 grid. However, each scene was generated with one of two consistent 6-tuples (six elements always appearing in a given configuration) and a noise element so that the participating elements of the 6-tuples had different occurrence frequencies and predictabilities. In particular, some elements of the 6-tuples appeared 1.5 times more often during familiarization, but were less predictive of neighboring elements, while some less frequent elements predicted their neighbors better. A 2AFC post-exposure test revealed that subjects remembered the embedded shape-pairs of the 6-tuple with lower element occurrence but higher predictability better than shape-pairs with higher element occurrence and lower predictability [t(29)=2.25, p<.05]. This result provides direct evidence that, other dimensions being controlled, encoding the sub-features of a larger coherent visual structure is determined by the predictability between elements rather than their higher occurrence frequency in the scenes. Moreover, elements with lower predictability in the structure serve as breakpoints for "chunking" the object structure into parts.