Abstract
Within a few hundred milliseconds, we transform light falling on our retina into percepts of visual scenes. The computational efficiency of this process may be explained by the presence of a rapid "gist" percept that accompanies detailed feature extraction. It is unclear, however, 1) how gist information, often described using biologically implausible image transformations, is computed by the visual system and 2) when and under what circumstances this information is extracted. We addressed these questions using a novel method that combines computational modeling of visual responses with regression analysis on single-image EEG activity. Subjects judged a specific instance of a gist property, "naturalness", for a large set of natural images. For each scene, we computationally estimated visual responses to local contrast, from which we derived two summary parameters: contrast energy and spatial coherence. Behaviorally, we found that spatial coherence correlated with perceived naturalness, whereas contrast energy correlated with reaction times. We then tested whether these two parameters predicted brain activity over the course of visual processing. In experiment 1, we found that contrast energy and spatial coherence predicted differences between single-image event-related potentials (sERPs) both early (80-150 ms) and later (180-250 ms) in time. In experiment 2, we manipulated task relevance, showing that early correlations with sERP differences persist when an orthogonal task is performed, whereas late correlations are only present when gist categorization is required. These results show that 1) gist information in natural scenes can be approximated by integrating local contrast responses, present in areas like LGN or V1, and 2) this process consists of at least two stages: an early, automatic stage and a later, task-dependent stage. Remarkably, this suggests that information maintained by top-down influences can still be described by summary parameters derived from the stimulus, supporting their relevance for perceptual decision-making about visual gist.
Meeting abstract presented at VSS 2013