Feedback-driven tuning of statistical summary representations

Abstract

The amount of sensory data encountered by the visual system often exceeds its processing capacity. One solution is to exploit statistical structure in the natural environment to generate a more efficient representation of the information (Simoncelli & Olshausen, 2001). For example, the visual system may construct a “statistical summary representation” over groups of visual objects, reflecting their general properties (Alvarez, 2011). Indeed, it has been shown that observers are able to quickly and accurately extract average values over a range of visual feature dimensions, including size (Chong & Treisman, 2003), orientation (Parkes et al. 2001), and emotional expression (Haberman & Whitney, 2007). However, it remains an open question as to how observers learn to produce such accurate estimates of these summary statistics. Although good performance on these tasks suggests that summary features are readily accessible, it is not clear to what extent these statistical operations are performed automatically—integrating over sensory information in an unsupervised fashion, or are penetrable to task demands—flexibly incorporating observer goals and error-related feedback to maximize performance (Bauer, 2009; Myczek & Simons, 2008). In the present study, we sought to understand the role of learning in statistical summary representations. Specifically, we examined the contribution of task practice and performance feedback to perceptual discrimination of the centroid (i.e., mean location) of a set of objects (Alvarez & Oliva, 2008). We hypothesized that providing vector error feedback (i.e., containing both distance and direction information) while observers practiced making pointing movements toward the centroid would improve the fidelity of their centroid representations. This improvement might be reflected in reduced error during training and lower discrimination thresholds in an independent perceptual test.