Abstract
Classical Hebbian models of learning depend on weighting sensory inputs on the basis of their association with the feature to be learned. Performance improvement due to reweighting mechanisms arises from a more optimal use of discriminating features in the stimuli. In the domain of perceptual learning, two distinct algorithms have been proposed for iterative reweighting of sensory data in the presence of feedback. The first algorithm is a Hebbian weighting increment that consists of the product of the sensory input and a feedback constant (Petrov et al., 2005; Dosher & Lu, 2009). The second is a Bayesian approach in which reweighting is seen as the process of using posterior probabilities, a non-linear function of the sensory data, as priors in subsequent trials (Eckstein et al., 2004; Trenti et al., 2009). The purpose of this study is to compare these models of reweighting with human-observers in a simple visual task. Three observers participated in yes-no detection of a contrast increment to a Gaussian spatial profile appearing at any one of four possible locations. Each location included additive Gaussian-distributed contrast noise. When present, the increment appeared at the same location for a block of four consecutive trials. This allowed observers the chance to learn the relevant location and thereby exclude irrelevant locations from consideration, leading to improved performance. At the end of each block, subjects were asked to identify the target location as a secondary measure of learning. We find that for this visual task, classical models of learning based on a simple sensory-feedback product do not capture the rapid gains in performance and magnitude of noise-classification weights (Ahumada, 2002) as well as the ideal Bayesian reweighting algorithm. Our results suggest that, at the behavioral level, the process of learning can incorporate optimal reweighting schemes that go beyond simple association.
Meeting abstract presented at VSS 2012