| The ventral visual pathway is widely viewed as having a hierarchical, feed-forward structure, in which individual layers attempt to represent a basis set of visual primitives (Field and Olshausen, 1997; Bell and Sejnowski, 1997). As information passes through these layers, receptive fields become larger and the complexity of the visual primitives increases until neurons become selective for visual objects such as houses or faces. This feed-forward view of the ventral visual pathway has been highly successful, but the mechanisms for top-down feedback in the ventral visual cortex are not as well understood. In this dissertation, we explore a controversial mechanism for top-down supervised learning in the ventral visual pathway.;Previous behavioral and neurophysiological studies have demonstrated that visual category learning results in the enhancement of relevant and the suppression of irrelevant visual information (Goldstone and Steyvers, 2001; Sigala and Logothetis, 2002; Meyers et al., 2008). The learning mechanisms that give rise to these effects, however, are not fully understood. The present state of the literature argues for two hypotheses, which explain these effects by appealing to either supervised learning or selective attention. First, it is possible that task feedback provides a supervisory learning signal that affects representations in visual cortex, changing them to reflect the behavioral relevance of each feature. This supervised learning mechanism would result in persistent task-independent enhancement of visual information that is relevant to making categorical distinctions. Alternatively, it is possible that the learning is limited to the allocation of attention and does not affect the underlying visual representations. According to this hypothesis, enhancement of relevant visual features is a transient property of visual cortex, driven by a dynamic allocation of selective attention. The key distinction between these hypotheses lies in whether category learning results in persistent, task-independent changes in the representations in the ventral visual stream.;In this dissertation, we investigate the learning mechanisms that shape visual representations by determining whether category learning results in persistent, task-independent enhancement of relevant features, and by identifying the neural correlates of these effects. We hypothesize that category learning results in persistent, task-independent modulation of the selectivity of neural populations in the ventral visual pathway. In addressing this hypothesis, we report the results of a behavioral experiment and two functional imaging experiments. We additionally discuss a novel methodology for estimating the mutual information for multivariate spheres of fMRI data. The results of these studies are discussed in the context of the larger debate surrounding learning algorithms in the ventral visual pathway. |
