Modeling neural dynamics caused by visual perceptual learning
Visual perceptual learning can be seen as a modification of the visual system which improves the performance in a visual task (Sasaki et al., 2009). To observe change in the neural activities, stimuli usually have to be shown for many thousands of trials indicating that the visual system is much less plastic in adulthood than during development. In the visual cortex, perceptual learning has been observed while performing certain visual tasks, for instance, a task requiring a fine orientation discrimination (Schoups et al., 2001; Zivari Adab and Vogels, 2011), or requiring the ability of contour integration (Li et al., 2004; Li et al., 2006).
Interestingly, it was found recently that performance improvement in such visual tasks in macaque monkeys is accompanied by changes in the response properties of local neural populations which can be observed after days of training in the primary visual cortex(Li et al., 2004; Li et al., 2006). However, since a whole hierarchy of brain areas, from primary visual cortex up to cognitive decision making areas, are involved during these tasks, and because higher visual areas (closer to the decision making areas) are reciprocally connected to early visual areas via feedback connections (Ahissar et al., 2009; Hochstein et al., 2002), it is still hotly debated where and how in the visual system modifications are made; that is, the neural mechanisms of perceptual learning remain unclear.
Mathematical ideas and abstract models have shown that task improvements can be achieved in principle if weights are adapted in abstract “channels” (Dosher and Lu, 1998) or by using back-propagation-like algorithms in artificial neural network models (Roelfsema and Ooyen, 2005; Roelfsema et al., 2010). However, since the interaction of multiple areas are involved while performing the tasks, it is difficult to pin down whether measured changes in the local activity in early visual cortices are due to local synaptic modifications or due to modification of feedback connections originating from higher areas. Current mathematical models of the learning process cannot answer this question, because they are purely conceptual, and thus cannot be related directly to the actual neural dynamics and activities as measured by experiments.
In this sub-project, we plan to use computational modeling to understand the dynamical changes in neural population activity induced by visual perceptual learning.