Constructing modular architectures with Boltzmann machines

A striking feature of animal brains is their modular organization. For example, different brain areas exist for visual processing, auditory processing as well as areas for coordination of motor tasks. The majority of modeling work in the visual areas considers uni-directional connections between areas, creating multi-layered perceptron like architectures. The notion of receptive field is strongly dependent on this feed-forward assumption. However, anatomical studies show, that there are extensive feedback connections between areas as well. The result of such symmetric connectivity is also evident from observed correlations between the firing patterns of neurons in different areas at zero or small time delay (Engel et al. 1991). Although learning for feed-forward networks is well established, learning for networks of recurrently connected modular networks is less understood.