The long term goals of this proposal are to build a computational model for the activity dependent development of hierarchical cortical structure. This model will be used to elucidate how anatomical and functional separation, combination and abstraction of afferent information may be controlled, and how various activity independent and dependent mechanisms could interact. Understanding these processes will provide Insight into the normal mechanisms for interpreting input and abnormal development in the face of peripheral or central nervous dysfunction. Cortical processing of afferent input is hierarchical, with more abstract information being extracted and represented in areas higher in the hierarchy. Recent experiments have focused on the anatomical and functional properties of some of these higher areas and the connections between them, and have revealed salient architectural features such as rich reciprocal connectivity between areas. Activity-dependent mechanisms are known to play a significant role in the development of primary sensory areas in vertebrate cortex, and have been a substantial focus for computational modeling. Unfortunately, most existing models have focused exclusively on primary sensory cortices, ignoring key features of hierarchical processing, such as reciprocal connections. It is not even clear that they would work in the face of such connections. Further, there are no existing models that reveal how the statistical structure of the inputs could be extracted in a hierarchical manner. The proposal is based on a recent theoretical model for hierarchical unsupervised learning, which uses a simple local rule for synaptic adaptation, embodies well-founded statistical principles and gives a precise role to reciprocal feedforward and feedback connections between areas. This model will be fleshed out in respect of structural properties of cortex, and tested using standard tasks for such models, including ocular dominance stripes and orientation domains. It will then be studied from the perspective of how properties of cells in different areas, of the inter- and intra-areal connections between them, and properties of the correlational structure of the inputs, can affect the development of the hierarchy. Area V2 and its reciprocal connections with V1 will be a main focus. Mathematical analysis and computer simulations will be used.
| Li, Z; Dayan, P (1999) Computational differences between asymmetrical and symmetrical networks. Network 10:59-77 |
| Dayan, P (1999) Recurrent sampling models for the Helmholtz machine. Neural Comput 11:653-78 |
| Zemel, R S; Dayan, P; Pouget, A (1998) Probabilistic interpretation of population codes. Neural Comput 10:403-30 |
| Dayan, P (1998) A hierarchical model of binocular rivalry. Neural Comput 10:1119-35 |
| Neal, R M; Dayan, P (1997) Factor analysis using delta-rule wake-sleep learning. Neural Comput 9:1781-803 |
