= Correlation-based development of visual cortical maps
Consider afferent activities $z_i$, cortical activity $y_a$ and a constrained <Hebbian learning> rule $\dot w_{ai}=\epsilon\langle y_az_i\rangle-\Lambda_a$. If $y_a=\sum_{b,j}K_{ab}w_{bj}z_j$, then $\langle y_az_i\rangle=\sum_{b,j}K_{ab}w_{bj}C_{ji}$, where $C$ is the input correlation matrix. Thus input correlations and lateral interactions select growing weight modes, while <synaptic normalization> and bounds stop unlimited growth. The model tests how organization can emerge without assigning every final cortical feature in advance.
Back to article page
