The response properties of neurons in many sensory cortical areas are ordered topographically, i.e. nearby neurons respond to nearby areas of the receptor surface. Such topographic maps form by the self-organization of afferent connections to the cortex, driven by external input [5,11,16,18]. Several neural network models have demonstrated how the global topographic order can emerge from local cooperative and competitive lateral interactions within the cortex [1,7,8,10,20]). These models are based on predetermined lateral interaction and focus on explaining how the afferent connections become ordered.
A number of recent neurobiological experiments indicate that lateral connections self-organize like the afferent connections: (1) The lateral connectivity is not uniform or genetically predetermined, but forms during the early development based on external input [6,9]. (2) In the primary visual cortex, lateral connections are initially widespread, but develop into clustered patches at the same time as the orientation and ocular dominance columns form [2,6,3]. (3) Lateral connections primarily connect areas with similar response properties, such as columns with the same orientation or eye preference [4,9]. To fully account for cortical self-organization, a cortical map model must demonstrate that both afferent and lateral connections can organize simultaneously, from the same external input, and in a mutually supportive manner.
We have previously shown how Kohonen's self-organizing feature maps [7] can be generalized to include self-organizing lateral connections (the Laterally Interconnected Synergetically Self-Organizing Map (LISSOM); [14,15]). LISSOM is a low-dimensional abstraction of cortical self-organizing processes and models a small region of the cortex where all neurons receive the same input vector. In contrast, this paper shows how realistic, high-dimensional receptive fields develop as part of the self-organization, and in essence scales up the LISSOM approach to large areas of the cortex where different parts of the cortical network receive inputs from different parts of the receptor surface. The new model shows how (1) topographically ordered receptive fields develop from simple retinal images, (2) lateral connections self-organize cooperatively and simultaneously with the afferents, (3) long-range lateral connections store correlations in activity across the topographic map (e.g. low-level Gestalt rules), and (4) the resulting lateral connection patterns closely follow receptive field properties such as ocular dominance.