This thesis is a first functional study of the RF-LISSOM model (Receptive-Field Laterally Interconnected Synergetically Self-Organizing Map) developed by Sirosh and Miikkulainen (1994a); Sirosh and Miikkulainen (1996, 1997). This chapter describes the architecture of the model in detail, summarizes previous results using RF-LISSOM to model the cortex, and examines the biological underpinnings of the model. Later chapters will show how this basic architecture can account for the psychological phenomenon of the tilt aftereffect.
The RF-LISSOM model is a descendent of computational models developed by von der Malsburg (1973; see also Amari 1980 and Grossberg 1976). These early models demonstrated that simple computational rules could account for the development of oriented receptive fields from visual input. Since these models, new discoveries about intracortical connectivity have fundamentally altered our understanding of the primary visual cortex (Gilbert and Wiesel, 1983; Gilbert et al., 1990). In addition to the afferent input leading from the eye to cortical areas, there are highly-specific patterns of lateral connectivity that develop in response to visual experience (Burkhalter et al., 1993; Dalva and Katz, 1994; Fisken et al., 1975; Gilbert, 1992; Katz and Callaway, 1992; Löwel and Singer, 1992). RF-LISSOM incorporates this new evidence into a computational model of structure and function.
In previous work with the RF-LISSOM model, Sirosh and Miikkulainen (1994a); Sirosh and Miikkulainen (1996, 1997) showed that Hebbian self-organization in a large recurrent network of simple, laterally-connected neural elements can provide a unified account of self-organization and plasticity in the visual cortex. They demonstrated computationally (1) how receptive fields develop selectivity to orientation, ocular dominance, and size, (2) how such receptive fields organize into global structures of intertwined columnar areas, (3) how the lateral connections develop synergetically with the afferent connections and follow their global organization, and (4) how such structures are maintained in a dynamic equilibrium with the input, resulting in reorganization after retinal and cortical lesions. The model also suggests a functional role for the lateral connections: during development, they learn the activity correlations between cortical neurons, and during visual processing, filter out these correlations from cortical activity to form a redundancy-reduced sparse coding of the visual input.