3.5 Biological basis of the model

  RF-LISSOM models the interactions of small groups of neurons and the connections between them rather than individual neurons and synapses. This is appropriate because detailed low-level knowledge of the cortex is still very patchy, so its function cannot simply be extrapolated from the fragments of anatomy and physiology which have been established. Instead, a more promising approach at present is to work backward from observed psychophysical and other aggregate phenomena to describe the basic computations that are being performed. When unequivocal low-level physiological or anatomical data is available, it is used to constrain the possible models.

Many of the fundamental assumptions of the model, such as the computation of the input activity as a weighted sum and the sigmoidal activation function, are common to most neural network models. Their biological validity has been examined in detail previously by other researchers. Representing activity as a scalar value rather than a spike train is a common abstraction made for computational convenience; RF-LISSOM can be extended to model the low-level time-dependent behavior of neurons when studying phenomena that depend upon it (Choe and Miikkulainen, 1996; Miikkulainen et al., 1997).

However, there are two key aspects of the model that remain particularly controversial: whether the long-range horizontal connections are primarily inhibitory in effect, and the balance between genetic factors (i.e., hardwiring) and environmental factors (i.e., due to visual input) in the organization of the cortex. These topics will be discussed in detail in the remaining sections to support the claim that RF-LISSOM represents a biologically realistic model of cortical phenomena, including the tilt aftereffect. Furthermore, recent experimental data has clarified many of the issues involved since the discussion by Sirosh (1995), and thus it is worthwhile to reexamine these issues.