TABLE OF CONTENTS Selectivity is a feature displayed by many cortical cells. In visual cortex (V1) this manifests itself most strikingly as selectivity to preferred orientations. There is substantial evidence that the normal development of this orientation selectivity requires rearing in a patterned visual environment[Wiesel and Hubel, 1965; Blakemore, 1976]. In previous chapters we used mostly low dimensional input environments in our explorations. In this chapter we employ more realistic visual environments in order to investigate how and why such selectivity develops, comparing BCM and several related statistically derived learning algorithms.
In what follows we present simulations of single cells in realistic environments composed of natural scene images to explore the normal development of orientation and direction selectivity. In subsequent chapters we introduce environments to represent the input activity to the cortex in deprived scenarios, such as monocular and binocular deprivation.
In this section we investigate a simple single cell model for developing monocular, orientation selective cells. Inputs d are chosen from preprocessed natural images. The output of the cell c, is a sigmoid of the weighted sum of inputs c = ?( ? i m id i). By convention the output activity c is measured with respect to the level of spontaneous activity, such that c < 0 represents firing below spontaneous. Typically we choose a non-symmetric sigmoidal, this asymmetry is a reasonable assumption for cortical neurons with a low level of spontaneous activity, which can have an output rate...
