5.3: Orientation Selectivity
5.3 Orientation Selectivity
Receptive fields in the visual cortex are usually orientation selective. But they receive inputs from LGN cells that are not orientation selective. Can synaptic plasticity account for this transformation?
In this section we review different cases in which the principal components of the input environment can become orientation selective.
Two distinct types of models will be described. In the first part of this section models in which artificial assumptions about the correlation function of the inputs are made in order to produce orientation selectivity, in the second we describe the evolution of orientation selectivity under realistic natural image environments.
We first describe a simple soluble 1D model in which true orientation selectivity can not be obtained; however symmetry breaking analogous to orientation selectivity does appear. This model can give us intuition as to what assumptions about the input environment must be made in order to obtain symmetry breaking. We next describe more elaborate 2D models which under some conditions produce orientation selective PC's; we discuss their symmetry properties as well. We then treat PCA models trained in a realistic natural image environment, display the results of analysis, and show under what conditions orientation selectivity is attained.
5.3.1 An exactly soluble 1D Model
This example given a simple demonstration of how symmetry in the correlation function can be broken when principal components do not have the same symmetry as the correlation function. This example is one dimensional (1D) in the sense that x denotes points on a...