Bernal S D_2010.pdf - University of Plymouth

2.1. OBJECT RECOGNJTION
properties arise from retinal ganglion cells and LGN cells, which also exhibit some selectivity
to contrast, velocity, colour and spatial frequency, V1 neurons with similar tuning properties
lend to group together, leading t(i Ihe classical columnar organization of ocular dominance and
orientation preference in cortex (Hubel and Wiesel 1965).
Hubel and Wiesel (1965) were the first to propose the hierarchical organization of receptive
tields, such thai VI simple cells are built from converging LGN cells aligned in space to pro­
duce the elongated on-off subregions observed. Additionally, the model provided the first clas­
sification of VI ceils, dividing them into simple and complex. Cells fell into the simple category
if their receptive fields could be separated into on and off subregions, which could be linearly
summaled to predict the cell's response to different artificial stimuli. The re,st of the cells, which
did not have separate subregions, were categorized by exclusion as complex cells. However, the
majority of VI cells fall into die complex category. As will be described further down, there are
also numerous variants within the simple and complex categories.
Several extensions improved the initial Hubel and Wiesel receptive held model of VI neurons.
l-irstly, the linear filler was expanded to include a temporal dimension. Spatiolemporal receptive
fields not only lake into account the spatial profile, but also the temporal course of the response,
and have proved to be crucial in understanding direction selectivity. This first filtering stage
was .shown to be well approximated by 2-dimensional Gabor filters (Jones and Palmer 1987).
Secondly, a nonlinear stage was added, which described how the linear filter outputs were trans­
formed into an instantaneous firing rate via a nonlinear Poisson process. The two-stage model
was therefore called the linear-nonlinear (l.N) model and provided a much better prediction of
neuron responses than strictly linear lilters, specially for retina and thalamic cells (Carandini
elal.2(M)5).
Nonetheless, the model still had significant limitations (Ringach 2004). It was unable to account
for the dependence on contrast of several response properties, such as saturation and summation
size. For example, the greater the contrast of the stimulus, the smaller the degree of spatial
summation, and thus the receptive field size. Furthermore, the LN model could not explain
surround suppression, such as stimuli at an orthogonal orientation inhibiting the cells' response.
10