Neurobiological Background - Hierarchical Neural Networks for Image Interpretation - page 32

Fig. 2.14. Binding by shrinkage of receptive fields to attended stimuli, as proposed by
Reynolds and Desimone [190]. The orientation selective neuron responds to a vertical bar
anywhere in its classical receptive field while the color sensitive neuron responds to any dark
bar, regardless of its orientation. Thus, when attention is drawn away, the response of the
neurons to the two objects is ambiguous since both are active. In contrast, when attention is
directed to one of the two stimuli, both neurons respond as if only the attended object were
present.
Because of this connectivity rule, neurons at low levels of the hierarchy have small
and relatively simple receptive fields, whereas neurons at the highest stages have
large and very specialized receptive fields. Activity of neurons at the highest stages
of the hierarchy is important for conscious vision, as suggested by the results of
imaging studies in humans and recordings in monkeys with bistable visual stimuli.
Although this model explains a large number of observations in visual perception,
it fails to account for the very dense network of feedback connections that connect
cortical areas in the reverse direction.
Super et al. [222] investigated what goes wrong when salient stimuli sometimes
go undetected. They showed that figure/ground contextual modulation in V1 [130]
is influenced strongly by whether stimuli are either 'seen' or 'not seen'. The fig-
ure/ground contextual modulation not only makes V1 neurons respond better, but
this enhancement is spatially uniform within the figure. Both 'detected' and 'non
detected' stimuli evoke similar early neuronal activity. In both cases, the visual in-
put reaches V1 and produces a clear neural response. Only the contextual modula-
tion reflects in a qualitative manner whether the stimulus has been processed up to
the level of 'detection'. They conclude that this perceptual level is situated between
purely sensory and decision or motor stages of processing.
In line with these findings, Lee et al. [139] present neurophysiological data,
which shows that the later part of V1 neuron responses reflects higher order percep-
tual computations, such as figure/ground segmentation. They propose that, because
of V1 neurons precise encoding of orientation and spatial information, higher level
perceptual computations that require high resolution details, fine geometry and spa-