Neurobiological Background - Hierarchical Neural Networks for Image Interpretation - page 33

Fig. 2.15. Illustration of the model of Lee et al . for the role of V1. Image segmentation, fig-
ure/ground, shape computation and object recognition in this framework occur concurrently
and interactively in a constant feed-forward and feedback loop that involves the entire hier-
archical circuit in the visual system. Signals of higher level visual representations, such as a
2.5D surface sketch, 3D model or view-based object memory, are likely reflected in the later
part of V1's activities. (Adapted from [139]).
tial precision would necessarily involve V1 and be reflected in the later part of its
neurons activities. This is illustrated in Figure 2.15.
This model is supported by the report of Doninger et al. [54], who found that
electric potentials reflecting closure have a latency of 290ms when incomplete pat-
terns must be recognized. Since higher ventral areas are activated much earlier, this
initial activity does not produce a coherent percept of the incomplete object. They
suggest that the objects must be first completed by feed-forward/feedback interac-
tions with lower visual areas before they can be recognized. Modulated activity in
lower areas may reflect these interactions.
Furthermore, a recent report by Pascual-Leone and Walsh [174] using transcra-
nial magnetic stimulation (TMS) suggests that activation of feedback connections
to the lowest stages of the hierarchy might be essential for conscious vision. They
stimulated area V5/MT and area V1 asynchronously and investigated how the in-
teraction of both stimuli affected perceived phosphenes (moving flashes of light).
They found that TMS over V1 with a latency of 5 to 45ms after TMS over V5 dis-
rupted the perception of the phosphene, while neither earlier nor later V1 stimuli
nor a conditioning V5 stimulus did affect the percept.
Based on these findings, Bullier [37] proposed that areas V1 and V2, instead of
simply transmitting information, might act as 'active blackboards' that integrate the
results of computations performed in higher order areas, at least for the early stages
of processing. This would be an efficient way to solve the problem of computations
that involve interactions between features which are not present in neighboring neu-
rons in any one cortical area.