Related Work - Hierarchical Neural Networks for Image Interpretation - page 61

(a) (b)
Fig. 3.19. Integrating top-down and bottom-up sensory processing by somato-dendritic in-
teractions proposed by Siegel et al. [210]: (a) areas A and B reside at different hierarchical
levels and are reciprocally connected (each area consists of excitatory and inhibitory neurons;
inhibitory neurons project to all neurons within an area; excitatory neurons from both areas
are connected by bottom-up and top-down projections); (b) somato-dendritic interaction and
burst generation (if excitatory input of top-down projections · is strong enough and bottom-
up input

· initiates an action potential that propagates back into the apical dendrite ·, a

dendritic calcium spike is triggered · that in turn causes a burst of action potentials ·)

(images adapted from [210]).

The model reflects recently discovered physiological properties, such as the back-
propagation of action potentials into the apical dendrite and the generation of slow
dendritic calcium spikes that drastically lower the threshold for burst generation.
Siegel et al. [210] propose a functional interpretation for these somato-dendritic
interactions. In the presence of a backpropagating action potential, the subthreshold
top-down input at the apical dendrite can trigger a dendritic calcium spike leading
to a burst of axonal action potentials, as illustrated in Figure 3.19(b). This burst
signal is much more robust to noise than the total number of action potentials. It
indicates a match between bottom-up and top-down stimuli. The authors also found
that priming an interpretation of the bottom-up stimulus by additional input to the
higher area leads to faster and more reliable recognition and biases processing if
multiple stimuli are present.
The model accounts for the asymmetry of bottom-up and top-down pathways
where feed-forward inputs mainly drive the activity of cells, whereas feedback has
rather modulatory effects on the total spike counts. Nevertheless, the integration
of top-down and bottom-up information leads to a robust burst signal. The authors
propose that this bursting pattern could be a basis for the binding of corresponding
high-level and low-level features.

3.2.3 Models with Lateral and Vertical Feedback

Neither lateral interactions nor top-down/bottom-up recurrent interactions alone are
sufficient to explain the performance of the visual system. Because both types are
present in the cortex, models that incorporate horizontal as well as vertical loops are