Summary and Conclusions - Hierarchical Neural Networks for Image Interpretation - page 217

have dynamic synapses. Spiking neurons could be used to implement fast tempo-
ral dynamics, as illustrated in the activity-driven update example. They would also
allow for codes which are based on the precise timing of spikes. The use of synchro-
nization for feature binding and segmentation could be explored. Dynamic synapses
could be employed to decode temporal codes. Furthermore, they could be used to
change the strength of cooperation/competition between feature cells dynamically.
Another possible line of future research is to give the network activities a prob-
abilistic interpretation. One could view the abstraction pyramid as a graphical be-
lief network and apply belief propagation. This proposal is motivated by the re-
cent success of the belief propagation algorithm in cyclic graphs [76]. Unlike belief
propagation in acyclic graphs, the algorithm only approximates inference and is not
guaranteed to converge. Generalized belief propagation [247] has been proposed to
implement better approximations with a moderate increase in computational costs.
11.3.3 Integration into Complete Systems
The goal of visual processing, in many contexts, is to ultimately control the behavior
of a system based on the sensed state of the environment. This calls for an integrated
treatment of perception and action. Since not only object identity, but object loca-
tion is needed for action, the perception network would need to not only model the
ventral visual pathway, but the dorsal one as well. Furthermore, an inverse hierarchi-
cal network could be used to expand abstract action decisions into low-level action
commands. Such an integrated system could be employed to implement active vi-
sion. It would also allow for the use of reinforcement learning techniques.