Connectionist and other computational models of vision frequently
assume, for simplicity or for other reasons, that the flow of
information through the visual cortex is unidirectional, and that its
architecture is purely feedforward. Although this self-imposed
constraint is understandable in the light of the PDP group's
[33] original concentration on feedforward networks, it is
unwarranted both from the computational and from the biological
standpoint. Anatomical and physiological evidence suggests that
lateral connections constitute an important part of the local cortical
circuitry, and, moreover, that these connections possess a clearly
nonrandom (albeit not yet perfectly understood) structure
[1,26]. The present paper argues that
this ubiquity of lateral connections is to be expected, given the
computational advantages such connections confer onto a feedforward
layered network architecture.
This argument is
supported by a number of successful models of various cortical
functions, all of which involve lateral connections. The functions
addressed by the models range from low-level (the formation of the
receptive fields in the primary visual cortical area V1
(the Shaping the RFs section); combining receptive fields to enhance their
utility for recognition (the Pairs of RFs section)) to high-level
(recognition of 3D objects (the Emergence of ... section); shape
representation (the Lateral Comparisons... section)). The Discussion section
contains a brief discussion of the functional value of lateral
connections, considered from the standpoint of philosophy of
representation.