Page 56 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

7. Homologous Laminar Cortical Circuits for All Biological Intelligence 43

Fig. 2.7 also shows that bottom-up signals from the LGN use the same modulatory
on-center, off-surround network to activate layer 4 in V1 that is used by the top-down
attentional feedback pathway. In addition, there is a direct bottom-up excitatory
pathway from LGN to layer 4 so that the LGN can activate V1 in response to inputs
from the retina. Taken together, the direct LGN-to-4 pathway and the LGN-to-6-to-4
modulatory on-center, off-surround network ensure that bottom-up inputs from the
LGN to V1 are contrast-normalized at layer 4 cells.
The sharing of the layer 6-to-4 modulatory on-center, off-surround network by
bottom-up and top-down pathways converts this network into a decision interface
where preattentive automatic bottom-up processing and attentive task-selective
top-down processing can cooperate and compete to choose the combination of
signals that is most salient at any given moment.
Such a cooperative-competitive decision interface exists in every granular
neocortical area. As a result, a top-down task-selective priming signal from a higher
cortical area can propagate through multiple lower cortical areas via their layer 6,
which can then activate their layer 6-to-4 modulatory on-center, off-surround
networks. In this way, an entire cortical hierarchy may get ready to process incoming
bottom-up signals to accommodate the bias imposed by the prime.
Fig. 2.7 also shows that layer 2/3 in each cortical area also projects back to layer 6,
and then up to layer 4 via the folded feedback network. The horizontal connections in
layer 2/3 carry out a variety of functions in different cortical areas. In V2, they carry
out perceptual grouping and boundary completion [29], a process whose so-called
bipole grouping properties were predicted before the neurophysiological data of
von der Heydt et al. were reported [30,31] and which were subsequently extensively
modeled by LAMINART (e.g., Refs. [27,28,32,33]). In cognitive processing regions,
such as the ventrolateral prefrontal cortex, it has been suggested that such horizontal
connections enable learning of categories, also called list chunks,that respond
selectively to sequences of items that are stored in working memory [34,35].
The development of these horizontal connections begins before birth and
continues in response to the statistics of visual environments after birth. The fact
that the layer 2/3-to-6-to-4-to-2/3 pathway satisfies the ART matching rule enables
this development, as well as that of other cortical circuits, to dynamically

=
circular reaction [17e19], is learned from spatial-to-motor and motor-to-spatial
representations at the two adaptive pathways in the model, which are denoted by
hemispherical synapses. In particular, the spatial direction vector (DV s ) is adaptively
mapped into the motor direction vector (DV m ), thereby carrying out the transformation from
visual direction into joint rotation that gives the DIRECT model its name.
(Left panel) Adapted with permission from D. Bullock, S. Grossberg, Neural dynamics of planned arm move-
ments: emergent invariants and speed-accuracy properties during trajectory formation, Psychological Review 95
(1988) 49e90. (Right panel) Reprinted with permission from D. Bullock, S. Grossberg, F.H. Guenther, A self-
organizing neural model of motor equivalent reaching and tool use by a multijoint arm, Journal of Cognitive
Neuroscience 5 (1993) 408e435.