Page 54 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

7. Homologous Laminar Cortical Circuits for All Biological Intelligence 41

top-down processing [13,25]. For example, in response to an unambiguous scene, a
fast feedforward sweep can occur through the entire visual hierarchy, leading to
rapid recognition, as reported by Ref. [26]. Such a feedforward sweep can occur,
for example, in the LAMINART architecture [27,28] that is shown in Fig. 2.7 by
leaping from retina to the lateral geniculate nucleus (LGN), then through layers 6,
4, and 2/3 in cortical area V1 to layers 6, 4, and 2/3 in cortical area V2, and beyond.
If, however, a scene contains ambiguous information, for example, in the form of
multiple possible groupings of the same sets of features in a complex textured scene,
then the network can automatically use its feedback loops to make the best decision
in the face of this uncertainty. In particular, competition among these groupings can
occur due to inhibitory interneurons in layers 4 and 2/3 (black cells and synapses in
Fig. 2.7). This competition can cause all cell activities to become smaller because
the competitive circuits in the model are self-normalizing; that is, they tend to
conserve the total activity of the circuit. This self-normalizing property arises
from the ability of the shunting on-center off-surround networks that realize the
competitive circuits to process input contrasts over a large dynamic range without
saturation, and thereby solve what I have called the noise-saturation dilemma
[45,46].
Self-normalizing competition among alternative cortical interpretations of the
data may hereby reduce the activation amplitude and coherence of each grouping
alternative, thereby slowing down its processing. This slowing down of processing
rate occurs as interlaminar, but intracortical, feedback between layers 2/3-to-6-to-4-
to-2/3 (Fig. 2.7), among other feedback pathways, contrast-enhances and amplifies
the grouping that is supported by the most evidence. The amplification of the win-
ning grouping’s activity automatically speeds up its ability to send output signals to
the next cortical region.
This example illustrates an important sense in which the cortex “runs as fast as it
can” in response to the degree of uncertainty in the data, automatically switching
from fast feedforward processing in response to unambiguous data to slower feed-
back processing to resolve uncertainties in the data to the degree that the data allow.
Our brains hereby go beyond current Bayesian models to implement a kind of real-
time probability theory and hypothesis testing that trade uncertainty against speed to
make the best decisions in response to probabilistic environments whose rules can
change rapidly through time.
Fig. 2.7 also depicts how the ART matching rule circuit in Fig. 2.4 is realized
within the laminar circuits of neocortex. For example, the top-down pathway
from layer 6 in V2 projects to layer 6 in V1, which sends bottom-up signals to layer
4. These bottom-up signals are sent via a modulatory on-center (note the balanced
excitatory and inhibitory pathways to layer 4) surrounded by a driving off-
surround network. The top-down signals from V2 are hereby “folded” at layer 6
in V1 in order to reach layer 4. I have accordingly called this property folded
feedback.
Because the ART matching rule is realized within laminar neocortical circuits, they
cansolvethestability-plasticitydilemmaandsupportrapidlearningandstablememory.