Networks of Neurons - Computational Explorations in Cognitive Neuroscience - page 73

2

Hidden
(Layers 2,3)

Input
(Layer 4) Output

(Layers 5,6)

1 3

Hidden
(Layers 2,3)

Hidden
(Layers 2,3)

Input
(Layer 4)

Output
(Layers 5,6)

Output
(Layers 5,6)

Sensation
(Thalamus)

Subcortex
BG

Thalamus
Subcortex

Thalamus Subcortex

Motor/BG

Figure 3.5: Larger scale version of cortical circuitry, showing the three different types of cortical areas: 1) is an input area ,which
has a well-developed layer 4 receiving sensory input from the thalamus, but not producing motor output directly. 2) is a hidden area
(often called a higher-level association area), which receives input from input areas (or other hidden areas) and sends outputs to
output areas (or other hidden areas). It has reduced input and output layers, and communicates primarily via layer 2-3 connectivity.
3) is an output area (motor control area), projecting to subcortical brain areas that drive the motor system, which has no real layer
4 (and a larger layer 5). Dashed lines indicate layers and connections that are reduced in importance for a given area, and dotted
lines again represent connections which may exist but are not consistent with the input-hidden-output model of information flow.

areas have reduced but still extant input and output lay-
ers. We will explore some ideas about what function
these might be serving in later chapters.
The picture that emerges from the above view of cor-
tical structure is a very sensible one — information
comes into the network in specialized areas and layers,
is processed by a potentially long sequence of internal
processing areas (the hidden layers), and the results are
then output to drive the motor system. However, we
must add two important wrinkles of complexity to this
otherwise simple picture — the excitatory connections
are almost universally bidirectional within the cortex,
and within each cortical layer there are a number of in-
hibitory neurons whose function has not yet been dis-
cussed. The importance of these features will be made
clear later. Finally, the possible role of the connectiv-
ity with the thalamus and other subcortical structures
in the hidden areas, which is somewhat puzzling under
this simple view, is discussed in later chapters.

3.3 Unidirectional Excitatory Interactions:

Transformations

In this and subsequent sections we explore the basic
types of computations that networks of neurons with
excitatory interactions can perform. We start with the
simpler case of unidirectional or feedforward connec-
tions. Though these are rare in the cortex (see chapter 9
for a relevant exception in the hippocampus ), we will
see that the basic computations generalize to the bidi-
rectionally connected case.
Unidirectional or feedforward excitatory connec-
tivity causes information to flow “forward” in one direc-
tion ( bottom-up ) through the network. We covered one
essential aspect of the computation performed by this
type of connectivity in chapter 2, when we discussed the
role of a neuron as a detector, and explored the process
of detection in section 2.6.3. In this section we explore
the collective effects of a number of detectors operating
in parallel, in terms of how they transform patterns of