Conclusions - Computational Explorations in Cognitive Neuroscience - page 408

We will take up the issue of phases of learning and co-
ordination of learning more generally in a later section.

Serotonin is another important neurotransmitter that has
been linked to the regulation of states of arousal (i.e.,
sleeping versus waking), and also to mood.
It is possible to ignore these neuromodulators to the
extent that we assume an awake cortex that is com-
pletely focused on the task at hand. However, a richer
and more complete model of cognition will require the
inclusion of such neuromodulatory factors, which play
a critical role in human and animal performance both in
the “wild” and in psychological task situations.
Aside from the basal ganglia and thalamus, proba-
bly the most cognitively relevant brain area, which has
received scant attention in this topic, is the cerebellum.
Long known to be important for motor control, the cere-
bellum has been recently implicated as also playing a
more cognitive role (e.g., Gao et al., 1996; Doyon, Gau-
dreau, & Bouchard, 1997). One useful focus of future
research would be to understand the exact contributions
of the cerebellum at a mechanistic level. Another area
that is likely to be important is the amygdala, which is
important for assigning emotional salience to stimuli,
and has been the subject of some recent modeling ef-
forts (Armony et al., 1997).

Missing Brain Areas

The models in this text have focused on the cortex (in-
cluding the hippocampus), with the exception of a few
that dealt in a very superficial way with the thalamus
and the basal ganglia. There are a number of good
reasons for this focus on cortical networks, including:
(1) Much of cognition depends most critically on the
cortex; (2) a cortical model can usually be initialized
rather simply, because learning apparently plays a very
strong role in structuring the cortical network; (3) the
cortical pyramidal neuron, with supporting inhibitory
interneurons, provides a common structural basis for
neural computation throughout the cortex, so a common
algorithm can be used for modeling many different cog-
nitive phenomena.
In reality of course, the cortex operates in the context
of a large number of other brain areas. There are sev-
eral difficulties in including these other areas in compu-
tational models, including the fact that they are gener-
ally evolutionarily older and correspondingly more de-
termined by genetics. This usually means that they have
more complex, specialized neuron types that cannot be
easily parameterized in terms of just patterns of weight
values in an otherwise generic unit model. At a more
practical level, adding other brain areas to a model can
substantially increase computational complexity of the
model.
Many of the missing brain areas serve as extensions
of the sensory input or motor output pathways, and are
often not particularly relevant for cognitive-level mod-
els. However, others have more profound effects on
the nature of processing in the cortex because they se-
crete neuromodulatory substances. We explored the
role of the neuromodulator dopamine in the context of
reinforcement learning (chapter 6) and prefrontal active
maintenance and learning (chapters 9 and 11). A num-
ber of other neuromodulators probably have similarly
important roles. For example, norepinephrine, secreted
by the locus ceruleus (LC), may regulate the ability of
cortical neurons to remain focused on a particular task
as a function of task performance or other variables.

Scaling

We have argued that the brain possesses a kind of self-
similar, fractal structure, so that coarse-grained models
of multiple brain areas using relatively few units should
employ the same principles as more fine-grained mod-
els of individual columns using a large number of units.
Although a number of factors support this fractal as-
sumption, the multiple models approach is an ideal one
to test the validity of these scaling assumptions. That
is, one could literally test how well a very detailed, fine-
grained model can be approximated by a coarse-grained
one. Clearly, the coarse-grained model can only repre-
sent a small fraction of the information compared to the
fine-grained one, but its overall dynamics and some rel-
evant behavioral characteristics should be comparable.
Ultimately, the limitation is one of computational
power — it is nearly impossible to use fine-grained
models of multiple brain areas on today's computers.
However, the computational horizons are rapidly ex-
panding (for example, computer power more than dou-