Biomedical Engineering Reference
In-Depth Information
of biological realism as computers have become more capable and the interests
of researchers have turned toward biological understanding, rather than just
simulation, of brain functions (41).
A standard for what was to come was established by Traub and Miles with
their model of the CA3 network of the hippocampus (42), which featured stan-
dardized multicompartment pyramidal and GABA
A
- and GABA
B
-type inhibitory
cells. Properties of several types of active channels (e.g., Ca
2+
-dependent K
+
channels) were combined into composite channels to keep the amount of com-
putation to a practical level. Synaptic connections of four basic types took into
account axonal delays, a composite activation process, and a first-order relaxa-
tion process. This model was very successful in helping to elucidate the roles of
the various channel types in complex behaviors such as bursting and network
synchronization.
More recently, and perhaps of more relevance to the main subject matter of
this topic, Tagamets and Horwitz (43) constructed a large-scale model of the
visual and forebrain circuitry thought to be involved in delayed match-to-sample
tasks in humans. Their goal was to explicate the neural basis for the rCBF sig-
nals recorded in PET (positron emission tomography) studies of this and related
tasks. Their hypothesis was that the rCBF signal is simply proportional to the
sum of the absolute values (whether excitatory or inhibitory) of all the inputs to
all the neuronal units in a region of interest. Because this hypothesis did not de-
pend on the complex details of neuronal activation as in the Traub and Miles
study, simplified canonical units representing local assemblies of neurons were
modeled, representing a total of four areas along the occipitotemporal pathway
from lateral geniculate nucleus to prefrontal cortex. Good agreement with ex-
perimental rCBF data was obtained, suggesting that synaptic activity indeed
accounts for at least a major portion of differential neural metabolic demand,
and thus of blood flow, although some influence of glial activity should not be
discounted. Arbib et al. (44), working from a similar hypothesis, showed how to
calculate simulated functional MRI images from network models, using neural
systems for imitative behavior as their exemplar. See also the study of temporal
patterns of spontaneous activity in the developing spinal cord in the next chapter
of this volume (by Tabak and Rinzel).
1.3. Learning
An essential element of network models is the incorporation of some
scheme for adaptive change or "learning." Parallel distributed processing models
only really took off with the popularization of the "back-propagation" learning
rule (19,45), a form of gradient-descent optimization based on adjusting the
strengths of connections between neurons to reduce their contribution to the
error measured at the output, as determined by the calculus chain rule for differ-
entiation. However, implementation of this rule requires that nonlocal informa-
tion be available at each synapse, and this is widely considered to be unavailable
