Biomedical Engineering Reference
In-Depth Information
As mentioned above, a typical paradigm for analyzing specific aspects of
motor learning is the method of classical conditioning of the eyeblink or with-
drawal reflexes. For determining the corresponding area within the cerebellum, a
transient inactivation of the cerebellar nuclei was performed in well-trained
animals by injection of the GABA
A
agonist muscimol (e.g., in the rabbit (11)
and in the cat (40)). The feedback error cerebellar learning model of Kawato
(35) simulates such a situation as follows: the signal from the cerebral cortex is
conveyed to a summing point projecting to spinal motor centers and to the cere-
bellum including an inverse model, the output of which represents the feedfor-
ward motor command in motor coordinates. In a "well-trained" situation, the
resulting trajectory is equal to the desired trajectory such that the output of the
feedback controller from the spinal motor centers representing the error signal is
typically zero. After injection with the "cerebellar nuclei blocked," the output of
the cerebellum is zero and the motor command generated by the cerebral cortex
is the only signal at the summing point. During recovery of the cerebellum, the
contribution of the motor cortex progressively attenuates. Although this simula-
tion by Kawato (35) does not include cerebellar pathology, such as ataxia, it
shows automatic substitution of the cerebral cortex when cerebellar functions
are disabled.
An update of these models is summarized by Wolpert et al. (79). In their
cerebellar-feedback-error-learning model (an inverse internal model), the trajec-
tory error trains the internal modal such that the actual trajectory (motor output)
finally will be fairly close to the desired trajectory. Further, the cerebellum also
generates a forward representation of the motor apparatus, known as a forward
model. This allows for simplification: the inputs of such a forward model are the
current state of, e.g., the arm (processed reafference signal) and the efference
copy of the motor command producing an estimate of the new state of the arm.
Such a predictor system has to take into account corresponding "transport de-
lays," which may be long with respect to movement duration. For the motor
learning aspect, a computational model is provided that includes multiple
paired
forward and inverse modules. The inverse model generates the motor command
representing the desired trajectory, whereas a single forward model predicts the
consequence of a performed action and can thus be used to estimate responsibil-
ity. This, however, can be achieved earlier when a movement has been initiated
within corresponding sensory contextual cues and the result of the action is
known (79).
A general model of the cerebellum presented by Arbib et al. (5) can simu-
late results from prism adaptation in control subjects and patients during dart-
throwing (68). The model includes a movement pattern generator, the concept of
modifiable synapses (29), and, again, an assumption that the inferior olive con-
veys the error signal via climbing fibers. The simulation results coincide well
with experimental results in humans (5). The same group presented a model in
