 
| |
        |
Department of Computer Science ResearchComputational Neuroscience
Biological motor control demonstrates an amazing ability to deal with
complex and ever-changing bodies and environments. Motor control
systems are ideal for us to study for several reasons. The problem of
manupulating a complex system such as a multi-link arm to solve a task
is very useful in testing artificial intelligence algorithms and
applications. Since reinforcement learning has biological parallels,
we can compare how an RL learning agent controls a limb to how humans
and animals control their limbs. This comparison is much more direct
than emotional or motivational aspects of biological systems. In
addition, animals control their limbs in a stereotyped manner,
indicating that the central nervous system of a certain animal has
evolved to control motor functions in a specific way. Using
computational neuroscience methods such as neural networks and
physiologically based models, we investigate how the brain learns and
controls movements such as reaching and grasping. We also collaborate
with neuroscientist specialists in motor control from UMass and other
universities. Our computational neuroscience research has
connections with not only artificial intelligence, but also the
developmental psychology and robotics research we are involved
with. What follows are brief descriptions of some of the projects
to which members of our lab have contributed. Specific papers,
proceedings, and posters are referenced in the ALL publications page.
|
We use neural network models to theorize possible mechanisms the brain
use for selecting muscles and how the primary cortex (MI) encodes
movement. Specifically, we use a gradient descent algorithm to train a
neural network to select muscles based on their ability to produce the
desired movement while minimizing the total effort required by all
muscles to implement the movement. We also show that MI neurons that
encode in extrinsic space (such as direction of movement) can directly
activate muscles appropriately. This is in contrast to other theories
in which only MI neurons that encode movement in muscle space can
directly activate muscles. The illustration on the left graphs
muscle activity as a function of movement direction as produced by the
models (thick dark line) and as recorded from monkey subjects (thin
light lines). The agreement between model and monkey data suggest that
these theories are reasonable
The stretch reflex can be described by non-linear equations of length
and velocity. In the fractional power damping model of the elbow
musculature, antagonist muscle bursts that function to decelerate the
ongoing movement are generated by the stretch reflex, with central
commands preserving some influence over the precise timing and
magnitude of this response. Although descending control is achieved by
setting equilibrium points of the viscoelastic muscle-reflex system,
the nonlinear damping behavior of opposing reflexes interacts to
produce a region of stiction around the specified equilibrium. This
slows movement so dramatically that the equilibrium point is
effectively never reached in most movements. The complexities of these
nonlinearities do not necessarily increase the difficulty of the
central control problem. Effective control can be achieved by
precisely timing the offset of the agonist burst, and the tendency for
oscillations around the endpoint is greatly diminished by the stiction
that results from fractional power damping.
Rapid human arm movements often have velocity profiles consisting of
several bell-shaped acceleration-deceleration phases, sometimes
overlapping in time and sometimes appearing separately. We show how
such sub-movement sequences can emerge naturally as a result of an
optimal control policy learned by a reinforcement learning system in
the face of uncertainty and feedback delay. The system learns to
generate sequences of pulse-step commands, producing fast initial
sub-movements followed by several slow corrective sub-movements that
often begin before the initial sub-movement has completed. These
results suggest how the nervous system might efficiently control a
stochastic motor plant under uncertainty and feedback delay.
A simplified model of the cerebellum was developed to explore its
potential for adaptive, predictive control based on delayed feedback
information. An abstract representation of a single Purkinje cell with
multistable properties was interfaced, via a formalized premotor
network, with a simulated single degree-of-freedom limb. The limb
actuator was a nonlinear spring-mass system based on the nonlinear
velocity dependence of the stretch reflex. By including realistic
mossy fiber signals, as well as realistic conduction delays in
afferent and efferent pathways, the model allowed the investigation of
timing and predictive processes relevant to cerebellar involvement in
the control of movement. The model regulates movement by learning to
react in an anticipatory fashion to sensory feedback. Learning depends
on training information generated from corrective movements and uses a
temporally-asymmetric form of plasticity for the parallel fiber
synapses on Purkinje cells