 |
 |
 |
 |
 |
 |
 |
 |
Research on Computational models of movement
-Using Information Theory to Quantify Movement Deficits
- Persistence of Excitation and Error-Dependent Noise
-Separate Modules for Skill Learning and Environment Adaptation
-Interpretation of Cortical Population Codes
-Large-scale Neural Simulation
Using Information Theory to Quantify Movement Deficits
The translation of sensation into action can be understood as the transmission of information from sensory to motor areas of the central nervous system. Therefore injury to either sensory or motor systems will reduce the transmission of information, and this will be reflected in a decrease in the speed or controllability of responses to stimuli. Although information theory has been applied to human movement for more than 50 years, our work is the first to apply it to children with movement disorders and the first to measure multiple contributors to poor information transmission simultaneously. We believe that the use of Information Theory is particularly relevant since reduced information transmission may directly reflect loss of neurons, and therefore it may provide a measure that is specifically relevant to injury to sensorimotor systems in children. We have demonstrated decreased information transmission; we have demonstrated that this is due to both decreased speed and to increased “noise” in the system. This result is related to the increased variability in movement for children with hyperkinetic dystonia, since variability introduces a form of unpredictable noise. We have shown that quantification of information rate in individual children can be used to design individualized computer interfaces that can double the speed of interaction with a computer or communication device.
Persistence of Excitation and Error-Dependent Noise
Ongoing theoretical research examines the effects of limitations in computation on learning and performance in children with dystonia. An important recent result is that learning can be augmented by adding a specific type of random noise to the system, where the magnitude of the noise depends on the movement error. This provides an exciting new avenue for therapy. These results are based upon extensions of the “Feedback-Error Learning” (FEL) model of motor learning, and we have proven convergence criteria as well as failure criteria for this model. Persistence of Excitation describes tasks that are sufficiently rich that they can be used to train many variables in a motor system, and we use this concept to study the training tasks that may be helpful to children with dystonia. It may be important to select tasks that are persistently exciting in order to permit improvement in overall motor performance rather than simply learning an individual task.
Separate Modules for Skill Learning and Environment Adaptation
Humans can learn many different skills, and they can perform each skill in multiple different environments without needing to relearn each one. For instance, it is possible to walk under water, or to ride instantly upon a new bicycle. This can be modeled by a combination of two networks: one that learns the sequence of movements needed to accomplish a skill, and another that adapts to changes in the environment. Ongoing work seeks to determine the effect of injury on these two systems. For instance, if a child has a brain injury that impairs the ability to adapt to the environment, the lack of adaptation might also impair the ability to learn new skills. Conversely, if a child has a brain injury that reduces the ability to learn new complex skills, he or she may not adapt to the environment because there may not be enough accurate movement to provide adequate training examples for the adaptation mechanism. Understanding these modules may provide new options for therapy, since under the right circumstances, modification of the environment might be used to train new skills.
Interpretation of Cortical Population Codes
Models of computational learning almost always involve the use of real numbers that can be well-represented by computers but which do not relate directly to the signals that we can measure during an experiment in electrophysiology. I have developed algorithms for the analysis of recordings from multiple cortical cells, as well as algorithms for analyzing what it is that the cells represent. The algorithms are based on a Bayesian analysis of the spike firing pattern, and they make use of approximate integration methods for stochastic differential equations.
Large-scale Neural Simulation
We have started a new initiative in large-scale neural simulation. The goal is to use a large network of computers (a supercomputer cluster) to simulate activity and learning in the hundreds of millions of neurons responsible for movement. The purpose of the simulation will be to model both normal and abnormal motor control, and to study the effect of development and learning on motor disorders. The simulation will allow developmental processes that normally occur over years to be examined over days in order to understand the evolution of motor disorders. Most importantly, the simulation will allow us to predict the long-term effect of injury and the long-term effects of training, medical, or surgical interventions. We plan to test the predictions of the simulation through clinical trials of children at different ages.