We move our eyes to facilitate excellent vision. The motion of objects poses a major threat to excellent vision, because it tends to cause images to slip across the retina, degrading visual acuity. Fortunately, object motion causes visual inputs to a neural circuit, resulting in eye movements that smoothly pursue moving objects. Such smooth pursuit eye movement does a remarkable job of estimating the speed and direction of object motion and creating eye motion that keeps the fovea pointed at moving objects. In my laboratory, we use pursuit eye movements as a model system to understand how neural circuits in the brain use what we see to guide how we move. We also ask how learning within those circuits optimizes motor performance.
From Vision to Action
We are able to use pursuit eye movements to reveal basic mechanisms of brain operation because the basic neural circuits for pursuit are known. Visual inputs arise in the primary visual cortex. They are transmitted through the middle temporal visual area in the extrastriate cortex (area MT) to the parietal and frontal cortex and, in parallel, from these three cortical areas to the cerebellum and brain stem for the assembly of motor commands.
The sensory events that drive pursuit are produced by moving objects and are represented in the electrical activity of populations of neurons in area MT. Each MT neuron is tuned so that it responds most strongly to objects that move in its preferred direction at its preferred speed and less strongly in proportion to how far target speed and direction are from the neuron's preferences.
We know that area MT is the right place to look for the sensory information that drives pursuit, and it appears that the motor system reliably follows the instructions given by MT. Indeed, even the trial-by-trial imprecision in the initiation of pursuit appears to result largely from correlated fluctuations in neural responses within the MT population.
A general challenge for the brain is to estimate the parameters of a sensory stimulus from the brief responses of a large population of neurons. For the visual motion signals that drive pursuit, the best way to estimate (or decode) the speed and direction of target motion is to compute the preferred speed and direction of the neurons that are responding the most strongly. Because accurate pursuit results from only 100 ms of target motion, the decoding computation must be automatic and machine-like: there is no time for "thinking."
To probe for the circuit organization (or organizations) that decodes the sensory representation in MT to estimate target direction and speed, we present the exact same target motion over and over and analyze the trial-by-trial correlations between the firing rate in MT neurons and the estimate of eye speed at the initiation of pursuit. Remarkably, for many MT neurons, the fluctuations of firing rate reasonably predict the eye movement that will ensue 40 ms later. Further, the details of the MT-pursuit correlations contradict the predictions of several popular hypotheses for the neural circuit that decodes the MT population response and support a single neural circuit decoder.
Many sensory and motor systems face the same decoding problem, and we imagine that they use a similar solution. Information about the parameters of a sensory event, or about the next desired motor act, resides in the responses across populations of neurons. The brain must solve the same problem we have studied for each sensory or motor modality: it must estimate the features of the sensory stimulus or the desired movement from a population response.
Because the neural circuit for smooth pursuit includes many areas of the cerebral cortex as well as the cerebellum, pursuit provides an excellent system for understanding how we learn motor skills. Our research over the past 15 years has demonstrated that the learned responses of cerebellar Purkinje cells play a causal role in motor learning in pursuit. Recently, we have been able to go still further and demonstrate that short-term changes in synaptic strength (plasticity) at a specific site in the cerebellum probably cause short-term behavioral motor learning.
Purkinje cells receive two kinds of inputs that are very different in their anatomy and physiology. We found that the occurrence of the rare "climbing fiber input" during one movement is linked to a depression of the firing rate of the conventional and frequent "simple spikes" during the next movement. This single-trial depression has a very short lifetime; both it and the companion single-trial learning in pursuit eye movements are present when the interval between learning trials is 2.5 s, but they vanish when the interval is extended to 6 s. Preliminary data support our hypothesis that single-trial learning is a prerequisite for some but not all of the long-term learning that develops over the course of 100 repetitions of the learning conditions. Our analysis provides strong support for the popular theory that climbing fiber inputs instruct learning in Purkinje cells and provides new evidence that this mechanism operates during learning in awake animals, and on a very brief timescale.
General organizing principles of motor learning have emerged from our studies on learning in pursuit, along with our earlier demonstration of the neural circuit mechanisms of learning in the vestibulo-ocular reflex. Behavioral learning appears, at least superficially, to work as a single, unified operation. We now understand, however, that behavioral learning is mediated by neural changes that occur at several sites in the brain, including changes in the cerebellum and the cerebral cortex. We also see that learning engages multiple mechanisms of modification of synaptic strength that operate in multiple sites over a range of times—from seconds to days. The same principles seem likely to operate in other forms of learning that occur in all brain systems.
This work was supported in part by the National Eye Institute and the National Institute of Mental Health.
As of March 22, 2013