Chapter 4 is the companion paper to the Global VOR Model in chapter 3. This model was presented at the Neural Control of Movement conference in April 1996 and published in the Proceedings of the 3rd Joint Symposium on Neural Computation in June 1996, organized by Institute of Neural Computation, University of California, San Diego, and the California Institute of Technology. The chapter introduces a model of the cerebellum to explain the predictive modulation of the VOR gain with vergence eye movements. During vergence eye movements, such as movements elicited when looking from a far to a close target which typically last about 600 ms in rhesus monkeys, the VOR gain is not modulated with current vergence eye position, but rather is modulated in anticipation of the final vergence position in order to optimize stabilization of the future target on the fovea ([Snyder, Lawrence and King1992]). The cerebellum in this model constructs a predictive representation of vergence angle which modulates the vergence input to the Global VOR Model of the previous chapter. The cerebellum constructs vergence angle predictions by learning to combine the dynamical responses during vergence eye movements of vergence-disparity neurons of primary cortical visual area V1 ([Trotter1995]; [Trotter et al.1992]).
In this chapter, it is argued that the cerebellum is a predictive machine with ubiquitous participation in brain functions. In the model, cerebellar predictions of neural activity are constructed from previous neural activities occurring consistently prior to the neural activity to predict. For example, in Pavlovian conditioning, if a bell rings consistently prior to the presentation of food, the bell alone will induce salivation after a training period. In this example, the cerebellum would use the neural activity representing the ringing of the bell to construct a prediction of the neural activity related to salivation. After the training period, the brain may use this prediction to initiate salivation in anticipation of the food.
As a predictive machine, the cerebellum learns to anticipate the temporal sequence of events that are experienced repeatedly. Since predictions of neural activity can contribute to many aspects of behavior from motor control to cognitive strategies, this potentially explains the ubiquitous activation of the cerebellum observed with a wide range of tasks during recordings with noninvasive techniques such as positron-emission tomography (PET) and functional magnetic resonance imaging (fMRI). The model provides an explanation for the changes in cerebellar activation during learning observed with these imaging techniques as well as for the changes in activation observed with task complexity and with the processing of unexpected events.
The feasibility of constructing predictive signals from known cell responses in the central nervous system is tested with the Predictive Cerebellar Model, where detailed responses of neural cells are used as inputs. Specifically, the model was constructed to demonstrate the possibility of establishing predictions of eye position that are directly relevant for the accurate function of the VOR system.
Cerebellar predictions may influence the VOR in several different ways. For example, the cerebellum could be involved with predicting eye position which would explain its participation in the eye movement neural integrator. The neural integrator is a group of nuclei in the brain stem which integrate eye velocity to yield an eye position signal. At another stage of processing, one part of the cerebellum may be specialized in predicting eye vergence position (vergence angle) and another part of the cerebellum may use this prediction to modulate predictively the VOR gain for future eye vergence position. This is how the Predictive Cerebellar Model achieves predictive modulation of the VOR gain during vergence eye movements.
The Predictive Cerebellar Model can predict the vergence angle up to 225 ms in advance. This prediction was used as the vergence input to the Global VOR Model of chapter 3 to accurately reproduce VOR psychophysical data from monkeys during changes in vergence ([Snyder, Lawrence and King1992]). To predict the vergence angle, the Predictive Cerebellar Model learned to combine the dynamical responses of hundreds of vergence-disparity cells of cortical visual area V1 ([Trotter1995]; [Trotter et al.1992]). These cells have tuning curves to both vergence and horizontal disparity and therefore express stereotypical responses during vergence eye movements caused by changes in target fixation distance. A dynamical vergence model developed by Schor (Schor92) was used to generate the vergence eye movements.