A Control Systems Approach to Understanding Saccadic Eye Movements

Abstract

The oculomotor system is extensively studied using concepts of control theory like optimal control, feedback control, and predictive control. But these studies have largely been restricted to investigating the central tendencies of movements like mean kinematics or the main sequence. But owing to the inherent noise in the motor system, eye movements generated by the oculomotor system exhibit considerable variability. Hence, it is imperative to study variability in these movements and consider noise to be an important aspect of oculomotor control. Therefore, in this thesis, a type of voluntary eye movement called saccades is studied, with emphasis on mean trajectories as well as inter-trial variability in the trajectories. A stochastic modelling approach has been used to gain deeper insights into control architecture of the saccade execution system. Firstly, using the stochastic optimal control framework it is shown that the saccadic system may have an explicit velocity plan. A new trajectory tracking stochastic optimal control model is proposed, which is observed to outperform the existing endpoint model in predicting the mean velocity profiles. Further evidence of this explicit velocity planning in saccade execution was provided using the saccade data obtained from an eye-hand coordination task. This is an important finding given that the dominant view in the field is that saccades are planned based on target displacement only. Secondly, using a stochastic saccade generation model with internal feedback, it is shown that the saccadic system uses both displacement and velocity information. The proposed model had a displacement block and velocity block with separate input and internal feedback loop. This framework was validated using behavioral data of horizontal saccades. Thirdly, the proposed stochastic dual model was extended for oblique saccade generation using an error coupling scheme. In case of oblique saccades too, the behavioral data was best fit by a dual model which used a convex combination of both...