Abstract
Several studies investigating the development of amblyopia and strabismus suggest a strong interaction between vergence and accommodation. For example, patients suffering from strabismus often develop amblyopia and subjects with amblyopia show decreased vergence and accommodation performance. Here we present the first computational model for the joint development of accommodation and vergence control in the active efficient coding framework. We use an online sparse coding algorithm to learn binocular receptive fields similar to those in V1 simple cells. These adapt online to the input statistics by maximizing coding efficiency. Simultaneously, the learned sparse representation is used to determine the reward for two actor-critic reinforcement learners (RLs), which control accommodation and vergence, respectively. By optimizing coding complexity (for accommodation control) and efficiency (for vergence control) the system learns to focus images with zero disparity under healthy conditions. Interestingly, the accommodation RL learns to deduce the correct command from the input disparity. We simulate an anisometropic case where the refraction power of one eye is decreased. In this situation our model chooses to focus close objects with the healthy and distant objects with the hyperopic eye. Vergence performance remains high as long as the refraction difference stays small. However, when focusing the object with one eye leads to a highly blurred input for the other eye, the receptive fields become more and more monocular. Thus, the RLs are no longer able to assess the exact input disparity, which ultimately leads to a decrease of both the vergence and accommodation performance. In conclusion, we present, to the best of our knowledge, the first model for the joint learning of vergence and accommodation control. The model explains how the brain might learn to exploit disparity signals to control both vergence and accommodation and how refractive errors could derail this process.
Meeting abstract presented at VSS 2017