Incorporating Developmental Factors into Models of Accommodation and Vergence

Purpose: To investigate how known developmental factors can be incorporated into neural network models of the accommodation and vergence systems. To demonstrate that models of this type can learn adult-like linkages between accommodation and vergence by simple associative mechanisms. Methods: Existing linear systems models were used as a starting point for the design of a series of multi-layer neural network models of accommodation and vergence. Unlike the systems models that are generally specified in detail by hand, the neural models learn for themselves how to perform appropriately. A general connectionist framework was set up and trained, by updating initially small random connection weights using gradient descent, so that the network outputs for realistic patterns of input cues (blur, disparity, etc.) corresponded to appropriate patterns of accommodation and vergence responses. Maturational factors were incorporated into the models by using empirical infant data to define starting points, by varying parameters such as depth of field during training and by bringing in cues at appropriate developmental stages in the model. The outputs and internal connection weights were recorded throughout the training to determine the developmental progression of the model. Results: Training the models in this way resulted in developmental progressions which matched empirical infant data and automatically generated linkages between the accommodation and vergence subsystems with strengths within the normal adult range. Conclusions: A range of known maturational factors together with empirical infant data can be used to constrain developmental models of accommodation and vergence. Starting with small random weights, these neural network models can learn to associate changes in accommodation with related changes in vergence, and develop adult-like linkages between accommodation and vergence. This provides a demonstration of how such linkages might develop through simple associative mechanisms.