Characterizing the Temporal Dynamics of Information in Visually Guided Predictive Control Using LSTM Recurrent Neural Networks

Theories for visually guided action account for online control in the presence of reliable sources of visual information, and predictive control to compensate for visuomotor delay and temporary occlusion. In this study, we characterize the temporal relationship between information integration window and prediction distance using computational models. Subjects were immersed in a simulated environment and attempted to catch virtual balls that were transiently "blanked" during flight. Recurrent neural networks were trained to reproduce subject's gaze and hand movements during blank. The models successfully predict gaze behavior within 3 degrees, and hand movements within 8.5 cm as far as 500 ms in time, with integration window as short as 27 ms. Furthermore, we quantified the contribution of each input source of information to motor output through an ablation study. The model is a proof of concept for prediction as a discrete mapping between information integrated over time and a temporally distant motor output.