Fig. 4: The recurrent network performs next-frame forecasting of a natural video input.
From: Waves traveling over a map of visual space can ignite short-term predictions of sensory input
a Training follows as in the moving bump example (Fig. 2a). b Next-frame closed-loop forecasting follows as in the moving bump example (Fig. 2b). c Video frames of the data: a person walking. d Corresponding closed-loop forecasts generated by the network model in the case of optimal recurrence. e Corresponding network states for the optimal-recurrence case (d). Cosine of phase is shown. f Same as (d), but in the absence of recurrence.
