Fig. 3: Moving bump forecast performance depends on specific properties of the recurrent connections.
From: Waves traveling over a map of visual space can ignite short-term predictions of sensory input
a Structural similarity (SSIM) between a forecast frame and the ground truth as a function of the closed-loop forecast video frame. Each curve corresponds to a different network parameter implementation. Curves have been smoothed by a moving-average filter (filter width of 30 time steps). Shaded error is the absolute difference between filtered and unfiltered. b Total structural similarity, in which a single SSIM is calculated for the whole movie as a function of the recurrence-to-input ratio. In the parameter space, each point differs only in recurrent strength. Smoothing and error shading is the same as in (a). c Total structural similarity as a function of recurrent length, which is the fraction of the network’s side length spanned by one standard deviation of the Gaussian connectivity kernel. In the three-dimensional parameter space comprising the recurrent strength (rs), recurrent length (rl), and input strength (is), averages (n = 89) across rs-is planes at fixed rl were computed (gray curve). Solid gray line: average. The peak coincides with the standard-deviation width of the Gaussian bump stimulus (dashed vertical line). Shaded area: variance. Solid black curve: maximum structural similarity at each recurrent length. Source data are provided as a Source Data file.
