Dynamical emergence of a neural solution for motion integration

摘要: FIGURE 1:(a) Measuring motion of an object from a localized receptive field RF may lead to ambiguous results (red arrow) compared to physical motion (blue arrow): it’s the aperture problem.(b) In natural scenes, we may constrain motion representation by using the prior knowledge that motion is roughly conserved along path-lines [Burgi et al., 2000]. We denote it as a “liquid-like” prior since it is equivalent to the Navier-Stokes advection term that controls the flow of incompressible fluids.(c) Using a markovian model of noise such as is represented by the blurry area, we may infer from a previous motion representation at time ta new measurement at time t+ dt using Eq.(1) and the prior predictive knowledge: For all possible positions and speeds, it is computed by integrals over all velocities with the corresponding predicted position. This predictive field thus defines a simple yet computationally extensive dynamical system.