An important problem for both biological and machine vision is the construction of scene representations from 2-D image data that are useful for recognition. One problem is that there can be more than one world out there giving rise to the image data at hand. Additional constraints regarding the nature of the environment have to be used to narrow the range of solutions. Although effort has gone into understanding these constraints, relatively little has been done to understand how neurallike learning networks may be used to solve scene-from-image problems. A paradigm is proposed in which stochastic models of scene properties are used to provide samples of image and scene representations. Distributed associative networks are taught, by example, the statistical constraints relating the image to the representation of the scene. This technique is applied to problems in optic flow, shape-from-shading, and stereo.