Adaptive coding of microstructure of natural images. Finding perceptually meaningful, highly informative, and computationally efficient representations of minuscule parts of digitized natural scenes is a key to designing efficient machine learning algorithms to solve high level vision tasks. We will discuss an approach to constructing such representations based on the basic principles of information theory and on a large data set of natural images. The resulting descriptors of image microstructure are simply a compact code for the results of a set of appropriate significance tests performed on image contrasts within a tiny image neighborhood (the primary, or central,"receptive field"). The distributions of the test statistics are, however, learned from larger surrounding neighborhoods ("the summation receptive fields"), thus adapting the tests to medium range dependences in the underlying random field. The resulting code for the central receptive field can also be thought of as an estimator of the local orientation of curves, and it also appears suitable for efficient integration of the local geometric structure ("contour integration"). Finally, we will attempt to explain how the above approach is consistent with the recently proposed views on architectural and functional properties of the visual cortical network in the brain.