We propose to study how information about objects and events in the environment can be reconstructed from collections of signals each arising from a small portion of space and/or time. The brain is continually faced with this task: eg, each retinal photoreceptor responds to light from a small bit of space compared with the size of real-world objects. In the time domain, early sensory neurons (visual, auditory and tactile) signal only instantaneous external stimuli, leaving the brain with the task of reconstructing events that extend over time (eg, speech). Neural computation of global constructs from local signals is therefore fundamental and ubiquitous, yet our understanding of it is still rudimentary. We will use neural modeling to study the topic. We have two specific goals, addressing two tasks which we have previously characterized experimentally, behaviorally and with brain imaging. (i) Recover the global motion of objects from local motion signals. For example, a spinning wheel's global motion corresponds to a single quantity, its angular velocity; but each point on it generates a local motion signal of a different direction and speed. The task is further complicated since other nearby local motion signals may arise from independently moving objects. The modeling will implement and test our theory that global motion computation is achieved by continual neural computation between two modules, one specializing in integration and the other in segmentation. (ii) Build models of neural networks with a hierarchical organization of Temporal Receptive Windows (TRWs). Defining a neuron's TRW as the length of time prior to a response during which sensory information can affect that response, we have recently shown that the brain is organized with a hierarchy of TRWs, ranging from 3 sec in sensory areas to 40 sec in higher cortical areas. The models will be used for studying basic properties of such networks and for starting to apply them to real-world tasks.