Biological support exists for the idea that large-scale models of the brain should be parallel, distributed, and associative. Some of this neurobiology is reviewed. It is then assumed that state vectors, large patterns of activity of groups of individual somewhat selective neurons, are the appropriate elementary entities to use for cognitive computation. Simple neural models using this approach are presented that will associate and will respond to prototypes of sets of related inputs. Some experimental evidence supporting the latter model is discussed. A model for categorization is then discussed. Educating the resulting systems and the use of error correcting techniques are discussed, and an example is presented of the behavior of the system when diffuse damage occurs to the memory, with and without compensatory learning. Finally, a simulation is presented which can learn partial information, integrate it with other material, and use that information to reconstruct missing information.