The authors study learning in a parallel, neural-network implementation of an image-recognition network recently constructed to synthesize model-based and data-driven approaches to the recognition problem. Learning in this context includes three considerations: (i) learning the basic implications of a hierarchical model-based description, (ii) learning the weights in analogy to conventional neural nets, and (iii) learning new features to update the model. The authors present examples as well as simulation results on new models of learning suggested by optimal control techniques