Probing the critic: approaches to connectionist pattern synthesis

Considers the application of connectionist algorithms to pattern synthesis. The emphasis is on general pattern synthesis approaches which may be tailored to model arbitrary phenomena to some degree of accuracy. One pattern synthesis task is to create novel samples of a given class of patterns, where class membership is specified by example rather than by rule. A recently introduced paradigm, termed creation by refinement (CBR), accomplishes this task. The CBR paradigm employs a standard neural net which has been trained to recognize the desired class of patterns. CBR probabilistically explores the pseudouniverse of the function learned by this critic net. The author considers several pattern synthesis approaches based on CBR. A variation of CBR based on a competitive learning algorithm is introduced.<>