Class Discovery in Galaxy Classification

In recent years, automated, supervised classification techniques have beenfruitfully applied to labeling and organizing large astronomical databases.These methods require off-line classifier training, based on labeled examplesfrom each of the (known) object classes. In practice, only a small batch oflabeled examples, hand-labeled by a human expert, may be available fortraining. Moreover, there may be no labeled examples for some classes presentin the data, i.e. the database may contain several unknown classes. Unknownclasses may be present due to 1) uncertainty in or lack of knowledge of themeasurement process, 2) an inability to adequately ``survey'' a massivedatabase to assess its content (classes), and/or 3) an incomplete scientifichypothesis. In recent work, new class discovery in mixed labeled/unlabeled datawas formally posed, with a proposed solution based on mixture models. In thiswork we investigate this approach, propose a competing technique suitable forclass discovery in neural networks, and evaluate both methods forclassification and class discovery on several astronomical data sets. Ourresults demonstrate up to a 57% reduction in classification error compared to astandard neural network classifier that uses only labeled data