Class Discovery in Galaxy Classification
David Bazell, David John Miller · The Astrophysical Journal · 2005
In recent years, automated, supervised classification techniques have been fruitfully applied to labeling and organizing large astronomical databases. These methods require off-line classifier training, based on labeled examples from each of the (known) object classes. In practice, only a small batch of labeled examples, hand-labeled by a human expert, may be available for training. Moreover, there may be no labeled examples for some classes present in the data; i.e., the database may contain several unknown classes . Unknown classes may be present because of (1) uncertainty in or lack of knowledge of the measurement process, (2) an inability to adequately "survey" a massive database to assess its content (classes), and/or (3) an incomplete scientific hypothesis. In recent work, the question of new class discovery in mixed labeled/unlabeled data was formally posed, with a proposed solution based on mixture models. In this work we investigate this approach, propose a competing technique suitable for class discovery in neural networks, and evaluate methods for both classification and class discovery in several astronomical data sets. Our results demonstrate up to a 57% reduction in classification error compared to a standard neural network classifier that uses only labeled data.