Partial Discriminative Training of Neural Networks for Classification of Overlapping Classes

In applications such as character recognition, some classes are heavily overlapped but are not necessarily to be separated. For classification of such overlapping classes, either discriminating between them or merging them into a metaclass does not satisfy. Merging the overlapping classes into a metaclass implies that within-metaclass substitution is considered as correct classification. For such classification problems, I propose a partial discriminative training (PDT) scheme for neural networks, in which, a training pattern of an overlapping class is used as a positive sample of its labeled class, and neither positive nor negative sample for its allied classes (classes overlapping with the labeled class). In experiments of handwritten letter recognition using neural networks and support vector machines, the PDT scheme mostly outperforms crosstraining (a scheme for multi-labeled classification), ordinary discriminative training and metaclass classification.

• Type: Conference
• Year: 2008
• Where: ANNPR
• Authors: Cheng-Lin Liu
• Keywords: ANNPR 2008, Discriminative Training, Neural Networks, Overlapping Classes