Abstract:
|
All successful classification tasks depend critically on a representation of the data for which there exists a learnable function distinguishing the classes. Without such a “class-informative” set of features, classification is not possible. A methodology is proposed which provides non-linear, semi-supervised class-informative feature set learning using a novel loss function and a dynamic training scheme. In the semi-supervised case, this methodology is shown to improve classification performance by incorporating the structure of unlabeled observations into the learned feature map.
|