Loading...
Please wait, while we are loading the content...
Similar Documents
Serendipitous Learning: Learning Beyond the Predefined Label Space
| Content Provider | CiteSeerX |
|---|---|
| Author | Liu, Yan Si, Luo Zhang, Dan |
| Abstract | In machine learning, most supervised learning methods are developed for learning from training examples within a predefined label space and make predictions on which classes (among those predefined labels) each test example belongs to. However, in many real world applications, such as text and image categorization, we are often confronted with the learning scenarios in which the label space needs to be enlarged during testing phase, that is, the test examples may belong to some new classes which have not been defined during the training phase. How to make accurate predictions for examples in existing classes and at the same time identify novel examples from new classes is an extremely challenging problem, which has not been addressed before. This paper explores this novel and practical learning scenario, which is named as Serendipitous Learning (SL). The basic idea is to leverage the knowledge in the labeled examples to help identify the unknown classes. In particular, a maximum margin formulation is proposed to model both the classification loss on the known classes and the clustering performance on the unknown ones. An efficient optimization algorithm is designed based on Constrained Concave-Convex Procedure (CCCP) and the bundle method to solve the corresponding optimization problem. Furthermore, an efficient online learning algorithm is proposed for large-scale applications of the proposed problem with guaranteed bounds on regret. The experimental results on two synthetic datasets and two real world datasets demonstrate the advantages of the proposed method over other baseline algorithms. |
| File Format | |
| Access Restriction | Open |
| Subject Keyword | Labeled Example Efficient Optimization Algorithm Training Phase Guaranteed Bound Unknown One Test Example Belongs Machine Learning Image Categorization Practical Learning Scenario Baseline Algorithm Test Example Many Real World Application Unknown Class Label Space Accurate Prediction Basic Idea Maximum Margin Formulation Supervised Learning Method Serendipitous Learning Efficient Online Bundle Method Real World Datasets Classification Loss Predefined Label Space Synthetic Datasets Corresponding Optimization Problem Constrained Concave-convex Procedure Large-scale Application Challenging Problem Experimental Result Novel Example |
| Content Type | Text |
| Resource Type | Article |
