Representative Vector Machines: A Unified Framework for Classical Classifiers.

Area of Science:

• Computer Science
• Machine Learning
• Pattern Recognition

Background:

• Nearest Neighbor (NN), Support Vector Machine (SVM), and Sparse Representation-based Classification (SRC) are established pattern classification methods.
• These classifiers are typically viewed as independent solutions with distinct theoretical or application origins.

Purpose of the Study:

• To introduce a novel, unified framework for pattern classification called Representative Vector Machines (RVMs).
• To reveal the underlying relationships among classical classifiers by interpreting them as special cases of RVMs.
• To inspire the development of new, advanced classification methods within the RVM framework.

Main Methods:

• The proposed Representative Vector Machines (RVMs) assign class labels based on the nearest representative vector.
• Classical classifiers (NN, SVM, SRC) are shown to be special cases of RVMs with specific definitions of representative vectors.
• A novel classifier, the Discriminant Vector Machine (DVM), is developed, utilizing k-Nearest Neighbors (k-NNs), M-estimator, and manifold regularization.

Main Results:

• RVMs provide a unified perspective, demonstrating that NN, SVM, and SRC are special instances of this general framework.
• The RVM framework facilitates the creation of new classifiers, exemplified by the robust DVM.
• Extensive experiments on face recognition, object categorization, and action recognition tasks show DVM outperforms other classifiers.

Conclusions:

• Representative Vector Machines (RVMs) offer a unifying framework for understanding and developing pattern classification methods.
• The RVM approach reveals inherent connections between existing classifiers and provides a foundation for novel algorithms like DVM.
• The Discriminant Vector Machine (DVM) demonstrates superior performance in various visual recognition tasks, validating the potential of the RVM framework.