Browse Prior Art Database

New Efficient Learning Architecture for Multi-Class Support Vector Machine Based Classification

IP.com Disclosure Number: IPCOM000127326D
Original Publication Date: 2005-Aug-23
Included in the Prior Art Database: 2005-Aug-23
Document File: 1 page(s) / 46K

Publishing Venue

IBM

Abstract

Disclosed is a program that implements a new and efficient learning architecture that reduces the number of Support Vector Machines (SVMs) needed to achive N-class classification where N is very large. Standard methods for learning N classes using SVMs, involve (i) learning N separate SVMs where each SVM learns to discriminate a unique class against all others, or (ii) learning (N*(N-1)) /2 SVMs where each SVM discriminates between a single pair of classes among all (N *(N-1))/2 pairwise combinations possible with N classes. However, in situations where N becomes large, these methods involve learning a number of SVMs that increases linearly or quadratically with N. This becomes undesirable owing to the high costs of training and executing an SVM. For a N-class problem, our architecture involves only training O(log_2 (N)) SVMs.

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 52% of the total text.

Page 1 of 1

New Efficient Learning Architecture for Multi-Class Support Vector Machine Based Classification

Disclosed is a program that implements a new and efficient learning architecture that reduces the num achive N-class classification where N is very large.

The problem of automatic classification involves assigning an item, usually described using its feature one of the many pre-defined categories using a classifier algorithm. Popular classifier algorithms inclu Bayesian classifiers, neural networks, Support Vector Machines, etc. When only two categories are p discriminated it is called a binary classification problem, while a problem consisting of more than two c referred to as multi-class classification.

In using Support Vector Machines (SVMs) which are learning systems that use linear functions in a hi space for multi-class classification problems, commonly used learning architectures do not scale well number of classes. Large number of classes are very often encountered in business data analysis app Standard methods for learning N classes using SVMs, involve (i) learning N separate SVMs where each SVM learns to discriminate a unique class against all others, or (ii) learning (N*(N-1)) /2 SVMs w SVM discriminates between a single pair of classes among all (N *(N-1))/2 pairwise combinations pos classes. However, in situations where N becomes large, these methods involve learning a number of that increases linearly or quadratically with N. This becomes undesirable owing to the high costs of training and executing an SVM.

Our invention propo...