Condensed vector machines: learning fast machine for large data

Area of Science:

• Machine Learning
• Artificial Intelligence
• Computer Science

Background:

• Scalability is a major challenge for kernel-based methods and Support Vector Machines (SVMs).
• The quadratic memory requirement for kernel matrices hinders training on large datasets.
• Existing decomposition methods for SVMs are limited by a large number of support vectors (SVs), impacting performance.

Purpose of the Study:

• To address the scalability limitations of SVMs.
• To reduce the number of support vectors (SVs) for efficient training and testing.
• To develop a more compact and faster SVM model without compromising generalization ability.

Main Methods:

• Integration of a vector combination scheme with incremental working set selection for SVM training.
• Simplifying the SVM solution to maintain a minimal set of final support vectors (SVs).

Main Results:

• The proposed method significantly reduces the number of support vectors (SVs).
• Achieved faster training and testing times on large benchmark datasets.
• Maintained generalization ability equivalent to conventional SVMs.

Conclusions:

• The integrated vector combination scheme enhances SVM scalability and efficiency.
• Condensed SVMs offer reduced resource demand and improved performance.
• This approach makes SVMs more practical for large-scale applications.