Online learning with (multiple) kernels: a review.

Area of Science:

• Machine Learning
• Computational Biology
• Computer Vision

Background:

• Kernel methods offer powerful tools for complex pattern recognition tasks.
• Online learning algorithms adapt to data sequentially, crucial for dynamic environments.
• Multiclass classification is a fundamental problem in machine learning with broad applications.

Purpose of the Study:

• To review and empirically validate various kernel methods for online multiclass classification.
• To compare margin-based, Gaussian process, and multiple kernel learning approaches.
• To assess method performance on real-world biological and object recognition datasets.

Main Methods:

• Examination of margin-based methods, including the perceptron algorithm.
• Analysis of nonparametric probabilistic approaches using Gaussian processes.
• Investigation of online multiple kernel learning techniques.

Main Results:

• Empirical validation of diverse kernel methods on protein fold recognition, Caltech101, and Caltech256 datasets.
• Demonstration of the effectiveness of different kernel combinations for multiclass classification.
• Comparative analysis highlighting the strengths of various approaches across different data types.

Conclusions:

• Kernel methods provide a robust framework for online multiclass classification.
• The choice of kernel and learning approach impacts performance based on data characteristics.
• Gaussian processes and multiple kernel learning show promise for complex recognition tasks.