Learning Data Dependent Composite Kernels for Robust Image Retrieval - A Genetic Programming Approach

摘要

Kernel methods are a class of pattern recognition and machine learning algorithms that map data to a high dimensional space and perform various learning tasks like clustering or regression in that space. The mapping from the low dimensional space to the high dimension is done implicitly by the use of a kernel function. But the question of how to choose the kernel is an interesting and intriguing one. The choice of the kernel and its parameters is usually done using cross-validation. We propose a methodology of learning a kernel from data using genetic programming. With the aid of genetic algorithms, we constructed composite kernels and compared their performance with an ad-hoc kernel in the domain of image retrieval. The learned composite kernels showed consistent better performance compared to the individual kernel.