Inner Product Optimization for Effective Multiple Kernel Learning

6th Annual 2018 International Conference on Geo-Spatial Knowledge and Intelligence

Niu, Guo^1 ; Duan, Zhikui^1

School of Electronics and Information Engineering, Foshan University, Foshan, China^1

关键词: Benchmark datasets;    Classification framework;    Closed form solutions;    Computational costs;    Iterative operation;    Multiple Kernel Learning;    Multiple kernels;    Symmetric positive definite matrices;   

Others  :  https://iopscience.iop.org/article/10.1088/1755-1315/234/1/012063/pdf
DOI  :  10.1088/1755-1315/234/1/012063

来源: IOP

【 摘 要 】

This paper proposes a novel and effective method for multiple kernel learning (MKL) that differs from standard MKL approaches. The new algorithm, named effective multiple kernel learning (EMKL), proposes a learn function space generated by multiple kernels with a group of parameters, as well as constructs a new inner-product with an optimized symmetric positive-definite matrix for the solution space of learning problem. We also describe the proposed EMKL algorithm within a general-purpose image classification framework. The EMKL algorithm has a closed-form solution, thus enabling optimization without the need for the additional iterative operations that are common in standard MKL algorithms. Hence, our proposed EMKL has a lower computational cost than that of conventional MKL techniques. The results of experiments on two benchmark datasets show the effectiveness of the proposed algorithm.