Abstract:

A multiple kernel learning (MKL) technique called lp regularized multiple kernel Fisher discriminant analysis (lp MK-FDA) was reviewed, and MKL′s performance was compared  fixed-norm and p-norm. According to the phenomenon that original feature space  noises exist, the effect of feature space denoising on MKL was investigated. Experiments on the VOC 2007 dataset show that with both the original kernels or denoised kernels, lp MKFDA outperforms its fixed-norm counterparts, and the feature space denoising boosts the performance of both single kernel FDA and lp MKFDA, and also there is a positive correlation between the learnt kernel weights and the amount of variance kept by feature space denoising.

Key words: kernel methods; multiple kernel learning; denoising space; kernel FDA; kernel PCA