L220, 11:00 21.7.2010

Multiple kernel learning is becoming a popular method of boosting the performance of classifiers. It involves combining different kernels and finding the appropriate mixing parameters to achieve performance improvement. We review a multiple kernel learning (MKL) technique called lp-regularised multiple kernel Fisher discriminant analysis (MK-FDA), and investigate the effect of feature space de-noising on MKL. Experiments in image and video retrieval show that with both, the original kernels or de-noised kernels, lp MK-FDA outperforms its fixed-norm counterparts. Experiments also show that feature space de-noising boosts the performance of both single kernel FDA and lp MK-FDA, and that there is a positive correlation between the learnt kernel weights and the amount of variance kept by feature space de-noising. Based on these observations, we argue that in the case where the base feature spaces are noisy, linear combination of kernels cannot be optimal. An MKL objective function which can take care of feature space de-noising automatically, and which can learn a truly optimal (non-linear) combination of the base kernels, is yet to be found.

Josef Kittler, Centre for Vision, Speech and Signal Processing, University of Surrey, Guildford GU2 7XH, United Kingdom

Speakers