A bi-level formulation for multiple kernel learning via self-paced training

Abstract

Multiple kernel learning (MKL) is a crucial issue which has been widely researched over the last two decades. Although existing MKL algorithms have achieved satisfactory performance in a broad range of applications, these methods do not adequately consider the adverse effects of unreliable or less reliable instances. To handle this shortcoming, we formulate multiple kernel learning in a bi-level learning paradigm consisting of the kernel combination weight learning (KWL) stage and the self-paced learning (SPL) stage, which alternatively negotiate with each other. The KWL stage dynamically absorbs reliable instances into model learning to accurately capture neighborhood relationships and obtains kernel coefficients via maximizing both global and local kernel alignment in a common schema. The SPL stage automatically evaluates the reliability of training samples via self-paced training. The extensive experiments indicate the robustness and superiority of the presented approach in comparison with existing MKL methods.