Multiple Kernel Learning with Hierarchical Feature Representations

Abstract

AbstractIn this paper, we suggest multiple kernel learning with hierarchical feature representations. Recently, deep learning represents excellent performance to extract hierarchical feature representations in unsupervised manner. However, since fine-tuning step of deep learning only considers global level of features for classification problems, it makes each layers hierarchical features intractable. Therefore, we propose a method to employ the combined multiple levels of pre-trained features via Multiple Kernel Learning (MKL). MKL is lately proposed optimization problem in classification and is applied to various machine learning problems. MKL automatically finds the best combination of kernels. By applying multiple kernel learning to hierarchical features pre-trained by deep learning, we obtain the optimal combinations of multiple levels of features for the classification task. Also, MKL is applied to analyze the contribution of each layer of features for classification by obtained weight of each kernel.KeywordsMultiple Kernel LearningDeep LearningDeep Belief Network