一种新的图像置乱算法 - 华东理工大学学报.doc

基于混合判别受限波兹曼机的音乐自动标注算法研究 王诗俊1 陈宁1 1华东理工大学，信息科学与工程学院，上海 200237 (MilkEffect@163.com) Annotating Music with Hybrid Discriminative Restricted Boltzmann Machines Wang Shijun1 and Chen Ning1 1 (Department of Information Science and Engineering, East China University of Science and Technology, Shanghai 200237) Abstract In the case of music annotation, the amount of unlabeled music data is often much smaller than the labeled music training data, which often results in unsatisfied training results. Generation model is able to fix the problem on some level, which can get more high-quality results on a small amount of training data, however, in the case of sufficient training datasets, the generation model is inferior to the discriminative model. This paper presents a model that combines the advantages of the generation model and the discriminative model to annotate music, which is called the hybrid discriminative Restricted Boltzmann Machines. The hybrid model can improve the accuracy of the music annotation tasks. The result of our experiment shows that the hybrid model is not only better than the traditional machine learning models but also better than the single discriminative Boltzmann Machines when the amount of training data is small. At the meantime, the experiment confirms that the hybrid model has similar results like the discriminative model in the case of sufficient amount of training data. In addition, the Dropout method is presented to improve the model, which can prevent the overfitting because of the small amount of the training datasets. Key words music annotation; hybrid discriminative Restricted Boltzmann Machines; machine learning; artificial intelligence 摘要 对于音乐自动标注任务，在很多情况下，未标注的歌曲量远远超过了已标注的歌曲数据，从而导致训练结果往往不理想。生成模型能够在某种程度上适应少量数据集的情况，得出较为满意的结果。然而，在有充分数据集的情况下生成模型的效果却劣于判别模型。本文提出了一种结合了生成模型与判别模型两者优势的面向音乐自动标注的混合判别波兹曼机模型，该模型可明显提升音乐自动标注的准确率。实验结果表明，混合波兹曼机的效果不仅好于传统的机器学习模型，同时，模型在拥有足够训练数据量的情况下与判别模型效果相当，且在训练集较少的情况下效果也好于判别模型。另外为了防止模型过拟合，本文引入了Dropout规则化方法以进一步加强