基于深度玻尔兹曼机的乐器分类问题研究

应用传统浅层模型处理乐器分类任务存在非线性拟合能力较差的问题，使分类准确率得不到有效保证,有必要引入深度学习方法提升复杂任务的非线性建模能力。将深度玻尔兹曼机作为特征提取器提取表达能力更强的数据特征，分别以SVM与Softmax分类器作为深度神经网络的顶层设置形成DBM SVM组合模型与DBM Softmax组合模型，引入平均场理论和动量项因子优化网络训练过程。将上述两组模型及单一SVM分类器在5类乐器音频数据上进行对比实验，两种深度学习组合模型的分类准确率分别达到89.29%和87.5%，与传统浅层分类方法SVM的73.21%的准确率相比优势明显。实验结果表明深度玻尔兹曼机在乐器分类领域的应用颇具前景。

Research on musical instrument classification based on deep Boltzmann machine

The application of traditional shallow model to instrument classification task had the problem of poor nonlinear fitting ability, so that the accuracy of classification was not guaranteed effectively. It was necessary to introduce deep learning method to improve the nonlinear modeling ability of complex tasks. Deep Boltzmann machine was used as feature extractor to abstract more expressive deep learning features. SVM and Softmax classifier were respectively used as top layer of deep neural network to form DBM SVM and DBM Softmax combined model. Besides, the mean field theory and momentum factor were introduced to optimize the network training process. The above two sets of models and single SVM classifier was compared on 5 kinds of musical instruments audio data. The classification accuracy of the two types of deep learning combination models reached 89.29% and 87.5%respectively, compared with the accuracy of the traditional shallow classification method SVM of 73.21% .The experimental results show that the application of deep Boltzmann machine in the field of musical instrument classification is very promising.